diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/chat_template.jinja b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..1bad6a0f648dccdbec523ca79ba90fbcfc806af0
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/chat_template.jinja
@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/special_tokens_map.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/special_tokens_map.json
new file mode 100644
index 0000000000000000000000000000000000000000..b43be96621d147110fb8a18b5776ec6e38516127
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/special_tokens_map.json
@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|eot_id|>"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/tokenizer_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..3beeacc86a6ca3cae14ad3004263ab74a4bac07a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/tokenizer/tokenizer_config.json
@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|eot_id|>",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..88a7ab65ecc03929fde705948a65e30260067bc3
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/config.json
@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "float32",
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "nemo_version": "0.2.0",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": 128009,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "vocab_size": 128256
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
new file mode 100644
index 0000000000000000000000000000000000000000..494ffc91e4983233a079912895e9dc90c8c11a79
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
@@ -0,0 +1,256 @@
+{
+  "model.embed_tokens.weight": 1,
+  "model.layers.0.input_layernorm.weight": 1,
+  "model.layers.0.mlp.down_proj.weight": 1,
+  "model.layers.0.mlp.gate_proj.weight": 1,
+  "model.layers.0.mlp.up_proj.weight": 1,
+  "model.layers.0.post_attention_layernorm.weight": 1,
+  "model.layers.0.self_attn.k_proj.weight": 1,
+  "model.layers.0.self_attn.o_proj.weight": 1,
+  "model.layers.0.self_attn.q_proj.weight": 1,
+  "model.layers.0.self_attn.v_proj.weight": 1,
+  "model.layers.1.input_layernorm.weight": 1,
+  "model.layers.1.mlp.down_proj.weight": 1,
+  "model.layers.1.mlp.gate_proj.weight": 1,
+  "model.layers.1.mlp.up_proj.weight": 1,
+  "model.layers.1.post_attention_layernorm.weight": 1,
+  "model.layers.1.self_attn.k_proj.weight": 1,
+  "model.layers.1.self_attn.o_proj.weight": 1,
+  "model.layers.1.self_attn.q_proj.weight": 1,
+  "model.layers.1.self_attn.v_proj.weight": 1,
+  "model.layers.10.input_layernorm.weight": 1,
+  "model.layers.10.mlp.down_proj.weight": 1,
+  "model.layers.10.mlp.gate_proj.weight": 1,
+  "model.layers.10.mlp.up_proj.weight": 1,
+  "model.layers.10.post_attention_layernorm.weight": 1,
+  "model.layers.10.self_attn.k_proj.weight": 1,
+  "model.layers.10.self_attn.o_proj.weight": 1,
+  "model.layers.10.self_attn.q_proj.weight": 1,
+  "model.layers.10.self_attn.v_proj.weight": 1,
+  "model.layers.11.input_layernorm.weight": 1,
+  "model.layers.11.mlp.down_proj.weight": 1,
+  "model.layers.11.mlp.gate_proj.weight": 1,
+  "model.layers.11.mlp.up_proj.weight": 1,
+  "model.layers.11.post_attention_layernorm.weight": 1,
+  "model.layers.11.self_attn.k_proj.weight": 1,
+  "model.layers.11.self_attn.o_proj.weight": 1,
+  "model.layers.11.self_attn.q_proj.weight": 1,
+  "model.layers.11.self_attn.v_proj.weight": 1,
+  "model.layers.12.input_layernorm.weight": 1,
+  "model.layers.12.mlp.down_proj.weight": 1,
+  "model.layers.12.mlp.gate_proj.weight": 1,
+  "model.layers.12.mlp.up_proj.weight": 1,
+  "model.layers.12.post_attention_layernorm.weight": 1,
+  "model.layers.12.self_attn.k_proj.weight": 1,
+  "model.layers.12.self_attn.o_proj.weight": 1,
+  "model.layers.12.self_attn.q_proj.weight": 1,
+  "model.layers.12.self_attn.v_proj.weight": 1,
+  "model.layers.13.input_layernorm.weight": 1,
+  "model.layers.13.mlp.down_proj.weight": 1,
+  "model.layers.13.mlp.gate_proj.weight": 1,
+  "model.layers.13.mlp.up_proj.weight": 1,
+  "model.layers.13.post_attention_layernorm.weight": 1,
+  "model.layers.13.self_attn.k_proj.weight": 1,
+  "model.layers.13.self_attn.o_proj.weight": 1,
+  "model.layers.13.self_attn.q_proj.weight": 1,
+  "model.layers.13.self_attn.v_proj.weight": 1,
+  "model.layers.14.input_layernorm.weight": 1,
+  "model.layers.14.mlp.down_proj.weight": 1,
+  "model.layers.14.mlp.gate_proj.weight": 1,
+  "model.layers.14.mlp.up_proj.weight": 1,
+  "model.layers.14.post_attention_layernorm.weight": 1,
+  "model.layers.14.self_attn.k_proj.weight": 1,
+  "model.layers.14.self_attn.o_proj.weight": 1,
+  "model.layers.14.self_attn.q_proj.weight": 1,
+  "model.layers.14.self_attn.v_proj.weight": 1,
+  "model.layers.15.input_layernorm.weight": 1,
+  "model.layers.15.mlp.down_proj.weight": 1,
+  "model.layers.15.mlp.gate_proj.weight": 1,
+  "model.layers.15.mlp.up_proj.weight": 1,
+  "model.layers.15.post_attention_layernorm.weight": 1,
+  "model.layers.15.self_attn.k_proj.weight": 1,
+  "model.layers.15.self_attn.o_proj.weight": 1,
+  "model.layers.15.self_attn.q_proj.weight": 1,
+  "model.layers.15.self_attn.v_proj.weight": 1,
+  "model.layers.16.input_layernorm.weight": 1,
+  "model.layers.16.mlp.down_proj.weight": 1,
+  "model.layers.16.mlp.gate_proj.weight": 1,
+  "model.layers.16.mlp.up_proj.weight": 1,
+  "model.layers.16.post_attention_layernorm.weight": 1,
+  "model.layers.16.self_attn.k_proj.weight": 1,
+  "model.layers.16.self_attn.o_proj.weight": 1,
+  "model.layers.16.self_attn.q_proj.weight": 1,
+  "model.layers.16.self_attn.v_proj.weight": 1,
+  "model.layers.17.input_layernorm.weight": 1,
+  "model.layers.17.mlp.down_proj.weight": 1,
+  "model.layers.17.mlp.gate_proj.weight": 1,
+  "model.layers.17.mlp.up_proj.weight": 1,
+  "model.layers.17.post_attention_layernorm.weight": 1,
+  "model.layers.17.self_attn.k_proj.weight": 1,
+  "model.layers.17.self_attn.o_proj.weight": 1,
+  "model.layers.17.self_attn.q_proj.weight": 1,
+  "model.layers.17.self_attn.v_proj.weight": 1,
+  "model.layers.18.input_layernorm.weight": 1,
+  "model.layers.18.mlp.down_proj.weight": 1,
+  "model.layers.18.mlp.gate_proj.weight": 1,
+  "model.layers.18.mlp.up_proj.weight": 1,
+  "model.layers.18.post_attention_layernorm.weight": 1,
+  "model.layers.18.self_attn.k_proj.weight": 1,
+  "model.layers.18.self_attn.o_proj.weight": 1,
+  "model.layers.18.self_attn.q_proj.weight": 1,
+  "model.layers.18.self_attn.v_proj.weight": 1,
+  "model.layers.19.input_layernorm.weight": 1,
+  "model.layers.19.mlp.down_proj.weight": 1,
+  "model.layers.19.mlp.gate_proj.weight": 1,
+  "model.layers.19.mlp.up_proj.weight": 1,
+  "model.layers.19.post_attention_layernorm.weight": 1,
+  "model.layers.19.self_attn.k_proj.weight": 1,
+  "model.layers.19.self_attn.o_proj.weight": 1,
+  "model.layers.19.self_attn.q_proj.weight": 1,
+  "model.layers.19.self_attn.v_proj.weight": 1,
+  "model.layers.2.input_layernorm.weight": 1,
+  "model.layers.2.mlp.down_proj.weight": 1,
+  "model.layers.2.mlp.gate_proj.weight": 1,
+  "model.layers.2.mlp.up_proj.weight": 1,
+  "model.layers.2.post_attention_layernorm.weight": 1,
+  "model.layers.2.self_attn.k_proj.weight": 1,
+  "model.layers.2.self_attn.o_proj.weight": 1,
+  "model.layers.2.self_attn.q_proj.weight": 1,
+  "model.layers.2.self_attn.v_proj.weight": 1,
+  "model.layers.20.input_layernorm.weight": 2,
+  "model.layers.20.mlp.down_proj.weight": 2,
+  "model.layers.20.mlp.gate_proj.weight": 1,
+  "model.layers.20.mlp.up_proj.weight": 1,
+  "model.layers.20.post_attention_layernorm.weight": 2,
+  "model.layers.20.self_attn.k_proj.weight": 1,
+  "model.layers.20.self_attn.o_proj.weight": 1,
+  "model.layers.20.self_attn.q_proj.weight": 1,
+  "model.layers.20.self_attn.v_proj.weight": 1,
+  "model.layers.21.input_layernorm.weight": 2,
+  "model.layers.21.mlp.down_proj.weight": 2,
+  "model.layers.21.mlp.gate_proj.weight": 2,
+  "model.layers.21.mlp.up_proj.weight": 2,
+  "model.layers.21.post_attention_layernorm.weight": 2,
+  "model.layers.21.self_attn.k_proj.weight": 2,
+  "model.layers.21.self_attn.o_proj.weight": 2,
+  "model.layers.21.self_attn.q_proj.weight": 2,
+  "model.layers.21.self_attn.v_proj.weight": 2,
+  "model.layers.22.input_layernorm.weight": 2,
+  "model.layers.22.mlp.down_proj.weight": 2,
+  "model.layers.22.mlp.gate_proj.weight": 2,
+  "model.layers.22.mlp.up_proj.weight": 2,
+  "model.layers.22.post_attention_layernorm.weight": 2,
+  "model.layers.22.self_attn.k_proj.weight": 2,
+  "model.layers.22.self_attn.o_proj.weight": 2,
+  "model.layers.22.self_attn.q_proj.weight": 2,
+  "model.layers.22.self_attn.v_proj.weight": 2,
+  "model.layers.23.input_layernorm.weight": 2,
+  "model.layers.23.mlp.down_proj.weight": 2,
+  "model.layers.23.mlp.gate_proj.weight": 2,
+  "model.layers.23.mlp.up_proj.weight": 2,
+  "model.layers.23.post_attention_layernorm.weight": 2,
+  "model.layers.23.self_attn.k_proj.weight": 2,
+  "model.layers.23.self_attn.o_proj.weight": 2,
+  "model.layers.23.self_attn.q_proj.weight": 2,
+  "model.layers.23.self_attn.v_proj.weight": 2,
+  "model.layers.24.input_layernorm.weight": 2,
+  "model.layers.24.mlp.down_proj.weight": 2,
+  "model.layers.24.mlp.gate_proj.weight": 2,
+  "model.layers.24.mlp.up_proj.weight": 2,
+  "model.layers.24.post_attention_layernorm.weight": 2,
+  "model.layers.24.self_attn.k_proj.weight": 2,
+  "model.layers.24.self_attn.o_proj.weight": 2,
+  "model.layers.24.self_attn.q_proj.weight": 2,
+  "model.layers.24.self_attn.v_proj.weight": 2,
+  "model.layers.25.input_layernorm.weight": 2,
+  "model.layers.25.mlp.down_proj.weight": 2,
+  "model.layers.25.mlp.gate_proj.weight": 2,
+  "model.layers.25.mlp.up_proj.weight": 2,
+  "model.layers.25.post_attention_layernorm.weight": 2,
+  "model.layers.25.self_attn.k_proj.weight": 2,
+  "model.layers.25.self_attn.o_proj.weight": 2,
+  "model.layers.25.self_attn.q_proj.weight": 2,
+  "model.layers.25.self_attn.v_proj.weight": 2,
+  "model.layers.26.input_layernorm.weight": 2,
+  "model.layers.26.mlp.down_proj.weight": 2,
+  "model.layers.26.mlp.gate_proj.weight": 2,
+  "model.layers.26.mlp.up_proj.weight": 2,
+  "model.layers.26.post_attention_layernorm.weight": 2,
+  "model.layers.26.self_attn.k_proj.weight": 2,
+  "model.layers.26.self_attn.o_proj.weight": 2,
+  "model.layers.26.self_attn.q_proj.weight": 2,
+  "model.layers.26.self_attn.v_proj.weight": 2,
+  "model.layers.27.input_layernorm.weight": 2,
+  "model.layers.27.mlp.down_proj.weight": 2,
+  "model.layers.27.mlp.gate_proj.weight": 2,
+  "model.layers.27.mlp.up_proj.weight": 2,
+  "model.layers.27.post_attention_layernorm.weight": 2,
+  "model.layers.27.self_attn.k_proj.weight": 2,
+  "model.layers.27.self_attn.o_proj.weight": 2,
+  "model.layers.27.self_attn.q_proj.weight": 2,
+  "model.layers.27.self_attn.v_proj.weight": 2,
+  "model.layers.3.input_layernorm.weight": 1,
+  "model.layers.3.mlp.down_proj.weight": 1,
+  "model.layers.3.mlp.gate_proj.weight": 1,
+  "model.layers.3.mlp.up_proj.weight": 1,
+  "model.layers.3.post_attention_layernorm.weight": 1,
+  "model.layers.3.self_attn.k_proj.weight": 1,
+  "model.layers.3.self_attn.o_proj.weight": 1,
+  "model.layers.3.self_attn.q_proj.weight": 1,
+  "model.layers.3.self_attn.v_proj.weight": 1,
+  "model.layers.4.input_layernorm.weight": 1,
+  "model.layers.4.mlp.down_proj.weight": 1,
+  "model.layers.4.mlp.gate_proj.weight": 1,
+  "model.layers.4.mlp.up_proj.weight": 1,
+  "model.layers.4.post_attention_layernorm.weight": 1,
+  "model.layers.4.self_attn.k_proj.weight": 1,
+  "model.layers.4.self_attn.o_proj.weight": 1,
+  "model.layers.4.self_attn.q_proj.weight": 1,
+  "model.layers.4.self_attn.v_proj.weight": 1,
+  "model.layers.5.input_layernorm.weight": 1,
+  "model.layers.5.mlp.down_proj.weight": 1,
+  "model.layers.5.mlp.gate_proj.weight": 1,
+  "model.layers.5.mlp.up_proj.weight": 1,
+  "model.layers.5.post_attention_layernorm.weight": 1,
+  "model.layers.5.self_attn.k_proj.weight": 1,
+  "model.layers.5.self_attn.o_proj.weight": 1,
+  "model.layers.5.self_attn.q_proj.weight": 1,
+  "model.layers.5.self_attn.v_proj.weight": 1,
+  "model.layers.6.input_layernorm.weight": 1,
+  "model.layers.6.mlp.down_proj.weight": 1,
+  "model.layers.6.mlp.gate_proj.weight": 1,
+  "model.layers.6.mlp.up_proj.weight": 1,
+  "model.layers.6.post_attention_layernorm.weight": 1,
+  "model.layers.6.self_attn.k_proj.weight": 1,
+  "model.layers.6.self_attn.o_proj.weight": 1,
+  "model.layers.6.self_attn.q_proj.weight": 1,
+  "model.layers.6.self_attn.v_proj.weight": 1,
+  "model.layers.7.input_layernorm.weight": 1,
+  "model.layers.7.mlp.down_proj.weight": 1,
+  "model.layers.7.mlp.gate_proj.weight": 1,
+  "model.layers.7.mlp.up_proj.weight": 1,
+  "model.layers.7.post_attention_layernorm.weight": 1,
+  "model.layers.7.self_attn.k_proj.weight": 1,
+  "model.layers.7.self_attn.o_proj.weight": 1,
+  "model.layers.7.self_attn.q_proj.weight": 1,
+  "model.layers.7.self_attn.v_proj.weight": 1,
+  "model.layers.8.input_layernorm.weight": 1,
+  "model.layers.8.mlp.down_proj.weight": 1,
+  "model.layers.8.mlp.gate_proj.weight": 1,
+  "model.layers.8.mlp.up_proj.weight": 1,
+  "model.layers.8.post_attention_layernorm.weight": 1,
+  "model.layers.8.self_attn.k_proj.weight": 1,
+  "model.layers.8.self_attn.o_proj.weight": 1,
+  "model.layers.8.self_attn.q_proj.weight": 1,
+  "model.layers.8.self_attn.v_proj.weight": 1,
+  "model.layers.9.input_layernorm.weight": 1,
+  "model.layers.9.mlp.down_proj.weight": 1,
+  "model.layers.9.mlp.gate_proj.weight": 1,
+  "model.layers.9.mlp.up_proj.weight": 1,
+  "model.layers.9.post_attention_layernorm.weight": 1,
+  "model.layers.9.self_attn.k_proj.weight": 1,
+  "model.layers.9.self_attn.o_proj.weight": 1,
+  "model.layers.9.self_attn.q_proj.weight": 1,
+  "model.layers.9.self_attn.v_proj.weight": 1,
+  "model.norm.weight": 2
+}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/generation_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/generation_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..125cc74759bc705ab8a0c48d07d7eac083f56011
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/policy/weights/model/.hf_metadata/generation_config.json
@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.1"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/training_info.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/training_info.json
new file mode 100644
index 0000000000000000000000000000000000000000..b98e660d2870de6bb42709371e3562228208d487
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_100/training_info.json
@@ -0,0 +1 @@
+{"current_epoch": 19, "current_step": 5, "total_steps": 100, "val_reward": 0.18229166666666666, "consumed_samples": 2000, "total_valid_tokens": 6628299.994122537}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/chat_template.jinja b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..1bad6a0f648dccdbec523ca79ba90fbcfc806af0
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/chat_template.jinja
@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/special_tokens_map.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/special_tokens_map.json
new file mode 100644
index 0000000000000000000000000000000000000000..b43be96621d147110fb8a18b5776ec6e38516127
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/special_tokens_map.json
@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|eot_id|>"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/tokenizer_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..3beeacc86a6ca3cae14ad3004263ab74a4bac07a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/tokenizer/tokenizer_config.json
@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|eot_id|>",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..88a7ab65ecc03929fde705948a65e30260067bc3
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/config.json
@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "float32",
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "nemo_version": "0.2.0",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": 128009,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "vocab_size": 128256
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
new file mode 100644
index 0000000000000000000000000000000000000000..494ffc91e4983233a079912895e9dc90c8c11a79
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
@@ -0,0 +1,256 @@
+{
+  "model.embed_tokens.weight": 1,
+  "model.layers.0.input_layernorm.weight": 1,
+  "model.layers.0.mlp.down_proj.weight": 1,
+  "model.layers.0.mlp.gate_proj.weight": 1,
+  "model.layers.0.mlp.up_proj.weight": 1,
+  "model.layers.0.post_attention_layernorm.weight": 1,
+  "model.layers.0.self_attn.k_proj.weight": 1,
+  "model.layers.0.self_attn.o_proj.weight": 1,
+  "model.layers.0.self_attn.q_proj.weight": 1,
+  "model.layers.0.self_attn.v_proj.weight": 1,
+  "model.layers.1.input_layernorm.weight": 1,
+  "model.layers.1.mlp.down_proj.weight": 1,
+  "model.layers.1.mlp.gate_proj.weight": 1,
+  "model.layers.1.mlp.up_proj.weight": 1,
+  "model.layers.1.post_attention_layernorm.weight": 1,
+  "model.layers.1.self_attn.k_proj.weight": 1,
+  "model.layers.1.self_attn.o_proj.weight": 1,
+  "model.layers.1.self_attn.q_proj.weight": 1,
+  "model.layers.1.self_attn.v_proj.weight": 1,
+  "model.layers.10.input_layernorm.weight": 1,
+  "model.layers.10.mlp.down_proj.weight": 1,
+  "model.layers.10.mlp.gate_proj.weight": 1,
+  "model.layers.10.mlp.up_proj.weight": 1,
+  "model.layers.10.post_attention_layernorm.weight": 1,
+  "model.layers.10.self_attn.k_proj.weight": 1,
+  "model.layers.10.self_attn.o_proj.weight": 1,
+  "model.layers.10.self_attn.q_proj.weight": 1,
+  "model.layers.10.self_attn.v_proj.weight": 1,
+  "model.layers.11.input_layernorm.weight": 1,
+  "model.layers.11.mlp.down_proj.weight": 1,
+  "model.layers.11.mlp.gate_proj.weight": 1,
+  "model.layers.11.mlp.up_proj.weight": 1,
+  "model.layers.11.post_attention_layernorm.weight": 1,
+  "model.layers.11.self_attn.k_proj.weight": 1,
+  "model.layers.11.self_attn.o_proj.weight": 1,
+  "model.layers.11.self_attn.q_proj.weight": 1,
+  "model.layers.11.self_attn.v_proj.weight": 1,
+  "model.layers.12.input_layernorm.weight": 1,
+  "model.layers.12.mlp.down_proj.weight": 1,
+  "model.layers.12.mlp.gate_proj.weight": 1,
+  "model.layers.12.mlp.up_proj.weight": 1,
+  "model.layers.12.post_attention_layernorm.weight": 1,
+  "model.layers.12.self_attn.k_proj.weight": 1,
+  "model.layers.12.self_attn.o_proj.weight": 1,
+  "model.layers.12.self_attn.q_proj.weight": 1,
+  "model.layers.12.self_attn.v_proj.weight": 1,
+  "model.layers.13.input_layernorm.weight": 1,
+  "model.layers.13.mlp.down_proj.weight": 1,
+  "model.layers.13.mlp.gate_proj.weight": 1,
+  "model.layers.13.mlp.up_proj.weight": 1,
+  "model.layers.13.post_attention_layernorm.weight": 1,
+  "model.layers.13.self_attn.k_proj.weight": 1,
+  "model.layers.13.self_attn.o_proj.weight": 1,
+  "model.layers.13.self_attn.q_proj.weight": 1,
+  "model.layers.13.self_attn.v_proj.weight": 1,
+  "model.layers.14.input_layernorm.weight": 1,
+  "model.layers.14.mlp.down_proj.weight": 1,
+  "model.layers.14.mlp.gate_proj.weight": 1,
+  "model.layers.14.mlp.up_proj.weight": 1,
+  "model.layers.14.post_attention_layernorm.weight": 1,
+  "model.layers.14.self_attn.k_proj.weight": 1,
+  "model.layers.14.self_attn.o_proj.weight": 1,
+  "model.layers.14.self_attn.q_proj.weight": 1,
+  "model.layers.14.self_attn.v_proj.weight": 1,
+  "model.layers.15.input_layernorm.weight": 1,
+  "model.layers.15.mlp.down_proj.weight": 1,
+  "model.layers.15.mlp.gate_proj.weight": 1,
+  "model.layers.15.mlp.up_proj.weight": 1,
+  "model.layers.15.post_attention_layernorm.weight": 1,
+  "model.layers.15.self_attn.k_proj.weight": 1,
+  "model.layers.15.self_attn.o_proj.weight": 1,
+  "model.layers.15.self_attn.q_proj.weight": 1,
+  "model.layers.15.self_attn.v_proj.weight": 1,
+  "model.layers.16.input_layernorm.weight": 1,
+  "model.layers.16.mlp.down_proj.weight": 1,
+  "model.layers.16.mlp.gate_proj.weight": 1,
+  "model.layers.16.mlp.up_proj.weight": 1,
+  "model.layers.16.post_attention_layernorm.weight": 1,
+  "model.layers.16.self_attn.k_proj.weight": 1,
+  "model.layers.16.self_attn.o_proj.weight": 1,
+  "model.layers.16.self_attn.q_proj.weight": 1,
+  "model.layers.16.self_attn.v_proj.weight": 1,
+  "model.layers.17.input_layernorm.weight": 1,
+  "model.layers.17.mlp.down_proj.weight": 1,
+  "model.layers.17.mlp.gate_proj.weight": 1,
+  "model.layers.17.mlp.up_proj.weight": 1,
+  "model.layers.17.post_attention_layernorm.weight": 1,
+  "model.layers.17.self_attn.k_proj.weight": 1,
+  "model.layers.17.self_attn.o_proj.weight": 1,
+  "model.layers.17.self_attn.q_proj.weight": 1,
+  "model.layers.17.self_attn.v_proj.weight": 1,
+  "model.layers.18.input_layernorm.weight": 1,
+  "model.layers.18.mlp.down_proj.weight": 1,
+  "model.layers.18.mlp.gate_proj.weight": 1,
+  "model.layers.18.mlp.up_proj.weight": 1,
+  "model.layers.18.post_attention_layernorm.weight": 1,
+  "model.layers.18.self_attn.k_proj.weight": 1,
+  "model.layers.18.self_attn.o_proj.weight": 1,
+  "model.layers.18.self_attn.q_proj.weight": 1,
+  "model.layers.18.self_attn.v_proj.weight": 1,
+  "model.layers.19.input_layernorm.weight": 1,
+  "model.layers.19.mlp.down_proj.weight": 1,
+  "model.layers.19.mlp.gate_proj.weight": 1,
+  "model.layers.19.mlp.up_proj.weight": 1,
+  "model.layers.19.post_attention_layernorm.weight": 1,
+  "model.layers.19.self_attn.k_proj.weight": 1,
+  "model.layers.19.self_attn.o_proj.weight": 1,
+  "model.layers.19.self_attn.q_proj.weight": 1,
+  "model.layers.19.self_attn.v_proj.weight": 1,
+  "model.layers.2.input_layernorm.weight": 1,
+  "model.layers.2.mlp.down_proj.weight": 1,
+  "model.layers.2.mlp.gate_proj.weight": 1,
+  "model.layers.2.mlp.up_proj.weight": 1,
+  "model.layers.2.post_attention_layernorm.weight": 1,
+  "model.layers.2.self_attn.k_proj.weight": 1,
+  "model.layers.2.self_attn.o_proj.weight": 1,
+  "model.layers.2.self_attn.q_proj.weight": 1,
+  "model.layers.2.self_attn.v_proj.weight": 1,
+  "model.layers.20.input_layernorm.weight": 2,
+  "model.layers.20.mlp.down_proj.weight": 2,
+  "model.layers.20.mlp.gate_proj.weight": 1,
+  "model.layers.20.mlp.up_proj.weight": 1,
+  "model.layers.20.post_attention_layernorm.weight": 2,
+  "model.layers.20.self_attn.k_proj.weight": 1,
+  "model.layers.20.self_attn.o_proj.weight": 1,
+  "model.layers.20.self_attn.q_proj.weight": 1,
+  "model.layers.20.self_attn.v_proj.weight": 1,
+  "model.layers.21.input_layernorm.weight": 2,
+  "model.layers.21.mlp.down_proj.weight": 2,
+  "model.layers.21.mlp.gate_proj.weight": 2,
+  "model.layers.21.mlp.up_proj.weight": 2,
+  "model.layers.21.post_attention_layernorm.weight": 2,
+  "model.layers.21.self_attn.k_proj.weight": 2,
+  "model.layers.21.self_attn.o_proj.weight": 2,
+  "model.layers.21.self_attn.q_proj.weight": 2,
+  "model.layers.21.self_attn.v_proj.weight": 2,
+  "model.layers.22.input_layernorm.weight": 2,
+  "model.layers.22.mlp.down_proj.weight": 2,
+  "model.layers.22.mlp.gate_proj.weight": 2,
+  "model.layers.22.mlp.up_proj.weight": 2,
+  "model.layers.22.post_attention_layernorm.weight": 2,
+  "model.layers.22.self_attn.k_proj.weight": 2,
+  "model.layers.22.self_attn.o_proj.weight": 2,
+  "model.layers.22.self_attn.q_proj.weight": 2,
+  "model.layers.22.self_attn.v_proj.weight": 2,
+  "model.layers.23.input_layernorm.weight": 2,
+  "model.layers.23.mlp.down_proj.weight": 2,
+  "model.layers.23.mlp.gate_proj.weight": 2,
+  "model.layers.23.mlp.up_proj.weight": 2,
+  "model.layers.23.post_attention_layernorm.weight": 2,
+  "model.layers.23.self_attn.k_proj.weight": 2,
+  "model.layers.23.self_attn.o_proj.weight": 2,
+  "model.layers.23.self_attn.q_proj.weight": 2,
+  "model.layers.23.self_attn.v_proj.weight": 2,
+  "model.layers.24.input_layernorm.weight": 2,
+  "model.layers.24.mlp.down_proj.weight": 2,
+  "model.layers.24.mlp.gate_proj.weight": 2,
+  "model.layers.24.mlp.up_proj.weight": 2,
+  "model.layers.24.post_attention_layernorm.weight": 2,
+  "model.layers.24.self_attn.k_proj.weight": 2,
+  "model.layers.24.self_attn.o_proj.weight": 2,
+  "model.layers.24.self_attn.q_proj.weight": 2,
+  "model.layers.24.self_attn.v_proj.weight": 2,
+  "model.layers.25.input_layernorm.weight": 2,
+  "model.layers.25.mlp.down_proj.weight": 2,
+  "model.layers.25.mlp.gate_proj.weight": 2,
+  "model.layers.25.mlp.up_proj.weight": 2,
+  "model.layers.25.post_attention_layernorm.weight": 2,
+  "model.layers.25.self_attn.k_proj.weight": 2,
+  "model.layers.25.self_attn.o_proj.weight": 2,
+  "model.layers.25.self_attn.q_proj.weight": 2,
+  "model.layers.25.self_attn.v_proj.weight": 2,
+  "model.layers.26.input_layernorm.weight": 2,
+  "model.layers.26.mlp.down_proj.weight": 2,
+  "model.layers.26.mlp.gate_proj.weight": 2,
+  "model.layers.26.mlp.up_proj.weight": 2,
+  "model.layers.26.post_attention_layernorm.weight": 2,
+  "model.layers.26.self_attn.k_proj.weight": 2,
+  "model.layers.26.self_attn.o_proj.weight": 2,
+  "model.layers.26.self_attn.q_proj.weight": 2,
+  "model.layers.26.self_attn.v_proj.weight": 2,
+  "model.layers.27.input_layernorm.weight": 2,
+  "model.layers.27.mlp.down_proj.weight": 2,
+  "model.layers.27.mlp.gate_proj.weight": 2,
+  "model.layers.27.mlp.up_proj.weight": 2,
+  "model.layers.27.post_attention_layernorm.weight": 2,
+  "model.layers.27.self_attn.k_proj.weight": 2,
+  "model.layers.27.self_attn.o_proj.weight": 2,
+  "model.layers.27.self_attn.q_proj.weight": 2,
+  "model.layers.27.self_attn.v_proj.weight": 2,
+  "model.layers.3.input_layernorm.weight": 1,
+  "model.layers.3.mlp.down_proj.weight": 1,
+  "model.layers.3.mlp.gate_proj.weight": 1,
+  "model.layers.3.mlp.up_proj.weight": 1,
+  "model.layers.3.post_attention_layernorm.weight": 1,
+  "model.layers.3.self_attn.k_proj.weight": 1,
+  "model.layers.3.self_attn.o_proj.weight": 1,
+  "model.layers.3.self_attn.q_proj.weight": 1,
+  "model.layers.3.self_attn.v_proj.weight": 1,
+  "model.layers.4.input_layernorm.weight": 1,
+  "model.layers.4.mlp.down_proj.weight": 1,
+  "model.layers.4.mlp.gate_proj.weight": 1,
+  "model.layers.4.mlp.up_proj.weight": 1,
+  "model.layers.4.post_attention_layernorm.weight": 1,
+  "model.layers.4.self_attn.k_proj.weight": 1,
+  "model.layers.4.self_attn.o_proj.weight": 1,
+  "model.layers.4.self_attn.q_proj.weight": 1,
+  "model.layers.4.self_attn.v_proj.weight": 1,
+  "model.layers.5.input_layernorm.weight": 1,
+  "model.layers.5.mlp.down_proj.weight": 1,
+  "model.layers.5.mlp.gate_proj.weight": 1,
+  "model.layers.5.mlp.up_proj.weight": 1,
+  "model.layers.5.post_attention_layernorm.weight": 1,
+  "model.layers.5.self_attn.k_proj.weight": 1,
+  "model.layers.5.self_attn.o_proj.weight": 1,
+  "model.layers.5.self_attn.q_proj.weight": 1,
+  "model.layers.5.self_attn.v_proj.weight": 1,
+  "model.layers.6.input_layernorm.weight": 1,
+  "model.layers.6.mlp.down_proj.weight": 1,
+  "model.layers.6.mlp.gate_proj.weight": 1,
+  "model.layers.6.mlp.up_proj.weight": 1,
+  "model.layers.6.post_attention_layernorm.weight": 1,
+  "model.layers.6.self_attn.k_proj.weight": 1,
+  "model.layers.6.self_attn.o_proj.weight": 1,
+  "model.layers.6.self_attn.q_proj.weight": 1,
+  "model.layers.6.self_attn.v_proj.weight": 1,
+  "model.layers.7.input_layernorm.weight": 1,
+  "model.layers.7.mlp.down_proj.weight": 1,
+  "model.layers.7.mlp.gate_proj.weight": 1,
+  "model.layers.7.mlp.up_proj.weight": 1,
+  "model.layers.7.post_attention_layernorm.weight": 1,
+  "model.layers.7.self_attn.k_proj.weight": 1,
+  "model.layers.7.self_attn.o_proj.weight": 1,
+  "model.layers.7.self_attn.q_proj.weight": 1,
+  "model.layers.7.self_attn.v_proj.weight": 1,
+  "model.layers.8.input_layernorm.weight": 1,
+  "model.layers.8.mlp.down_proj.weight": 1,
+  "model.layers.8.mlp.gate_proj.weight": 1,
+  "model.layers.8.mlp.up_proj.weight": 1,
+  "model.layers.8.post_attention_layernorm.weight": 1,
+  "model.layers.8.self_attn.k_proj.weight": 1,
+  "model.layers.8.self_attn.o_proj.weight": 1,
+  "model.layers.8.self_attn.q_proj.weight": 1,
+  "model.layers.8.self_attn.v_proj.weight": 1,
+  "model.layers.9.input_layernorm.weight": 1,
+  "model.layers.9.mlp.down_proj.weight": 1,
+  "model.layers.9.mlp.gate_proj.weight": 1,
+  "model.layers.9.mlp.up_proj.weight": 1,
+  "model.layers.9.post_attention_layernorm.weight": 1,
+  "model.layers.9.self_attn.k_proj.weight": 1,
+  "model.layers.9.self_attn.o_proj.weight": 1,
+  "model.layers.9.self_attn.q_proj.weight": 1,
+  "model.layers.9.self_attn.v_proj.weight": 1,
+  "model.norm.weight": 2
+}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/generation_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/generation_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..125cc74759bc705ab8a0c48d07d7eac083f56011
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/policy/weights/model/.hf_metadata/generation_config.json
@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.1"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/training_info.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/training_info.json
new file mode 100644
index 0000000000000000000000000000000000000000..1f5d2a61ecabf263b19bc48144c566912750f6d5
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_120/training_info.json
@@ -0,0 +1 @@
+{"current_epoch": 23, "current_step": 5, "total_steps": 120, "val_reward": 0.24479166666666666, "consumed_samples": 2400, "total_valid_tokens": 7602486.356696717}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/chat_template.jinja b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..1bad6a0f648dccdbec523ca79ba90fbcfc806af0
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/chat_template.jinja
@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/special_tokens_map.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/special_tokens_map.json
new file mode 100644
index 0000000000000000000000000000000000000000..b43be96621d147110fb8a18b5776ec6e38516127
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/special_tokens_map.json
@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|eot_id|>"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/tokenizer_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..3beeacc86a6ca3cae14ad3004263ab74a4bac07a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/tokenizer/tokenizer_config.json
@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|eot_id|>",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..88a7ab65ecc03929fde705948a65e30260067bc3
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/config.json
@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "float32",
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "nemo_version": "0.2.0",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": 128009,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "vocab_size": 128256
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
new file mode 100644
index 0000000000000000000000000000000000000000..494ffc91e4983233a079912895e9dc90c8c11a79
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
@@ -0,0 +1,256 @@
+{
+  "model.embed_tokens.weight": 1,
+  "model.layers.0.input_layernorm.weight": 1,
+  "model.layers.0.mlp.down_proj.weight": 1,
+  "model.layers.0.mlp.gate_proj.weight": 1,
+  "model.layers.0.mlp.up_proj.weight": 1,
+  "model.layers.0.post_attention_layernorm.weight": 1,
+  "model.layers.0.self_attn.k_proj.weight": 1,
+  "model.layers.0.self_attn.o_proj.weight": 1,
+  "model.layers.0.self_attn.q_proj.weight": 1,
+  "model.layers.0.self_attn.v_proj.weight": 1,
+  "model.layers.1.input_layernorm.weight": 1,
+  "model.layers.1.mlp.down_proj.weight": 1,
+  "model.layers.1.mlp.gate_proj.weight": 1,
+  "model.layers.1.mlp.up_proj.weight": 1,
+  "model.layers.1.post_attention_layernorm.weight": 1,
+  "model.layers.1.self_attn.k_proj.weight": 1,
+  "model.layers.1.self_attn.o_proj.weight": 1,
+  "model.layers.1.self_attn.q_proj.weight": 1,
+  "model.layers.1.self_attn.v_proj.weight": 1,
+  "model.layers.10.input_layernorm.weight": 1,
+  "model.layers.10.mlp.down_proj.weight": 1,
+  "model.layers.10.mlp.gate_proj.weight": 1,
+  "model.layers.10.mlp.up_proj.weight": 1,
+  "model.layers.10.post_attention_layernorm.weight": 1,
+  "model.layers.10.self_attn.k_proj.weight": 1,
+  "model.layers.10.self_attn.o_proj.weight": 1,
+  "model.layers.10.self_attn.q_proj.weight": 1,
+  "model.layers.10.self_attn.v_proj.weight": 1,
+  "model.layers.11.input_layernorm.weight": 1,
+  "model.layers.11.mlp.down_proj.weight": 1,
+  "model.layers.11.mlp.gate_proj.weight": 1,
+  "model.layers.11.mlp.up_proj.weight": 1,
+  "model.layers.11.post_attention_layernorm.weight": 1,
+  "model.layers.11.self_attn.k_proj.weight": 1,
+  "model.layers.11.self_attn.o_proj.weight": 1,
+  "model.layers.11.self_attn.q_proj.weight": 1,
+  "model.layers.11.self_attn.v_proj.weight": 1,
+  "model.layers.12.input_layernorm.weight": 1,
+  "model.layers.12.mlp.down_proj.weight": 1,
+  "model.layers.12.mlp.gate_proj.weight": 1,
+  "model.layers.12.mlp.up_proj.weight": 1,
+  "model.layers.12.post_attention_layernorm.weight": 1,
+  "model.layers.12.self_attn.k_proj.weight": 1,
+  "model.layers.12.self_attn.o_proj.weight": 1,
+  "model.layers.12.self_attn.q_proj.weight": 1,
+  "model.layers.12.self_attn.v_proj.weight": 1,
+  "model.layers.13.input_layernorm.weight": 1,
+  "model.layers.13.mlp.down_proj.weight": 1,
+  "model.layers.13.mlp.gate_proj.weight": 1,
+  "model.layers.13.mlp.up_proj.weight": 1,
+  "model.layers.13.post_attention_layernorm.weight": 1,
+  "model.layers.13.self_attn.k_proj.weight": 1,
+  "model.layers.13.self_attn.o_proj.weight": 1,
+  "model.layers.13.self_attn.q_proj.weight": 1,
+  "model.layers.13.self_attn.v_proj.weight": 1,
+  "model.layers.14.input_layernorm.weight": 1,
+  "model.layers.14.mlp.down_proj.weight": 1,
+  "model.layers.14.mlp.gate_proj.weight": 1,
+  "model.layers.14.mlp.up_proj.weight": 1,
+  "model.layers.14.post_attention_layernorm.weight": 1,
+  "model.layers.14.self_attn.k_proj.weight": 1,
+  "model.layers.14.self_attn.o_proj.weight": 1,
+  "model.layers.14.self_attn.q_proj.weight": 1,
+  "model.layers.14.self_attn.v_proj.weight": 1,
+  "model.layers.15.input_layernorm.weight": 1,
+  "model.layers.15.mlp.down_proj.weight": 1,
+  "model.layers.15.mlp.gate_proj.weight": 1,
+  "model.layers.15.mlp.up_proj.weight": 1,
+  "model.layers.15.post_attention_layernorm.weight": 1,
+  "model.layers.15.self_attn.k_proj.weight": 1,
+  "model.layers.15.self_attn.o_proj.weight": 1,
+  "model.layers.15.self_attn.q_proj.weight": 1,
+  "model.layers.15.self_attn.v_proj.weight": 1,
+  "model.layers.16.input_layernorm.weight": 1,
+  "model.layers.16.mlp.down_proj.weight": 1,
+  "model.layers.16.mlp.gate_proj.weight": 1,
+  "model.layers.16.mlp.up_proj.weight": 1,
+  "model.layers.16.post_attention_layernorm.weight": 1,
+  "model.layers.16.self_attn.k_proj.weight": 1,
+  "model.layers.16.self_attn.o_proj.weight": 1,
+  "model.layers.16.self_attn.q_proj.weight": 1,
+  "model.layers.16.self_attn.v_proj.weight": 1,
+  "model.layers.17.input_layernorm.weight": 1,
+  "model.layers.17.mlp.down_proj.weight": 1,
+  "model.layers.17.mlp.gate_proj.weight": 1,
+  "model.layers.17.mlp.up_proj.weight": 1,
+  "model.layers.17.post_attention_layernorm.weight": 1,
+  "model.layers.17.self_attn.k_proj.weight": 1,
+  "model.layers.17.self_attn.o_proj.weight": 1,
+  "model.layers.17.self_attn.q_proj.weight": 1,
+  "model.layers.17.self_attn.v_proj.weight": 1,
+  "model.layers.18.input_layernorm.weight": 1,
+  "model.layers.18.mlp.down_proj.weight": 1,
+  "model.layers.18.mlp.gate_proj.weight": 1,
+  "model.layers.18.mlp.up_proj.weight": 1,
+  "model.layers.18.post_attention_layernorm.weight": 1,
+  "model.layers.18.self_attn.k_proj.weight": 1,
+  "model.layers.18.self_attn.o_proj.weight": 1,
+  "model.layers.18.self_attn.q_proj.weight": 1,
+  "model.layers.18.self_attn.v_proj.weight": 1,
+  "model.layers.19.input_layernorm.weight": 1,
+  "model.layers.19.mlp.down_proj.weight": 1,
+  "model.layers.19.mlp.gate_proj.weight": 1,
+  "model.layers.19.mlp.up_proj.weight": 1,
+  "model.layers.19.post_attention_layernorm.weight": 1,
+  "model.layers.19.self_attn.k_proj.weight": 1,
+  "model.layers.19.self_attn.o_proj.weight": 1,
+  "model.layers.19.self_attn.q_proj.weight": 1,
+  "model.layers.19.self_attn.v_proj.weight": 1,
+  "model.layers.2.input_layernorm.weight": 1,
+  "model.layers.2.mlp.down_proj.weight": 1,
+  "model.layers.2.mlp.gate_proj.weight": 1,
+  "model.layers.2.mlp.up_proj.weight": 1,
+  "model.layers.2.post_attention_layernorm.weight": 1,
+  "model.layers.2.self_attn.k_proj.weight": 1,
+  "model.layers.2.self_attn.o_proj.weight": 1,
+  "model.layers.2.self_attn.q_proj.weight": 1,
+  "model.layers.2.self_attn.v_proj.weight": 1,
+  "model.layers.20.input_layernorm.weight": 2,
+  "model.layers.20.mlp.down_proj.weight": 2,
+  "model.layers.20.mlp.gate_proj.weight": 1,
+  "model.layers.20.mlp.up_proj.weight": 1,
+  "model.layers.20.post_attention_layernorm.weight": 2,
+  "model.layers.20.self_attn.k_proj.weight": 1,
+  "model.layers.20.self_attn.o_proj.weight": 1,
+  "model.layers.20.self_attn.q_proj.weight": 1,
+  "model.layers.20.self_attn.v_proj.weight": 1,
+  "model.layers.21.input_layernorm.weight": 2,
+  "model.layers.21.mlp.down_proj.weight": 2,
+  "model.layers.21.mlp.gate_proj.weight": 2,
+  "model.layers.21.mlp.up_proj.weight": 2,
+  "model.layers.21.post_attention_layernorm.weight": 2,
+  "model.layers.21.self_attn.k_proj.weight": 2,
+  "model.layers.21.self_attn.o_proj.weight": 2,
+  "model.layers.21.self_attn.q_proj.weight": 2,
+  "model.layers.21.self_attn.v_proj.weight": 2,
+  "model.layers.22.input_layernorm.weight": 2,
+  "model.layers.22.mlp.down_proj.weight": 2,
+  "model.layers.22.mlp.gate_proj.weight": 2,
+  "model.layers.22.mlp.up_proj.weight": 2,
+  "model.layers.22.post_attention_layernorm.weight": 2,
+  "model.layers.22.self_attn.k_proj.weight": 2,
+  "model.layers.22.self_attn.o_proj.weight": 2,
+  "model.layers.22.self_attn.q_proj.weight": 2,
+  "model.layers.22.self_attn.v_proj.weight": 2,
+  "model.layers.23.input_layernorm.weight": 2,
+  "model.layers.23.mlp.down_proj.weight": 2,
+  "model.layers.23.mlp.gate_proj.weight": 2,
+  "model.layers.23.mlp.up_proj.weight": 2,
+  "model.layers.23.post_attention_layernorm.weight": 2,
+  "model.layers.23.self_attn.k_proj.weight": 2,
+  "model.layers.23.self_attn.o_proj.weight": 2,
+  "model.layers.23.self_attn.q_proj.weight": 2,
+  "model.layers.23.self_attn.v_proj.weight": 2,
+  "model.layers.24.input_layernorm.weight": 2,
+  "model.layers.24.mlp.down_proj.weight": 2,
+  "model.layers.24.mlp.gate_proj.weight": 2,
+  "model.layers.24.mlp.up_proj.weight": 2,
+  "model.layers.24.post_attention_layernorm.weight": 2,
+  "model.layers.24.self_attn.k_proj.weight": 2,
+  "model.layers.24.self_attn.o_proj.weight": 2,
+  "model.layers.24.self_attn.q_proj.weight": 2,
+  "model.layers.24.self_attn.v_proj.weight": 2,
+  "model.layers.25.input_layernorm.weight": 2,
+  "model.layers.25.mlp.down_proj.weight": 2,
+  "model.layers.25.mlp.gate_proj.weight": 2,
+  "model.layers.25.mlp.up_proj.weight": 2,
+  "model.layers.25.post_attention_layernorm.weight": 2,
+  "model.layers.25.self_attn.k_proj.weight": 2,
+  "model.layers.25.self_attn.o_proj.weight": 2,
+  "model.layers.25.self_attn.q_proj.weight": 2,
+  "model.layers.25.self_attn.v_proj.weight": 2,
+  "model.layers.26.input_layernorm.weight": 2,
+  "model.layers.26.mlp.down_proj.weight": 2,
+  "model.layers.26.mlp.gate_proj.weight": 2,
+  "model.layers.26.mlp.up_proj.weight": 2,
+  "model.layers.26.post_attention_layernorm.weight": 2,
+  "model.layers.26.self_attn.k_proj.weight": 2,
+  "model.layers.26.self_attn.o_proj.weight": 2,
+  "model.layers.26.self_attn.q_proj.weight": 2,
+  "model.layers.26.self_attn.v_proj.weight": 2,
+  "model.layers.27.input_layernorm.weight": 2,
+  "model.layers.27.mlp.down_proj.weight": 2,
+  "model.layers.27.mlp.gate_proj.weight": 2,
+  "model.layers.27.mlp.up_proj.weight": 2,
+  "model.layers.27.post_attention_layernorm.weight": 2,
+  "model.layers.27.self_attn.k_proj.weight": 2,
+  "model.layers.27.self_attn.o_proj.weight": 2,
+  "model.layers.27.self_attn.q_proj.weight": 2,
+  "model.layers.27.self_attn.v_proj.weight": 2,
+  "model.layers.3.input_layernorm.weight": 1,
+  "model.layers.3.mlp.down_proj.weight": 1,
+  "model.layers.3.mlp.gate_proj.weight": 1,
+  "model.layers.3.mlp.up_proj.weight": 1,
+  "model.layers.3.post_attention_layernorm.weight": 1,
+  "model.layers.3.self_attn.k_proj.weight": 1,
+  "model.layers.3.self_attn.o_proj.weight": 1,
+  "model.layers.3.self_attn.q_proj.weight": 1,
+  "model.layers.3.self_attn.v_proj.weight": 1,
+  "model.layers.4.input_layernorm.weight": 1,
+  "model.layers.4.mlp.down_proj.weight": 1,
+  "model.layers.4.mlp.gate_proj.weight": 1,
+  "model.layers.4.mlp.up_proj.weight": 1,
+  "model.layers.4.post_attention_layernorm.weight": 1,
+  "model.layers.4.self_attn.k_proj.weight": 1,
+  "model.layers.4.self_attn.o_proj.weight": 1,
+  "model.layers.4.self_attn.q_proj.weight": 1,
+  "model.layers.4.self_attn.v_proj.weight": 1,
+  "model.layers.5.input_layernorm.weight": 1,
+  "model.layers.5.mlp.down_proj.weight": 1,
+  "model.layers.5.mlp.gate_proj.weight": 1,
+  "model.layers.5.mlp.up_proj.weight": 1,
+  "model.layers.5.post_attention_layernorm.weight": 1,
+  "model.layers.5.self_attn.k_proj.weight": 1,
+  "model.layers.5.self_attn.o_proj.weight": 1,
+  "model.layers.5.self_attn.q_proj.weight": 1,
+  "model.layers.5.self_attn.v_proj.weight": 1,
+  "model.layers.6.input_layernorm.weight": 1,
+  "model.layers.6.mlp.down_proj.weight": 1,
+  "model.layers.6.mlp.gate_proj.weight": 1,
+  "model.layers.6.mlp.up_proj.weight": 1,
+  "model.layers.6.post_attention_layernorm.weight": 1,
+  "model.layers.6.self_attn.k_proj.weight": 1,
+  "model.layers.6.self_attn.o_proj.weight": 1,
+  "model.layers.6.self_attn.q_proj.weight": 1,
+  "model.layers.6.self_attn.v_proj.weight": 1,
+  "model.layers.7.input_layernorm.weight": 1,
+  "model.layers.7.mlp.down_proj.weight": 1,
+  "model.layers.7.mlp.gate_proj.weight": 1,
+  "model.layers.7.mlp.up_proj.weight": 1,
+  "model.layers.7.post_attention_layernorm.weight": 1,
+  "model.layers.7.self_attn.k_proj.weight": 1,
+  "model.layers.7.self_attn.o_proj.weight": 1,
+  "model.layers.7.self_attn.q_proj.weight": 1,
+  "model.layers.7.self_attn.v_proj.weight": 1,
+  "model.layers.8.input_layernorm.weight": 1,
+  "model.layers.8.mlp.down_proj.weight": 1,
+  "model.layers.8.mlp.gate_proj.weight": 1,
+  "model.layers.8.mlp.up_proj.weight": 1,
+  "model.layers.8.post_attention_layernorm.weight": 1,
+  "model.layers.8.self_attn.k_proj.weight": 1,
+  "model.layers.8.self_attn.o_proj.weight": 1,
+  "model.layers.8.self_attn.q_proj.weight": 1,
+  "model.layers.8.self_attn.v_proj.weight": 1,
+  "model.layers.9.input_layernorm.weight": 1,
+  "model.layers.9.mlp.down_proj.weight": 1,
+  "model.layers.9.mlp.gate_proj.weight": 1,
+  "model.layers.9.mlp.up_proj.weight": 1,
+  "model.layers.9.post_attention_layernorm.weight": 1,
+  "model.layers.9.self_attn.k_proj.weight": 1,
+  "model.layers.9.self_attn.o_proj.weight": 1,
+  "model.layers.9.self_attn.q_proj.weight": 1,
+  "model.layers.9.self_attn.v_proj.weight": 1,
+  "model.norm.weight": 2
+}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/generation_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/generation_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..125cc74759bc705ab8a0c48d07d7eac083f56011
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/policy/weights/model/.hf_metadata/generation_config.json
@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.1"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/training_info.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/training_info.json
new file mode 100644
index 0000000000000000000000000000000000000000..bea84cc7610be8361c585b1fa108239f501880ea
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_140/training_info.json
@@ -0,0 +1 @@
+{"current_epoch": 27, "current_step": 5, "total_steps": 140, "val_reward": 0.21354166666666666, "consumed_samples": 2800, "total_valid_tokens": 8562606.381899573}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/chat_template.jinja b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..1bad6a0f648dccdbec523ca79ba90fbcfc806af0
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/chat_template.jinja
@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/special_tokens_map.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/special_tokens_map.json
new file mode 100644
index 0000000000000000000000000000000000000000..b43be96621d147110fb8a18b5776ec6e38516127
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/special_tokens_map.json
@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|eot_id|>"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/tokenizer_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..3beeacc86a6ca3cae14ad3004263ab74a4bac07a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/tokenizer/tokenizer_config.json
@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|eot_id|>",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..88a7ab65ecc03929fde705948a65e30260067bc3
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/config.json
@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "float32",
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "nemo_version": "0.2.0",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": 128009,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "vocab_size": 128256
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
new file mode 100644
index 0000000000000000000000000000000000000000..494ffc91e4983233a079912895e9dc90c8c11a79
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
@@ -0,0 +1,256 @@
+{
+  "model.embed_tokens.weight": 1,
+  "model.layers.0.input_layernorm.weight": 1,
+  "model.layers.0.mlp.down_proj.weight": 1,
+  "model.layers.0.mlp.gate_proj.weight": 1,
+  "model.layers.0.mlp.up_proj.weight": 1,
+  "model.layers.0.post_attention_layernorm.weight": 1,
+  "model.layers.0.self_attn.k_proj.weight": 1,
+  "model.layers.0.self_attn.o_proj.weight": 1,
+  "model.layers.0.self_attn.q_proj.weight": 1,
+  "model.layers.0.self_attn.v_proj.weight": 1,
+  "model.layers.1.input_layernorm.weight": 1,
+  "model.layers.1.mlp.down_proj.weight": 1,
+  "model.layers.1.mlp.gate_proj.weight": 1,
+  "model.layers.1.mlp.up_proj.weight": 1,
+  "model.layers.1.post_attention_layernorm.weight": 1,
+  "model.layers.1.self_attn.k_proj.weight": 1,
+  "model.layers.1.self_attn.o_proj.weight": 1,
+  "model.layers.1.self_attn.q_proj.weight": 1,
+  "model.layers.1.self_attn.v_proj.weight": 1,
+  "model.layers.10.input_layernorm.weight": 1,
+  "model.layers.10.mlp.down_proj.weight": 1,
+  "model.layers.10.mlp.gate_proj.weight": 1,
+  "model.layers.10.mlp.up_proj.weight": 1,
+  "model.layers.10.post_attention_layernorm.weight": 1,
+  "model.layers.10.self_attn.k_proj.weight": 1,
+  "model.layers.10.self_attn.o_proj.weight": 1,
+  "model.layers.10.self_attn.q_proj.weight": 1,
+  "model.layers.10.self_attn.v_proj.weight": 1,
+  "model.layers.11.input_layernorm.weight": 1,
+  "model.layers.11.mlp.down_proj.weight": 1,
+  "model.layers.11.mlp.gate_proj.weight": 1,
+  "model.layers.11.mlp.up_proj.weight": 1,
+  "model.layers.11.post_attention_layernorm.weight": 1,
+  "model.layers.11.self_attn.k_proj.weight": 1,
+  "model.layers.11.self_attn.o_proj.weight": 1,
+  "model.layers.11.self_attn.q_proj.weight": 1,
+  "model.layers.11.self_attn.v_proj.weight": 1,
+  "model.layers.12.input_layernorm.weight": 1,
+  "model.layers.12.mlp.down_proj.weight": 1,
+  "model.layers.12.mlp.gate_proj.weight": 1,
+  "model.layers.12.mlp.up_proj.weight": 1,
+  "model.layers.12.post_attention_layernorm.weight": 1,
+  "model.layers.12.self_attn.k_proj.weight": 1,
+  "model.layers.12.self_attn.o_proj.weight": 1,
+  "model.layers.12.self_attn.q_proj.weight": 1,
+  "model.layers.12.self_attn.v_proj.weight": 1,
+  "model.layers.13.input_layernorm.weight": 1,
+  "model.layers.13.mlp.down_proj.weight": 1,
+  "model.layers.13.mlp.gate_proj.weight": 1,
+  "model.layers.13.mlp.up_proj.weight": 1,
+  "model.layers.13.post_attention_layernorm.weight": 1,
+  "model.layers.13.self_attn.k_proj.weight": 1,
+  "model.layers.13.self_attn.o_proj.weight": 1,
+  "model.layers.13.self_attn.q_proj.weight": 1,
+  "model.layers.13.self_attn.v_proj.weight": 1,
+  "model.layers.14.input_layernorm.weight": 1,
+  "model.layers.14.mlp.down_proj.weight": 1,
+  "model.layers.14.mlp.gate_proj.weight": 1,
+  "model.layers.14.mlp.up_proj.weight": 1,
+  "model.layers.14.post_attention_layernorm.weight": 1,
+  "model.layers.14.self_attn.k_proj.weight": 1,
+  "model.layers.14.self_attn.o_proj.weight": 1,
+  "model.layers.14.self_attn.q_proj.weight": 1,
+  "model.layers.14.self_attn.v_proj.weight": 1,
+  "model.layers.15.input_layernorm.weight": 1,
+  "model.layers.15.mlp.down_proj.weight": 1,
+  "model.layers.15.mlp.gate_proj.weight": 1,
+  "model.layers.15.mlp.up_proj.weight": 1,
+  "model.layers.15.post_attention_layernorm.weight": 1,
+  "model.layers.15.self_attn.k_proj.weight": 1,
+  "model.layers.15.self_attn.o_proj.weight": 1,
+  "model.layers.15.self_attn.q_proj.weight": 1,
+  "model.layers.15.self_attn.v_proj.weight": 1,
+  "model.layers.16.input_layernorm.weight": 1,
+  "model.layers.16.mlp.down_proj.weight": 1,
+  "model.layers.16.mlp.gate_proj.weight": 1,
+  "model.layers.16.mlp.up_proj.weight": 1,
+  "model.layers.16.post_attention_layernorm.weight": 1,
+  "model.layers.16.self_attn.k_proj.weight": 1,
+  "model.layers.16.self_attn.o_proj.weight": 1,
+  "model.layers.16.self_attn.q_proj.weight": 1,
+  "model.layers.16.self_attn.v_proj.weight": 1,
+  "model.layers.17.input_layernorm.weight": 1,
+  "model.layers.17.mlp.down_proj.weight": 1,
+  "model.layers.17.mlp.gate_proj.weight": 1,
+  "model.layers.17.mlp.up_proj.weight": 1,
+  "model.layers.17.post_attention_layernorm.weight": 1,
+  "model.layers.17.self_attn.k_proj.weight": 1,
+  "model.layers.17.self_attn.o_proj.weight": 1,
+  "model.layers.17.self_attn.q_proj.weight": 1,
+  "model.layers.17.self_attn.v_proj.weight": 1,
+  "model.layers.18.input_layernorm.weight": 1,
+  "model.layers.18.mlp.down_proj.weight": 1,
+  "model.layers.18.mlp.gate_proj.weight": 1,
+  "model.layers.18.mlp.up_proj.weight": 1,
+  "model.layers.18.post_attention_layernorm.weight": 1,
+  "model.layers.18.self_attn.k_proj.weight": 1,
+  "model.layers.18.self_attn.o_proj.weight": 1,
+  "model.layers.18.self_attn.q_proj.weight": 1,
+  "model.layers.18.self_attn.v_proj.weight": 1,
+  "model.layers.19.input_layernorm.weight": 1,
+  "model.layers.19.mlp.down_proj.weight": 1,
+  "model.layers.19.mlp.gate_proj.weight": 1,
+  "model.layers.19.mlp.up_proj.weight": 1,
+  "model.layers.19.post_attention_layernorm.weight": 1,
+  "model.layers.19.self_attn.k_proj.weight": 1,
+  "model.layers.19.self_attn.o_proj.weight": 1,
+  "model.layers.19.self_attn.q_proj.weight": 1,
+  "model.layers.19.self_attn.v_proj.weight": 1,
+  "model.layers.2.input_layernorm.weight": 1,
+  "model.layers.2.mlp.down_proj.weight": 1,
+  "model.layers.2.mlp.gate_proj.weight": 1,
+  "model.layers.2.mlp.up_proj.weight": 1,
+  "model.layers.2.post_attention_layernorm.weight": 1,
+  "model.layers.2.self_attn.k_proj.weight": 1,
+  "model.layers.2.self_attn.o_proj.weight": 1,
+  "model.layers.2.self_attn.q_proj.weight": 1,
+  "model.layers.2.self_attn.v_proj.weight": 1,
+  "model.layers.20.input_layernorm.weight": 2,
+  "model.layers.20.mlp.down_proj.weight": 2,
+  "model.layers.20.mlp.gate_proj.weight": 1,
+  "model.layers.20.mlp.up_proj.weight": 1,
+  "model.layers.20.post_attention_layernorm.weight": 2,
+  "model.layers.20.self_attn.k_proj.weight": 1,
+  "model.layers.20.self_attn.o_proj.weight": 1,
+  "model.layers.20.self_attn.q_proj.weight": 1,
+  "model.layers.20.self_attn.v_proj.weight": 1,
+  "model.layers.21.input_layernorm.weight": 2,
+  "model.layers.21.mlp.down_proj.weight": 2,
+  "model.layers.21.mlp.gate_proj.weight": 2,
+  "model.layers.21.mlp.up_proj.weight": 2,
+  "model.layers.21.post_attention_layernorm.weight": 2,
+  "model.layers.21.self_attn.k_proj.weight": 2,
+  "model.layers.21.self_attn.o_proj.weight": 2,
+  "model.layers.21.self_attn.q_proj.weight": 2,
+  "model.layers.21.self_attn.v_proj.weight": 2,
+  "model.layers.22.input_layernorm.weight": 2,
+  "model.layers.22.mlp.down_proj.weight": 2,
+  "model.layers.22.mlp.gate_proj.weight": 2,
+  "model.layers.22.mlp.up_proj.weight": 2,
+  "model.layers.22.post_attention_layernorm.weight": 2,
+  "model.layers.22.self_attn.k_proj.weight": 2,
+  "model.layers.22.self_attn.o_proj.weight": 2,
+  "model.layers.22.self_attn.q_proj.weight": 2,
+  "model.layers.22.self_attn.v_proj.weight": 2,
+  "model.layers.23.input_layernorm.weight": 2,
+  "model.layers.23.mlp.down_proj.weight": 2,
+  "model.layers.23.mlp.gate_proj.weight": 2,
+  "model.layers.23.mlp.up_proj.weight": 2,
+  "model.layers.23.post_attention_layernorm.weight": 2,
+  "model.layers.23.self_attn.k_proj.weight": 2,
+  "model.layers.23.self_attn.o_proj.weight": 2,
+  "model.layers.23.self_attn.q_proj.weight": 2,
+  "model.layers.23.self_attn.v_proj.weight": 2,
+  "model.layers.24.input_layernorm.weight": 2,
+  "model.layers.24.mlp.down_proj.weight": 2,
+  "model.layers.24.mlp.gate_proj.weight": 2,
+  "model.layers.24.mlp.up_proj.weight": 2,
+  "model.layers.24.post_attention_layernorm.weight": 2,
+  "model.layers.24.self_attn.k_proj.weight": 2,
+  "model.layers.24.self_attn.o_proj.weight": 2,
+  "model.layers.24.self_attn.q_proj.weight": 2,
+  "model.layers.24.self_attn.v_proj.weight": 2,
+  "model.layers.25.input_layernorm.weight": 2,
+  "model.layers.25.mlp.down_proj.weight": 2,
+  "model.layers.25.mlp.gate_proj.weight": 2,
+  "model.layers.25.mlp.up_proj.weight": 2,
+  "model.layers.25.post_attention_layernorm.weight": 2,
+  "model.layers.25.self_attn.k_proj.weight": 2,
+  "model.layers.25.self_attn.o_proj.weight": 2,
+  "model.layers.25.self_attn.q_proj.weight": 2,
+  "model.layers.25.self_attn.v_proj.weight": 2,
+  "model.layers.26.input_layernorm.weight": 2,
+  "model.layers.26.mlp.down_proj.weight": 2,
+  "model.layers.26.mlp.gate_proj.weight": 2,
+  "model.layers.26.mlp.up_proj.weight": 2,
+  "model.layers.26.post_attention_layernorm.weight": 2,
+  "model.layers.26.self_attn.k_proj.weight": 2,
+  "model.layers.26.self_attn.o_proj.weight": 2,
+  "model.layers.26.self_attn.q_proj.weight": 2,
+  "model.layers.26.self_attn.v_proj.weight": 2,
+  "model.layers.27.input_layernorm.weight": 2,
+  "model.layers.27.mlp.down_proj.weight": 2,
+  "model.layers.27.mlp.gate_proj.weight": 2,
+  "model.layers.27.mlp.up_proj.weight": 2,
+  "model.layers.27.post_attention_layernorm.weight": 2,
+  "model.layers.27.self_attn.k_proj.weight": 2,
+  "model.layers.27.self_attn.o_proj.weight": 2,
+  "model.layers.27.self_attn.q_proj.weight": 2,
+  "model.layers.27.self_attn.v_proj.weight": 2,
+  "model.layers.3.input_layernorm.weight": 1,
+  "model.layers.3.mlp.down_proj.weight": 1,
+  "model.layers.3.mlp.gate_proj.weight": 1,
+  "model.layers.3.mlp.up_proj.weight": 1,
+  "model.layers.3.post_attention_layernorm.weight": 1,
+  "model.layers.3.self_attn.k_proj.weight": 1,
+  "model.layers.3.self_attn.o_proj.weight": 1,
+  "model.layers.3.self_attn.q_proj.weight": 1,
+  "model.layers.3.self_attn.v_proj.weight": 1,
+  "model.layers.4.input_layernorm.weight": 1,
+  "model.layers.4.mlp.down_proj.weight": 1,
+  "model.layers.4.mlp.gate_proj.weight": 1,
+  "model.layers.4.mlp.up_proj.weight": 1,
+  "model.layers.4.post_attention_layernorm.weight": 1,
+  "model.layers.4.self_attn.k_proj.weight": 1,
+  "model.layers.4.self_attn.o_proj.weight": 1,
+  "model.layers.4.self_attn.q_proj.weight": 1,
+  "model.layers.4.self_attn.v_proj.weight": 1,
+  "model.layers.5.input_layernorm.weight": 1,
+  "model.layers.5.mlp.down_proj.weight": 1,
+  "model.layers.5.mlp.gate_proj.weight": 1,
+  "model.layers.5.mlp.up_proj.weight": 1,
+  "model.layers.5.post_attention_layernorm.weight": 1,
+  "model.layers.5.self_attn.k_proj.weight": 1,
+  "model.layers.5.self_attn.o_proj.weight": 1,
+  "model.layers.5.self_attn.q_proj.weight": 1,
+  "model.layers.5.self_attn.v_proj.weight": 1,
+  "model.layers.6.input_layernorm.weight": 1,
+  "model.layers.6.mlp.down_proj.weight": 1,
+  "model.layers.6.mlp.gate_proj.weight": 1,
+  "model.layers.6.mlp.up_proj.weight": 1,
+  "model.layers.6.post_attention_layernorm.weight": 1,
+  "model.layers.6.self_attn.k_proj.weight": 1,
+  "model.layers.6.self_attn.o_proj.weight": 1,
+  "model.layers.6.self_attn.q_proj.weight": 1,
+  "model.layers.6.self_attn.v_proj.weight": 1,
+  "model.layers.7.input_layernorm.weight": 1,
+  "model.layers.7.mlp.down_proj.weight": 1,
+  "model.layers.7.mlp.gate_proj.weight": 1,
+  "model.layers.7.mlp.up_proj.weight": 1,
+  "model.layers.7.post_attention_layernorm.weight": 1,
+  "model.layers.7.self_attn.k_proj.weight": 1,
+  "model.layers.7.self_attn.o_proj.weight": 1,
+  "model.layers.7.self_attn.q_proj.weight": 1,
+  "model.layers.7.self_attn.v_proj.weight": 1,
+  "model.layers.8.input_layernorm.weight": 1,
+  "model.layers.8.mlp.down_proj.weight": 1,
+  "model.layers.8.mlp.gate_proj.weight": 1,
+  "model.layers.8.mlp.up_proj.weight": 1,
+  "model.layers.8.post_attention_layernorm.weight": 1,
+  "model.layers.8.self_attn.k_proj.weight": 1,
+  "model.layers.8.self_attn.o_proj.weight": 1,
+  "model.layers.8.self_attn.q_proj.weight": 1,
+  "model.layers.8.self_attn.v_proj.weight": 1,
+  "model.layers.9.input_layernorm.weight": 1,
+  "model.layers.9.mlp.down_proj.weight": 1,
+  "model.layers.9.mlp.gate_proj.weight": 1,
+  "model.layers.9.mlp.up_proj.weight": 1,
+  "model.layers.9.post_attention_layernorm.weight": 1,
+  "model.layers.9.self_attn.k_proj.weight": 1,
+  "model.layers.9.self_attn.o_proj.weight": 1,
+  "model.layers.9.self_attn.q_proj.weight": 1,
+  "model.layers.9.self_attn.v_proj.weight": 1,
+  "model.norm.weight": 2
+}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/generation_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/generation_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..125cc74759bc705ab8a0c48d07d7eac083f56011
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/policy/weights/model/.hf_metadata/generation_config.json
@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.1"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/training_info.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/training_info.json
new file mode 100644
index 0000000000000000000000000000000000000000..68d18b42ba0f8f5fca890e246a6ff576854633ca
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_160/training_info.json
@@ -0,0 +1 @@
+{"current_epoch": 31, "current_step": 5, "total_steps": 160, "val_reward": 0.21354166666666666, "consumed_samples": 3200, "total_valid_tokens": 9497172.517022463}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/chat_template.jinja b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..1bad6a0f648dccdbec523ca79ba90fbcfc806af0
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/chat_template.jinja
@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/special_tokens_map.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/special_tokens_map.json
new file mode 100644
index 0000000000000000000000000000000000000000..b43be96621d147110fb8a18b5776ec6e38516127
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/special_tokens_map.json
@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|eot_id|>"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/tokenizer_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..3beeacc86a6ca3cae14ad3004263ab74a4bac07a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/tokenizer/tokenizer_config.json
@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|eot_id|>",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..88a7ab65ecc03929fde705948a65e30260067bc3
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/config.json
@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "float32",
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "nemo_version": "0.2.0",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": 128009,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "vocab_size": 128256
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
new file mode 100644
index 0000000000000000000000000000000000000000..494ffc91e4983233a079912895e9dc90c8c11a79
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
@@ -0,0 +1,256 @@
+{
+  "model.embed_tokens.weight": 1,
+  "model.layers.0.input_layernorm.weight": 1,
+  "model.layers.0.mlp.down_proj.weight": 1,
+  "model.layers.0.mlp.gate_proj.weight": 1,
+  "model.layers.0.mlp.up_proj.weight": 1,
+  "model.layers.0.post_attention_layernorm.weight": 1,
+  "model.layers.0.self_attn.k_proj.weight": 1,
+  "model.layers.0.self_attn.o_proj.weight": 1,
+  "model.layers.0.self_attn.q_proj.weight": 1,
+  "model.layers.0.self_attn.v_proj.weight": 1,
+  "model.layers.1.input_layernorm.weight": 1,
+  "model.layers.1.mlp.down_proj.weight": 1,
+  "model.layers.1.mlp.gate_proj.weight": 1,
+  "model.layers.1.mlp.up_proj.weight": 1,
+  "model.layers.1.post_attention_layernorm.weight": 1,
+  "model.layers.1.self_attn.k_proj.weight": 1,
+  "model.layers.1.self_attn.o_proj.weight": 1,
+  "model.layers.1.self_attn.q_proj.weight": 1,
+  "model.layers.1.self_attn.v_proj.weight": 1,
+  "model.layers.10.input_layernorm.weight": 1,
+  "model.layers.10.mlp.down_proj.weight": 1,
+  "model.layers.10.mlp.gate_proj.weight": 1,
+  "model.layers.10.mlp.up_proj.weight": 1,
+  "model.layers.10.post_attention_layernorm.weight": 1,
+  "model.layers.10.self_attn.k_proj.weight": 1,
+  "model.layers.10.self_attn.o_proj.weight": 1,
+  "model.layers.10.self_attn.q_proj.weight": 1,
+  "model.layers.10.self_attn.v_proj.weight": 1,
+  "model.layers.11.input_layernorm.weight": 1,
+  "model.layers.11.mlp.down_proj.weight": 1,
+  "model.layers.11.mlp.gate_proj.weight": 1,
+  "model.layers.11.mlp.up_proj.weight": 1,
+  "model.layers.11.post_attention_layernorm.weight": 1,
+  "model.layers.11.self_attn.k_proj.weight": 1,
+  "model.layers.11.self_attn.o_proj.weight": 1,
+  "model.layers.11.self_attn.q_proj.weight": 1,
+  "model.layers.11.self_attn.v_proj.weight": 1,
+  "model.layers.12.input_layernorm.weight": 1,
+  "model.layers.12.mlp.down_proj.weight": 1,
+  "model.layers.12.mlp.gate_proj.weight": 1,
+  "model.layers.12.mlp.up_proj.weight": 1,
+  "model.layers.12.post_attention_layernorm.weight": 1,
+  "model.layers.12.self_attn.k_proj.weight": 1,
+  "model.layers.12.self_attn.o_proj.weight": 1,
+  "model.layers.12.self_attn.q_proj.weight": 1,
+  "model.layers.12.self_attn.v_proj.weight": 1,
+  "model.layers.13.input_layernorm.weight": 1,
+  "model.layers.13.mlp.down_proj.weight": 1,
+  "model.layers.13.mlp.gate_proj.weight": 1,
+  "model.layers.13.mlp.up_proj.weight": 1,
+  "model.layers.13.post_attention_layernorm.weight": 1,
+  "model.layers.13.self_attn.k_proj.weight": 1,
+  "model.layers.13.self_attn.o_proj.weight": 1,
+  "model.layers.13.self_attn.q_proj.weight": 1,
+  "model.layers.13.self_attn.v_proj.weight": 1,
+  "model.layers.14.input_layernorm.weight": 1,
+  "model.layers.14.mlp.down_proj.weight": 1,
+  "model.layers.14.mlp.gate_proj.weight": 1,
+  "model.layers.14.mlp.up_proj.weight": 1,
+  "model.layers.14.post_attention_layernorm.weight": 1,
+  "model.layers.14.self_attn.k_proj.weight": 1,
+  "model.layers.14.self_attn.o_proj.weight": 1,
+  "model.layers.14.self_attn.q_proj.weight": 1,
+  "model.layers.14.self_attn.v_proj.weight": 1,
+  "model.layers.15.input_layernorm.weight": 1,
+  "model.layers.15.mlp.down_proj.weight": 1,
+  "model.layers.15.mlp.gate_proj.weight": 1,
+  "model.layers.15.mlp.up_proj.weight": 1,
+  "model.layers.15.post_attention_layernorm.weight": 1,
+  "model.layers.15.self_attn.k_proj.weight": 1,
+  "model.layers.15.self_attn.o_proj.weight": 1,
+  "model.layers.15.self_attn.q_proj.weight": 1,
+  "model.layers.15.self_attn.v_proj.weight": 1,
+  "model.layers.16.input_layernorm.weight": 1,
+  "model.layers.16.mlp.down_proj.weight": 1,
+  "model.layers.16.mlp.gate_proj.weight": 1,
+  "model.layers.16.mlp.up_proj.weight": 1,
+  "model.layers.16.post_attention_layernorm.weight": 1,
+  "model.layers.16.self_attn.k_proj.weight": 1,
+  "model.layers.16.self_attn.o_proj.weight": 1,
+  "model.layers.16.self_attn.q_proj.weight": 1,
+  "model.layers.16.self_attn.v_proj.weight": 1,
+  "model.layers.17.input_layernorm.weight": 1,
+  "model.layers.17.mlp.down_proj.weight": 1,
+  "model.layers.17.mlp.gate_proj.weight": 1,
+  "model.layers.17.mlp.up_proj.weight": 1,
+  "model.layers.17.post_attention_layernorm.weight": 1,
+  "model.layers.17.self_attn.k_proj.weight": 1,
+  "model.layers.17.self_attn.o_proj.weight": 1,
+  "model.layers.17.self_attn.q_proj.weight": 1,
+  "model.layers.17.self_attn.v_proj.weight": 1,
+  "model.layers.18.input_layernorm.weight": 1,
+  "model.layers.18.mlp.down_proj.weight": 1,
+  "model.layers.18.mlp.gate_proj.weight": 1,
+  "model.layers.18.mlp.up_proj.weight": 1,
+  "model.layers.18.post_attention_layernorm.weight": 1,
+  "model.layers.18.self_attn.k_proj.weight": 1,
+  "model.layers.18.self_attn.o_proj.weight": 1,
+  "model.layers.18.self_attn.q_proj.weight": 1,
+  "model.layers.18.self_attn.v_proj.weight": 1,
+  "model.layers.19.input_layernorm.weight": 1,
+  "model.layers.19.mlp.down_proj.weight": 1,
+  "model.layers.19.mlp.gate_proj.weight": 1,
+  "model.layers.19.mlp.up_proj.weight": 1,
+  "model.layers.19.post_attention_layernorm.weight": 1,
+  "model.layers.19.self_attn.k_proj.weight": 1,
+  "model.layers.19.self_attn.o_proj.weight": 1,
+  "model.layers.19.self_attn.q_proj.weight": 1,
+  "model.layers.19.self_attn.v_proj.weight": 1,
+  "model.layers.2.input_layernorm.weight": 1,
+  "model.layers.2.mlp.down_proj.weight": 1,
+  "model.layers.2.mlp.gate_proj.weight": 1,
+  "model.layers.2.mlp.up_proj.weight": 1,
+  "model.layers.2.post_attention_layernorm.weight": 1,
+  "model.layers.2.self_attn.k_proj.weight": 1,
+  "model.layers.2.self_attn.o_proj.weight": 1,
+  "model.layers.2.self_attn.q_proj.weight": 1,
+  "model.layers.2.self_attn.v_proj.weight": 1,
+  "model.layers.20.input_layernorm.weight": 2,
+  "model.layers.20.mlp.down_proj.weight": 2,
+  "model.layers.20.mlp.gate_proj.weight": 1,
+  "model.layers.20.mlp.up_proj.weight": 1,
+  "model.layers.20.post_attention_layernorm.weight": 2,
+  "model.layers.20.self_attn.k_proj.weight": 1,
+  "model.layers.20.self_attn.o_proj.weight": 1,
+  "model.layers.20.self_attn.q_proj.weight": 1,
+  "model.layers.20.self_attn.v_proj.weight": 1,
+  "model.layers.21.input_layernorm.weight": 2,
+  "model.layers.21.mlp.down_proj.weight": 2,
+  "model.layers.21.mlp.gate_proj.weight": 2,
+  "model.layers.21.mlp.up_proj.weight": 2,
+  "model.layers.21.post_attention_layernorm.weight": 2,
+  "model.layers.21.self_attn.k_proj.weight": 2,
+  "model.layers.21.self_attn.o_proj.weight": 2,
+  "model.layers.21.self_attn.q_proj.weight": 2,
+  "model.layers.21.self_attn.v_proj.weight": 2,
+  "model.layers.22.input_layernorm.weight": 2,
+  "model.layers.22.mlp.down_proj.weight": 2,
+  "model.layers.22.mlp.gate_proj.weight": 2,
+  "model.layers.22.mlp.up_proj.weight": 2,
+  "model.layers.22.post_attention_layernorm.weight": 2,
+  "model.layers.22.self_attn.k_proj.weight": 2,
+  "model.layers.22.self_attn.o_proj.weight": 2,
+  "model.layers.22.self_attn.q_proj.weight": 2,
+  "model.layers.22.self_attn.v_proj.weight": 2,
+  "model.layers.23.input_layernorm.weight": 2,
+  "model.layers.23.mlp.down_proj.weight": 2,
+  "model.layers.23.mlp.gate_proj.weight": 2,
+  "model.layers.23.mlp.up_proj.weight": 2,
+  "model.layers.23.post_attention_layernorm.weight": 2,
+  "model.layers.23.self_attn.k_proj.weight": 2,
+  "model.layers.23.self_attn.o_proj.weight": 2,
+  "model.layers.23.self_attn.q_proj.weight": 2,
+  "model.layers.23.self_attn.v_proj.weight": 2,
+  "model.layers.24.input_layernorm.weight": 2,
+  "model.layers.24.mlp.down_proj.weight": 2,
+  "model.layers.24.mlp.gate_proj.weight": 2,
+  "model.layers.24.mlp.up_proj.weight": 2,
+  "model.layers.24.post_attention_layernorm.weight": 2,
+  "model.layers.24.self_attn.k_proj.weight": 2,
+  "model.layers.24.self_attn.o_proj.weight": 2,
+  "model.layers.24.self_attn.q_proj.weight": 2,
+  "model.layers.24.self_attn.v_proj.weight": 2,
+  "model.layers.25.input_layernorm.weight": 2,
+  "model.layers.25.mlp.down_proj.weight": 2,
+  "model.layers.25.mlp.gate_proj.weight": 2,
+  "model.layers.25.mlp.up_proj.weight": 2,
+  "model.layers.25.post_attention_layernorm.weight": 2,
+  "model.layers.25.self_attn.k_proj.weight": 2,
+  "model.layers.25.self_attn.o_proj.weight": 2,
+  "model.layers.25.self_attn.q_proj.weight": 2,
+  "model.layers.25.self_attn.v_proj.weight": 2,
+  "model.layers.26.input_layernorm.weight": 2,
+  "model.layers.26.mlp.down_proj.weight": 2,
+  "model.layers.26.mlp.gate_proj.weight": 2,
+  "model.layers.26.mlp.up_proj.weight": 2,
+  "model.layers.26.post_attention_layernorm.weight": 2,
+  "model.layers.26.self_attn.k_proj.weight": 2,
+  "model.layers.26.self_attn.o_proj.weight": 2,
+  "model.layers.26.self_attn.q_proj.weight": 2,
+  "model.layers.26.self_attn.v_proj.weight": 2,
+  "model.layers.27.input_layernorm.weight": 2,
+  "model.layers.27.mlp.down_proj.weight": 2,
+  "model.layers.27.mlp.gate_proj.weight": 2,
+  "model.layers.27.mlp.up_proj.weight": 2,
+  "model.layers.27.post_attention_layernorm.weight": 2,
+  "model.layers.27.self_attn.k_proj.weight": 2,
+  "model.layers.27.self_attn.o_proj.weight": 2,
+  "model.layers.27.self_attn.q_proj.weight": 2,
+  "model.layers.27.self_attn.v_proj.weight": 2,
+  "model.layers.3.input_layernorm.weight": 1,
+  "model.layers.3.mlp.down_proj.weight": 1,
+  "model.layers.3.mlp.gate_proj.weight": 1,
+  "model.layers.3.mlp.up_proj.weight": 1,
+  "model.layers.3.post_attention_layernorm.weight": 1,
+  "model.layers.3.self_attn.k_proj.weight": 1,
+  "model.layers.3.self_attn.o_proj.weight": 1,
+  "model.layers.3.self_attn.q_proj.weight": 1,
+  "model.layers.3.self_attn.v_proj.weight": 1,
+  "model.layers.4.input_layernorm.weight": 1,
+  "model.layers.4.mlp.down_proj.weight": 1,
+  "model.layers.4.mlp.gate_proj.weight": 1,
+  "model.layers.4.mlp.up_proj.weight": 1,
+  "model.layers.4.post_attention_layernorm.weight": 1,
+  "model.layers.4.self_attn.k_proj.weight": 1,
+  "model.layers.4.self_attn.o_proj.weight": 1,
+  "model.layers.4.self_attn.q_proj.weight": 1,
+  "model.layers.4.self_attn.v_proj.weight": 1,
+  "model.layers.5.input_layernorm.weight": 1,
+  "model.layers.5.mlp.down_proj.weight": 1,
+  "model.layers.5.mlp.gate_proj.weight": 1,
+  "model.layers.5.mlp.up_proj.weight": 1,
+  "model.layers.5.post_attention_layernorm.weight": 1,
+  "model.layers.5.self_attn.k_proj.weight": 1,
+  "model.layers.5.self_attn.o_proj.weight": 1,
+  "model.layers.5.self_attn.q_proj.weight": 1,
+  "model.layers.5.self_attn.v_proj.weight": 1,
+  "model.layers.6.input_layernorm.weight": 1,
+  "model.layers.6.mlp.down_proj.weight": 1,
+  "model.layers.6.mlp.gate_proj.weight": 1,
+  "model.layers.6.mlp.up_proj.weight": 1,
+  "model.layers.6.post_attention_layernorm.weight": 1,
+  "model.layers.6.self_attn.k_proj.weight": 1,
+  "model.layers.6.self_attn.o_proj.weight": 1,
+  "model.layers.6.self_attn.q_proj.weight": 1,
+  "model.layers.6.self_attn.v_proj.weight": 1,
+  "model.layers.7.input_layernorm.weight": 1,
+  "model.layers.7.mlp.down_proj.weight": 1,
+  "model.layers.7.mlp.gate_proj.weight": 1,
+  "model.layers.7.mlp.up_proj.weight": 1,
+  "model.layers.7.post_attention_layernorm.weight": 1,
+  "model.layers.7.self_attn.k_proj.weight": 1,
+  "model.layers.7.self_attn.o_proj.weight": 1,
+  "model.layers.7.self_attn.q_proj.weight": 1,
+  "model.layers.7.self_attn.v_proj.weight": 1,
+  "model.layers.8.input_layernorm.weight": 1,
+  "model.layers.8.mlp.down_proj.weight": 1,
+  "model.layers.8.mlp.gate_proj.weight": 1,
+  "model.layers.8.mlp.up_proj.weight": 1,
+  "model.layers.8.post_attention_layernorm.weight": 1,
+  "model.layers.8.self_attn.k_proj.weight": 1,
+  "model.layers.8.self_attn.o_proj.weight": 1,
+  "model.layers.8.self_attn.q_proj.weight": 1,
+  "model.layers.8.self_attn.v_proj.weight": 1,
+  "model.layers.9.input_layernorm.weight": 1,
+  "model.layers.9.mlp.down_proj.weight": 1,
+  "model.layers.9.mlp.gate_proj.weight": 1,
+  "model.layers.9.mlp.up_proj.weight": 1,
+  "model.layers.9.post_attention_layernorm.weight": 1,
+  "model.layers.9.self_attn.k_proj.weight": 1,
+  "model.layers.9.self_attn.o_proj.weight": 1,
+  "model.layers.9.self_attn.q_proj.weight": 1,
+  "model.layers.9.self_attn.v_proj.weight": 1,
+  "model.norm.weight": 2
+}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/generation_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/generation_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..125cc74759bc705ab8a0c48d07d7eac083f56011
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/policy/weights/model/.hf_metadata/generation_config.json
@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.1"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/training_info.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/training_info.json
new file mode 100644
index 0000000000000000000000000000000000000000..82018912a9aadcff0886ae16299433ff2b502a04
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_180/training_info.json
@@ -0,0 +1 @@
+{"current_epoch": 35, "current_step": 5, "total_steps": 180, "val_reward": 0.3229166666666667, "consumed_samples": 3600, "total_valid_tokens": 10439352.273438336}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/chat_template.jinja b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/chat_template.jinja
new file mode 100644
index 0000000000000000000000000000000000000000..1bad6a0f648dccdbec523ca79ba90fbcfc806af0
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/chat_template.jinja
@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/special_tokens_map.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/special_tokens_map.json
new file mode 100644
index 0000000000000000000000000000000000000000..b43be96621d147110fb8a18b5776ec6e38516127
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/special_tokens_map.json
@@ -0,0 +1,17 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|eot_id|>"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/tokenizer_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/tokenizer_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..3beeacc86a6ca3cae14ad3004263ab74a4bac07a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/tokenizer/tokenizer_config.json
@@ -0,0 +1,2063 @@
+{
+  "added_tokens_decoder": {
+    "128000": {
+      "content": "<|begin_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128001": {
+      "content": "<|end_of_text|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128002": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128003": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128004": {
+      "content": "<|finetune_right_pad_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128005": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128006": {
+      "content": "<|start_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128007": {
+      "content": "<|end_header_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128008": {
+      "content": "<|eom_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128009": {
+      "content": "<|eot_id|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128010": {
+      "content": "<|python_tag|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128011": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128012": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128013": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128014": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128015": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128016": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128017": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128018": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128019": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128020": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128021": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128022": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128023": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128024": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128025": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128026": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128027": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128028": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128029": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128030": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128031": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128032": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128033": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128034": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128035": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128036": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128037": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128038": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128039": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128040": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128041": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128042": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128043": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128044": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128045": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128046": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128047": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128048": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128049": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128050": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128051": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128052": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128053": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128054": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128055": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128056": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128057": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128058": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128059": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128060": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128061": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128062": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128063": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128064": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128065": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128066": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128067": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128068": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128069": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128070": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128071": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128072": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128073": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128074": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128075": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128076": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128077": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128078": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128079": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128080": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128081": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128082": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128083": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128084": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128085": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128086": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128087": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128088": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128089": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128090": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128091": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128092": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128093": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128094": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128095": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128096": {
+      "content": "<|reserved_special_token_88|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128097": {
+      "content": "<|reserved_special_token_89|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128098": {
+      "content": "<|reserved_special_token_90|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128099": {
+      "content": "<|reserved_special_token_91|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128100": {
+      "content": "<|reserved_special_token_92|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128101": {
+      "content": "<|reserved_special_token_93|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128102": {
+      "content": "<|reserved_special_token_94|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128103": {
+      "content": "<|reserved_special_token_95|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128104": {
+      "content": "<|reserved_special_token_96|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128105": {
+      "content": "<|reserved_special_token_97|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128106": {
+      "content": "<|reserved_special_token_98|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128107": {
+      "content": "<|reserved_special_token_99|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128108": {
+      "content": "<|reserved_special_token_100|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128109": {
+      "content": "<|reserved_special_token_101|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128110": {
+      "content": "<|reserved_special_token_102|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128111": {
+      "content": "<|reserved_special_token_103|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128112": {
+      "content": "<|reserved_special_token_104|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128113": {
+      "content": "<|reserved_special_token_105|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128114": {
+      "content": "<|reserved_special_token_106|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128115": {
+      "content": "<|reserved_special_token_107|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128116": {
+      "content": "<|reserved_special_token_108|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128117": {
+      "content": "<|reserved_special_token_109|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128118": {
+      "content": "<|reserved_special_token_110|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128119": {
+      "content": "<|reserved_special_token_111|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128120": {
+      "content": "<|reserved_special_token_112|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128121": {
+      "content": "<|reserved_special_token_113|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128122": {
+      "content": "<|reserved_special_token_114|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128123": {
+      "content": "<|reserved_special_token_115|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128124": {
+      "content": "<|reserved_special_token_116|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128125": {
+      "content": "<|reserved_special_token_117|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128126": {
+      "content": "<|reserved_special_token_118|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128127": {
+      "content": "<|reserved_special_token_119|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128128": {
+      "content": "<|reserved_special_token_120|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128129": {
+      "content": "<|reserved_special_token_121|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128130": {
+      "content": "<|reserved_special_token_122|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128131": {
+      "content": "<|reserved_special_token_123|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128132": {
+      "content": "<|reserved_special_token_124|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128133": {
+      "content": "<|reserved_special_token_125|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128134": {
+      "content": "<|reserved_special_token_126|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128135": {
+      "content": "<|reserved_special_token_127|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128136": {
+      "content": "<|reserved_special_token_128|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128137": {
+      "content": "<|reserved_special_token_129|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128138": {
+      "content": "<|reserved_special_token_130|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128139": {
+      "content": "<|reserved_special_token_131|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128140": {
+      "content": "<|reserved_special_token_132|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128141": {
+      "content": "<|reserved_special_token_133|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128142": {
+      "content": "<|reserved_special_token_134|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128143": {
+      "content": "<|reserved_special_token_135|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128144": {
+      "content": "<|reserved_special_token_136|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128145": {
+      "content": "<|reserved_special_token_137|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128146": {
+      "content": "<|reserved_special_token_138|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128147": {
+      "content": "<|reserved_special_token_139|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128148": {
+      "content": "<|reserved_special_token_140|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128149": {
+      "content": "<|reserved_special_token_141|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128150": {
+      "content": "<|reserved_special_token_142|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128151": {
+      "content": "<|reserved_special_token_143|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128152": {
+      "content": "<|reserved_special_token_144|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128153": {
+      "content": "<|reserved_special_token_145|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128154": {
+      "content": "<|reserved_special_token_146|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128155": {
+      "content": "<|reserved_special_token_147|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128156": {
+      "content": "<|reserved_special_token_148|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128157": {
+      "content": "<|reserved_special_token_149|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128158": {
+      "content": "<|reserved_special_token_150|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128159": {
+      "content": "<|reserved_special_token_151|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128160": {
+      "content": "<|reserved_special_token_152|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128161": {
+      "content": "<|reserved_special_token_153|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128162": {
+      "content": "<|reserved_special_token_154|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128163": {
+      "content": "<|reserved_special_token_155|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128164": {
+      "content": "<|reserved_special_token_156|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128165": {
+      "content": "<|reserved_special_token_157|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128166": {
+      "content": "<|reserved_special_token_158|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128167": {
+      "content": "<|reserved_special_token_159|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128168": {
+      "content": "<|reserved_special_token_160|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128169": {
+      "content": "<|reserved_special_token_161|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128170": {
+      "content": "<|reserved_special_token_162|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128171": {
+      "content": "<|reserved_special_token_163|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128172": {
+      "content": "<|reserved_special_token_164|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128173": {
+      "content": "<|reserved_special_token_165|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128174": {
+      "content": "<|reserved_special_token_166|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128175": {
+      "content": "<|reserved_special_token_167|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128176": {
+      "content": "<|reserved_special_token_168|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128177": {
+      "content": "<|reserved_special_token_169|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128178": {
+      "content": "<|reserved_special_token_170|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128179": {
+      "content": "<|reserved_special_token_171|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128180": {
+      "content": "<|reserved_special_token_172|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128181": {
+      "content": "<|reserved_special_token_173|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128182": {
+      "content": "<|reserved_special_token_174|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128183": {
+      "content": "<|reserved_special_token_175|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128184": {
+      "content": "<|reserved_special_token_176|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128185": {
+      "content": "<|reserved_special_token_177|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128186": {
+      "content": "<|reserved_special_token_178|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128187": {
+      "content": "<|reserved_special_token_179|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128188": {
+      "content": "<|reserved_special_token_180|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128189": {
+      "content": "<|reserved_special_token_181|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128190": {
+      "content": "<|reserved_special_token_182|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128191": {
+      "content": "<|reserved_special_token_183|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128192": {
+      "content": "<|reserved_special_token_184|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128193": {
+      "content": "<|reserved_special_token_185|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128194": {
+      "content": "<|reserved_special_token_186|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128195": {
+      "content": "<|reserved_special_token_187|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128196": {
+      "content": "<|reserved_special_token_188|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128197": {
+      "content": "<|reserved_special_token_189|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128198": {
+      "content": "<|reserved_special_token_190|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128199": {
+      "content": "<|reserved_special_token_191|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128200": {
+      "content": "<|reserved_special_token_192|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128201": {
+      "content": "<|reserved_special_token_193|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128202": {
+      "content": "<|reserved_special_token_194|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128203": {
+      "content": "<|reserved_special_token_195|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128204": {
+      "content": "<|reserved_special_token_196|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128205": {
+      "content": "<|reserved_special_token_197|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128206": {
+      "content": "<|reserved_special_token_198|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128207": {
+      "content": "<|reserved_special_token_199|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128208": {
+      "content": "<|reserved_special_token_200|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128209": {
+      "content": "<|reserved_special_token_201|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128210": {
+      "content": "<|reserved_special_token_202|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128211": {
+      "content": "<|reserved_special_token_203|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128212": {
+      "content": "<|reserved_special_token_204|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128213": {
+      "content": "<|reserved_special_token_205|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128214": {
+      "content": "<|reserved_special_token_206|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128215": {
+      "content": "<|reserved_special_token_207|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128216": {
+      "content": "<|reserved_special_token_208|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128217": {
+      "content": "<|reserved_special_token_209|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128218": {
+      "content": "<|reserved_special_token_210|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128219": {
+      "content": "<|reserved_special_token_211|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128220": {
+      "content": "<|reserved_special_token_212|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128221": {
+      "content": "<|reserved_special_token_213|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128222": {
+      "content": "<|reserved_special_token_214|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128223": {
+      "content": "<|reserved_special_token_215|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128224": {
+      "content": "<|reserved_special_token_216|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128225": {
+      "content": "<|reserved_special_token_217|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128226": {
+      "content": "<|reserved_special_token_218|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128227": {
+      "content": "<|reserved_special_token_219|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128228": {
+      "content": "<|reserved_special_token_220|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128229": {
+      "content": "<|reserved_special_token_221|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128230": {
+      "content": "<|reserved_special_token_222|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128231": {
+      "content": "<|reserved_special_token_223|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128232": {
+      "content": "<|reserved_special_token_224|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128233": {
+      "content": "<|reserved_special_token_225|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128234": {
+      "content": "<|reserved_special_token_226|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128235": {
+      "content": "<|reserved_special_token_227|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128236": {
+      "content": "<|reserved_special_token_228|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128237": {
+      "content": "<|reserved_special_token_229|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128238": {
+      "content": "<|reserved_special_token_230|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128239": {
+      "content": "<|reserved_special_token_231|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128240": {
+      "content": "<|reserved_special_token_232|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128241": {
+      "content": "<|reserved_special_token_233|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128242": {
+      "content": "<|reserved_special_token_234|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128243": {
+      "content": "<|reserved_special_token_235|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128244": {
+      "content": "<|reserved_special_token_236|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128245": {
+      "content": "<|reserved_special_token_237|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128246": {
+      "content": "<|reserved_special_token_238|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128247": {
+      "content": "<|reserved_special_token_239|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128248": {
+      "content": "<|reserved_special_token_240|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128249": {
+      "content": "<|reserved_special_token_241|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128250": {
+      "content": "<|reserved_special_token_242|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128251": {
+      "content": "<|reserved_special_token_243|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128252": {
+      "content": "<|reserved_special_token_244|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128253": {
+      "content": "<|reserved_special_token_245|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128254": {
+      "content": "<|reserved_special_token_246|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "128255": {
+      "content": "<|reserved_special_token_247|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|begin_of_text|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|eot_id|>",
+  "extra_special_tokens": {},
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|eot_id|>",
+  "tokenizer_class": "PreTrainedTokenizerFast"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..88a7ab65ecc03929fde705948a65e30260067bc3
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/config.json
@@ -0,0 +1,41 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 128000,
+  "dtype": "float32",
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "nemo_version": "0.2.0",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": 128009,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 32.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 8192,
+    "rope_type": "llama3"
+  },
+  "rope_theta": 500000.0,
+  "tie_word_embeddings": true,
+  "transformers_version": "4.57.1",
+  "use_cache": false,
+  "vocab_size": 128256
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
new file mode 100644
index 0000000000000000000000000000000000000000..494ffc91e4983233a079912895e9dc90c8c11a79
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/fqn_to_file_index_mapping.json
@@ -0,0 +1,256 @@
+{
+  "model.embed_tokens.weight": 1,
+  "model.layers.0.input_layernorm.weight": 1,
+  "model.layers.0.mlp.down_proj.weight": 1,
+  "model.layers.0.mlp.gate_proj.weight": 1,
+  "model.layers.0.mlp.up_proj.weight": 1,
+  "model.layers.0.post_attention_layernorm.weight": 1,
+  "model.layers.0.self_attn.k_proj.weight": 1,
+  "model.layers.0.self_attn.o_proj.weight": 1,
+  "model.layers.0.self_attn.q_proj.weight": 1,
+  "model.layers.0.self_attn.v_proj.weight": 1,
+  "model.layers.1.input_layernorm.weight": 1,
+  "model.layers.1.mlp.down_proj.weight": 1,
+  "model.layers.1.mlp.gate_proj.weight": 1,
+  "model.layers.1.mlp.up_proj.weight": 1,
+  "model.layers.1.post_attention_layernorm.weight": 1,
+  "model.layers.1.self_attn.k_proj.weight": 1,
+  "model.layers.1.self_attn.o_proj.weight": 1,
+  "model.layers.1.self_attn.q_proj.weight": 1,
+  "model.layers.1.self_attn.v_proj.weight": 1,
+  "model.layers.10.input_layernorm.weight": 1,
+  "model.layers.10.mlp.down_proj.weight": 1,
+  "model.layers.10.mlp.gate_proj.weight": 1,
+  "model.layers.10.mlp.up_proj.weight": 1,
+  "model.layers.10.post_attention_layernorm.weight": 1,
+  "model.layers.10.self_attn.k_proj.weight": 1,
+  "model.layers.10.self_attn.o_proj.weight": 1,
+  "model.layers.10.self_attn.q_proj.weight": 1,
+  "model.layers.10.self_attn.v_proj.weight": 1,
+  "model.layers.11.input_layernorm.weight": 1,
+  "model.layers.11.mlp.down_proj.weight": 1,
+  "model.layers.11.mlp.gate_proj.weight": 1,
+  "model.layers.11.mlp.up_proj.weight": 1,
+  "model.layers.11.post_attention_layernorm.weight": 1,
+  "model.layers.11.self_attn.k_proj.weight": 1,
+  "model.layers.11.self_attn.o_proj.weight": 1,
+  "model.layers.11.self_attn.q_proj.weight": 1,
+  "model.layers.11.self_attn.v_proj.weight": 1,
+  "model.layers.12.input_layernorm.weight": 1,
+  "model.layers.12.mlp.down_proj.weight": 1,
+  "model.layers.12.mlp.gate_proj.weight": 1,
+  "model.layers.12.mlp.up_proj.weight": 1,
+  "model.layers.12.post_attention_layernorm.weight": 1,
+  "model.layers.12.self_attn.k_proj.weight": 1,
+  "model.layers.12.self_attn.o_proj.weight": 1,
+  "model.layers.12.self_attn.q_proj.weight": 1,
+  "model.layers.12.self_attn.v_proj.weight": 1,
+  "model.layers.13.input_layernorm.weight": 1,
+  "model.layers.13.mlp.down_proj.weight": 1,
+  "model.layers.13.mlp.gate_proj.weight": 1,
+  "model.layers.13.mlp.up_proj.weight": 1,
+  "model.layers.13.post_attention_layernorm.weight": 1,
+  "model.layers.13.self_attn.k_proj.weight": 1,
+  "model.layers.13.self_attn.o_proj.weight": 1,
+  "model.layers.13.self_attn.q_proj.weight": 1,
+  "model.layers.13.self_attn.v_proj.weight": 1,
+  "model.layers.14.input_layernorm.weight": 1,
+  "model.layers.14.mlp.down_proj.weight": 1,
+  "model.layers.14.mlp.gate_proj.weight": 1,
+  "model.layers.14.mlp.up_proj.weight": 1,
+  "model.layers.14.post_attention_layernorm.weight": 1,
+  "model.layers.14.self_attn.k_proj.weight": 1,
+  "model.layers.14.self_attn.o_proj.weight": 1,
+  "model.layers.14.self_attn.q_proj.weight": 1,
+  "model.layers.14.self_attn.v_proj.weight": 1,
+  "model.layers.15.input_layernorm.weight": 1,
+  "model.layers.15.mlp.down_proj.weight": 1,
+  "model.layers.15.mlp.gate_proj.weight": 1,
+  "model.layers.15.mlp.up_proj.weight": 1,
+  "model.layers.15.post_attention_layernorm.weight": 1,
+  "model.layers.15.self_attn.k_proj.weight": 1,
+  "model.layers.15.self_attn.o_proj.weight": 1,
+  "model.layers.15.self_attn.q_proj.weight": 1,
+  "model.layers.15.self_attn.v_proj.weight": 1,
+  "model.layers.16.input_layernorm.weight": 1,
+  "model.layers.16.mlp.down_proj.weight": 1,
+  "model.layers.16.mlp.gate_proj.weight": 1,
+  "model.layers.16.mlp.up_proj.weight": 1,
+  "model.layers.16.post_attention_layernorm.weight": 1,
+  "model.layers.16.self_attn.k_proj.weight": 1,
+  "model.layers.16.self_attn.o_proj.weight": 1,
+  "model.layers.16.self_attn.q_proj.weight": 1,
+  "model.layers.16.self_attn.v_proj.weight": 1,
+  "model.layers.17.input_layernorm.weight": 1,
+  "model.layers.17.mlp.down_proj.weight": 1,
+  "model.layers.17.mlp.gate_proj.weight": 1,
+  "model.layers.17.mlp.up_proj.weight": 1,
+  "model.layers.17.post_attention_layernorm.weight": 1,
+  "model.layers.17.self_attn.k_proj.weight": 1,
+  "model.layers.17.self_attn.o_proj.weight": 1,
+  "model.layers.17.self_attn.q_proj.weight": 1,
+  "model.layers.17.self_attn.v_proj.weight": 1,
+  "model.layers.18.input_layernorm.weight": 1,
+  "model.layers.18.mlp.down_proj.weight": 1,
+  "model.layers.18.mlp.gate_proj.weight": 1,
+  "model.layers.18.mlp.up_proj.weight": 1,
+  "model.layers.18.post_attention_layernorm.weight": 1,
+  "model.layers.18.self_attn.k_proj.weight": 1,
+  "model.layers.18.self_attn.o_proj.weight": 1,
+  "model.layers.18.self_attn.q_proj.weight": 1,
+  "model.layers.18.self_attn.v_proj.weight": 1,
+  "model.layers.19.input_layernorm.weight": 1,
+  "model.layers.19.mlp.down_proj.weight": 1,
+  "model.layers.19.mlp.gate_proj.weight": 1,
+  "model.layers.19.mlp.up_proj.weight": 1,
+  "model.layers.19.post_attention_layernorm.weight": 1,
+  "model.layers.19.self_attn.k_proj.weight": 1,
+  "model.layers.19.self_attn.o_proj.weight": 1,
+  "model.layers.19.self_attn.q_proj.weight": 1,
+  "model.layers.19.self_attn.v_proj.weight": 1,
+  "model.layers.2.input_layernorm.weight": 1,
+  "model.layers.2.mlp.down_proj.weight": 1,
+  "model.layers.2.mlp.gate_proj.weight": 1,
+  "model.layers.2.mlp.up_proj.weight": 1,
+  "model.layers.2.post_attention_layernorm.weight": 1,
+  "model.layers.2.self_attn.k_proj.weight": 1,
+  "model.layers.2.self_attn.o_proj.weight": 1,
+  "model.layers.2.self_attn.q_proj.weight": 1,
+  "model.layers.2.self_attn.v_proj.weight": 1,
+  "model.layers.20.input_layernorm.weight": 2,
+  "model.layers.20.mlp.down_proj.weight": 2,
+  "model.layers.20.mlp.gate_proj.weight": 1,
+  "model.layers.20.mlp.up_proj.weight": 1,
+  "model.layers.20.post_attention_layernorm.weight": 2,
+  "model.layers.20.self_attn.k_proj.weight": 1,
+  "model.layers.20.self_attn.o_proj.weight": 1,
+  "model.layers.20.self_attn.q_proj.weight": 1,
+  "model.layers.20.self_attn.v_proj.weight": 1,
+  "model.layers.21.input_layernorm.weight": 2,
+  "model.layers.21.mlp.down_proj.weight": 2,
+  "model.layers.21.mlp.gate_proj.weight": 2,
+  "model.layers.21.mlp.up_proj.weight": 2,
+  "model.layers.21.post_attention_layernorm.weight": 2,
+  "model.layers.21.self_attn.k_proj.weight": 2,
+  "model.layers.21.self_attn.o_proj.weight": 2,
+  "model.layers.21.self_attn.q_proj.weight": 2,
+  "model.layers.21.self_attn.v_proj.weight": 2,
+  "model.layers.22.input_layernorm.weight": 2,
+  "model.layers.22.mlp.down_proj.weight": 2,
+  "model.layers.22.mlp.gate_proj.weight": 2,
+  "model.layers.22.mlp.up_proj.weight": 2,
+  "model.layers.22.post_attention_layernorm.weight": 2,
+  "model.layers.22.self_attn.k_proj.weight": 2,
+  "model.layers.22.self_attn.o_proj.weight": 2,
+  "model.layers.22.self_attn.q_proj.weight": 2,
+  "model.layers.22.self_attn.v_proj.weight": 2,
+  "model.layers.23.input_layernorm.weight": 2,
+  "model.layers.23.mlp.down_proj.weight": 2,
+  "model.layers.23.mlp.gate_proj.weight": 2,
+  "model.layers.23.mlp.up_proj.weight": 2,
+  "model.layers.23.post_attention_layernorm.weight": 2,
+  "model.layers.23.self_attn.k_proj.weight": 2,
+  "model.layers.23.self_attn.o_proj.weight": 2,
+  "model.layers.23.self_attn.q_proj.weight": 2,
+  "model.layers.23.self_attn.v_proj.weight": 2,
+  "model.layers.24.input_layernorm.weight": 2,
+  "model.layers.24.mlp.down_proj.weight": 2,
+  "model.layers.24.mlp.gate_proj.weight": 2,
+  "model.layers.24.mlp.up_proj.weight": 2,
+  "model.layers.24.post_attention_layernorm.weight": 2,
+  "model.layers.24.self_attn.k_proj.weight": 2,
+  "model.layers.24.self_attn.o_proj.weight": 2,
+  "model.layers.24.self_attn.q_proj.weight": 2,
+  "model.layers.24.self_attn.v_proj.weight": 2,
+  "model.layers.25.input_layernorm.weight": 2,
+  "model.layers.25.mlp.down_proj.weight": 2,
+  "model.layers.25.mlp.gate_proj.weight": 2,
+  "model.layers.25.mlp.up_proj.weight": 2,
+  "model.layers.25.post_attention_layernorm.weight": 2,
+  "model.layers.25.self_attn.k_proj.weight": 2,
+  "model.layers.25.self_attn.o_proj.weight": 2,
+  "model.layers.25.self_attn.q_proj.weight": 2,
+  "model.layers.25.self_attn.v_proj.weight": 2,
+  "model.layers.26.input_layernorm.weight": 2,
+  "model.layers.26.mlp.down_proj.weight": 2,
+  "model.layers.26.mlp.gate_proj.weight": 2,
+  "model.layers.26.mlp.up_proj.weight": 2,
+  "model.layers.26.post_attention_layernorm.weight": 2,
+  "model.layers.26.self_attn.k_proj.weight": 2,
+  "model.layers.26.self_attn.o_proj.weight": 2,
+  "model.layers.26.self_attn.q_proj.weight": 2,
+  "model.layers.26.self_attn.v_proj.weight": 2,
+  "model.layers.27.input_layernorm.weight": 2,
+  "model.layers.27.mlp.down_proj.weight": 2,
+  "model.layers.27.mlp.gate_proj.weight": 2,
+  "model.layers.27.mlp.up_proj.weight": 2,
+  "model.layers.27.post_attention_layernorm.weight": 2,
+  "model.layers.27.self_attn.k_proj.weight": 2,
+  "model.layers.27.self_attn.o_proj.weight": 2,
+  "model.layers.27.self_attn.q_proj.weight": 2,
+  "model.layers.27.self_attn.v_proj.weight": 2,
+  "model.layers.3.input_layernorm.weight": 1,
+  "model.layers.3.mlp.down_proj.weight": 1,
+  "model.layers.3.mlp.gate_proj.weight": 1,
+  "model.layers.3.mlp.up_proj.weight": 1,
+  "model.layers.3.post_attention_layernorm.weight": 1,
+  "model.layers.3.self_attn.k_proj.weight": 1,
+  "model.layers.3.self_attn.o_proj.weight": 1,
+  "model.layers.3.self_attn.q_proj.weight": 1,
+  "model.layers.3.self_attn.v_proj.weight": 1,
+  "model.layers.4.input_layernorm.weight": 1,
+  "model.layers.4.mlp.down_proj.weight": 1,
+  "model.layers.4.mlp.gate_proj.weight": 1,
+  "model.layers.4.mlp.up_proj.weight": 1,
+  "model.layers.4.post_attention_layernorm.weight": 1,
+  "model.layers.4.self_attn.k_proj.weight": 1,
+  "model.layers.4.self_attn.o_proj.weight": 1,
+  "model.layers.4.self_attn.q_proj.weight": 1,
+  "model.layers.4.self_attn.v_proj.weight": 1,
+  "model.layers.5.input_layernorm.weight": 1,
+  "model.layers.5.mlp.down_proj.weight": 1,
+  "model.layers.5.mlp.gate_proj.weight": 1,
+  "model.layers.5.mlp.up_proj.weight": 1,
+  "model.layers.5.post_attention_layernorm.weight": 1,
+  "model.layers.5.self_attn.k_proj.weight": 1,
+  "model.layers.5.self_attn.o_proj.weight": 1,
+  "model.layers.5.self_attn.q_proj.weight": 1,
+  "model.layers.5.self_attn.v_proj.weight": 1,
+  "model.layers.6.input_layernorm.weight": 1,
+  "model.layers.6.mlp.down_proj.weight": 1,
+  "model.layers.6.mlp.gate_proj.weight": 1,
+  "model.layers.6.mlp.up_proj.weight": 1,
+  "model.layers.6.post_attention_layernorm.weight": 1,
+  "model.layers.6.self_attn.k_proj.weight": 1,
+  "model.layers.6.self_attn.o_proj.weight": 1,
+  "model.layers.6.self_attn.q_proj.weight": 1,
+  "model.layers.6.self_attn.v_proj.weight": 1,
+  "model.layers.7.input_layernorm.weight": 1,
+  "model.layers.7.mlp.down_proj.weight": 1,
+  "model.layers.7.mlp.gate_proj.weight": 1,
+  "model.layers.7.mlp.up_proj.weight": 1,
+  "model.layers.7.post_attention_layernorm.weight": 1,
+  "model.layers.7.self_attn.k_proj.weight": 1,
+  "model.layers.7.self_attn.o_proj.weight": 1,
+  "model.layers.7.self_attn.q_proj.weight": 1,
+  "model.layers.7.self_attn.v_proj.weight": 1,
+  "model.layers.8.input_layernorm.weight": 1,
+  "model.layers.8.mlp.down_proj.weight": 1,
+  "model.layers.8.mlp.gate_proj.weight": 1,
+  "model.layers.8.mlp.up_proj.weight": 1,
+  "model.layers.8.post_attention_layernorm.weight": 1,
+  "model.layers.8.self_attn.k_proj.weight": 1,
+  "model.layers.8.self_attn.o_proj.weight": 1,
+  "model.layers.8.self_attn.q_proj.weight": 1,
+  "model.layers.8.self_attn.v_proj.weight": 1,
+  "model.layers.9.input_layernorm.weight": 1,
+  "model.layers.9.mlp.down_proj.weight": 1,
+  "model.layers.9.mlp.gate_proj.weight": 1,
+  "model.layers.9.mlp.up_proj.weight": 1,
+  "model.layers.9.post_attention_layernorm.weight": 1,
+  "model.layers.9.self_attn.k_proj.weight": 1,
+  "model.layers.9.self_attn.o_proj.weight": 1,
+  "model.layers.9.self_attn.q_proj.weight": 1,
+  "model.layers.9.self_attn.v_proj.weight": 1,
+  "model.norm.weight": 2
+}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/generation_config.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/generation_config.json
new file mode 100644
index 0000000000000000000000000000000000000000..125cc74759bc705ab8a0c48d07d7eac083f56011
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/policy/weights/model/.hf_metadata/generation_config.json
@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 128000,
+  "do_sample": true,
+  "eos_token_id": [
+    128001,
+    128008,
+    128009
+  ],
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.57.1"
+}
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/training_info.json b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/training_info.json
new file mode 100644
index 0000000000000000000000000000000000000000..5ca21660023679329dbdd5b4c95251156bb6c8fb
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_20/training_info.json
@@ -0,0 +1 @@
+{"current_epoch": 3, "current_step": 5, "total_steps": 20, "val_reward": 0.0, "consumed_samples": 400, "total_valid_tokens": 1721411.7689668084}
\ No newline at end of file
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_80/config.yaml b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_80/config.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..4f36007f4f838f0b33cc5310689faaeda7683dde
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/step_80/config.yaml
@@ -0,0 +1,255 @@
+checkpointing:
+  checkpoint_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716
+  checkpoint_must_save_by: null
+  enabled: true
+  higher_is_better: true
+  keep_top_k: null
+  metric_name: null
+  model_save_format: safetensors
+  save_consolidated: false
+  save_optimizer: false
+  save_per_epoch: false
+  save_period: 20
+cluster:
+  gpus_per_node: 4
+  num_nodes: 1
+data:
+  default:
+    env_name: knowledge
+    prompt_file: null
+    system_prompt_file: null
+  max_input_seq_length: 10000
+  shuffle: true
+  train:
+  - dataset_name: knowledge_opsd
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  validation:
+  - dataset_name: knowledge_opsd_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+  - dataset_name: knowledge_opsd_comp_val
+    env_name: knowledge
+    processor: knowledge_opsd
+    prompt_file: null
+    system_prompt_file: null
+distillation:
+  ema_decay: 1.0
+  max_num_epochs: 40
+  max_num_steps: 1000
+  max_rollout_turns: 1
+  max_val_samples: 200
+  num_generations_per_prompt: 8
+  num_prompts_per_step: 20
+  rollout_policy: student
+  seed: 42
+  topk_logits_k: 64
+  val_at_end: true
+  val_at_start: true
+  val_batch_size: 32
+  val_period: 20
+env:
+  knowledge:
+    num_workers: 8
+logger:
+  gpu_monitoring:
+    collection_interval: 10
+    flush_interval: 10
+  log_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001
+  mlflow_enabled: false
+  monitor_gpus: true
+  num_val_samples_to_print: 5
+  swanlab_enabled: false
+  tensorboard:
+    log_dir: tb_logs-opsd-knowledge-forward-kl-gt-d5-e10-llama3-2-3b
+  tensorboard_enabled: true
+  wandb:
+    entity: mathrl-jiarui
+    name: opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+    project: mix-teachers-opsd
+  wandb_enabled: true
+loss_fn:
+  kl_type: forward
+  mixed_kl_weight: 0.5
+  zero_outside_topk: false
+paths:
+  base_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/knowledge_opsd
+  checkpoint_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl
+  data_dir: /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  log_dir: /home/fs01/yw2349/jiarui/mathrl_data/mix_teachers/checkpoints/knowledge/opsd_nemo_rl/logs
+  model_search_dirs:
+  - /home/fs01/yw2349/jiarui/mathrl_data/models
+policy:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    _pad_token_id: 128009
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+    stop_strings: null
+    stop_token_ids:
+    - 128009
+    temperature: 0.8
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      hf_overrides: {}
+      kv_cache_dtype: auto
+      load_format: dummy
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      skip_tokenizer_init: true
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 20
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 3.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 20
+  train_micro_batch_size: 1
+teacher:
+  dtensor_cfg:
+    _v2: true
+    activation_checkpointing: true
+    context_parallel_size: 1
+    cpu_offload: true
+    custom_parallel_plan: null
+    enabled: true
+    sequence_parallel: false
+    tensor_parallel_size: 4
+  dynamic_batching:
+    enabled: true
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  generation:
+    backend: vllm
+    colocated:
+      enabled: true
+      resources:
+        gpus_per_node: null
+        num_nodes: null
+    max_new_tokens: 8192
+    stop_strings: null
+    stop_token_ids: null
+    temperature: 1.0
+    top_k: null
+    top_p: 1.0
+    vllm_cfg:
+      async_engine: false
+      distributed_executor_backend: null
+      enforce_eager: false
+      expert_parallel_size: 1
+      gpu_memory_utilization: 0.7
+      kv_cache_dtype: auto
+      max_model_len: 10000
+      num_first_layers_in_bf16: 0
+      num_last_layers_in_bf16: 0
+      pipeline_parallel_size: 1
+      precision: bfloat16
+      tensor_parallel_size: 4
+      use_deep_gemm: false
+  generation_batch_size: 16
+  logprob_batch_size: 1
+  logprob_chunk_size: null
+  make_sequence_length_divisible_by: 8
+  max_grad_norm: 1.0
+  max_total_sequence_length: 10000
+  megatron_cfg:
+    enabled: false
+  model_name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  offload_optimizer_for_logprob: false
+  optimizer:
+    kwargs:
+      betas:
+      - 0.9
+      - 0.95
+      eps: 1.0e-08
+      foreach: false
+      fused: false
+      lr: 1.0e-05
+      weight_decay: 0.01
+    name: torch.optim.AdamW
+  precision: bfloat16
+  scheduler:
+  - kwargs:
+      factor: 1.0
+      total_iters: 10000000000
+    name: torch.optim.lr_scheduler.ConstantLR
+  - milestones: []
+  sequence_packing:
+    algorithm: modified_first_fit_decreasing
+    enabled: false
+    logprob_mb_tokens: 10000
+    sequence_length_round: 64
+    train_mb_tokens: 10000
+  tokenizer:
+    chat_template_kwargs:
+      enable_thinking: false
+    name: /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+  train_global_batch_size: 16
+  train_micro_batch_size: 1
diff --git a/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/train_log_20260502.203716.txt b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/train_log_20260502.203716.txt
new file mode 100644
index 0000000000000000000000000000000000000000..f2ca10c52113b1b5ea7fa564dc2168d70c805a8a
--- /dev/null
+++ b/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/train_log_20260502.203716.txt
@@ -0,0 +1,85245 @@
+grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+  warnings.warn(
+/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+We recommend installing via `pip install torch-c-dlpack-ext`
+  warnings.warn(
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+  warnings.warn(
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+  warnings.warn(
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+  warnings.warn(
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+  warnings.warn("Apex is not installed. Falling back to Torch Norm")
+[run_opsd] Using teacher-context prompt variant: knowledge_base_answer
+[INFO] Dataset processor knowledge_opsd registered
+Registered knowledge_opsd datasets from /home/pj24003162/ku40003404/jiarui/TeacherMix/data/knowledge_d5_e10/atomic_sft
+  -> Patching _build_cot_gt_texts with knowledge GT variant
+  -> Patching _build_teacher_inputs_with_cot_in_user_turn with smart truncation (max_new_tokens=8192, teacher input budget=1808)
+2026-05-02 20:40:15,938	INFO worker.py:2004 -- Started a local Ray instance. View the dashboard at http://127.0.0.1:8265 
+/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/worker.py:2052: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+  warnings.warn(
+No chat template provided, using tokenizer's default
+
+▶ Setting up envs...
+
+▶ Setting up data...
+  - Loaded training dataset atomic_sft-train_messages with 100 samples.
+TaskDataSpec(task_name='atomic_sft-train_messages', prompt_file=None, system_prompt_file=None) <function _knowledge_opsd_processor at 0x14a40c5a9f80>
+  ✓ Training dataset loaded with 100 samples.
+
Generating train split: 0 examples [00:00, ? examples/s]
Generating train split: 100 examples [00:00, 13256.33 examples/s]
+  - Loaded validation dataset atomic_sft-val_paraphrased_messages with 100 samples.
+
Generating train split: 0 examples [00:00, ? examples/s]
Generating train split: 93 examples [00:00, 25076.84 examples/s]
+  - Loaded validation dataset compositional_2step_sft-test_messages with 93 samples.
+  ✓ Validation dataset loaded with 193 samples.
+wandb: [wandb.login()] Loaded credentials for https://api.wandb.ai from WANDB_API_KEY.
+wandb: Currently logged in as: jiaruil5 (monitorability) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: Tracking run with wandb version 0.25.1
+wandb: Run data is saved locally in /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/wandb/wandb/run-20260502_204022-jaynkchc
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Syncing run opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct
+wandb: ⭐️ View project at https://wandb.ai/mathrl-jiarui/mix-teachers-opsd
+wandb: 🚀 View run at https://wandb.ai/mathrl-jiarui/mix-teachers-opsd/runs/jaynkchc
+Unexpected error during git code logging: Artifact type 'code' is reserved for internal use. Please use a different type.
+No uncommitted changes found
+Error during git operations: Command '['git', 'diff', 'main']' returned non-zero exit status 128.
+Initialized WandbLogger for project mix-teachers-opsd, run opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct at /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/wandb
+Initialized TensorboardLogger at /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/tensorboard
+GPU monitoring started with collection interval=10s, flush interval=10s
+  ✓ Training dataloader loaded with 100 samples
+  ✓ Validation dataloader loaded with 193 samples
+
+▶ Setting up compute cluster...
+  ✓ Ray cluster initialized with 1 nodes
+
Initializing vllm_policy workers:   0%|          | 0/4 [00:00<?, ?worker/s](IsolatedWorkerInitializer pid=9751) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=9750) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=9748) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=9752) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+
Initializing vllm_policy workers:  25%|██▌       | 1/4 [00:12<00:36, 12.27s/worker]
Initializing vllm_policy workers: 100%|██████████| 4/4 [00:12<00:00,  2.39s/worker]
Initializing vllm_policy workers: 100%|██████████| 4/4 [00:12<00:00,  3.13s/worker]
+  ✓ 4 workers initialized in 12.55s
+(pid=10140) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=10139) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=10141) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=10142) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(VllmGenerationWorker pid=10140) VLLM_RAY_BUNDLE_INDICES environment variable set to: 0,1,2,3
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:25 [utils.py:253] non-default args: {'served_model_name': '/home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct', 'skip_tokenizer_init': True, 'trust_remote_code': True, 'load_format': 'dummy', 'dtype': 'bfloat16', 'seed': 0, 'max_model_len': 10000, 'distributed_executor_backend': 'ray', 'tensor_parallel_size': 4, 'enable_prefix_caching': True, 'gpu_memory_utilization': 0.7, 'logprobs_mode': 'processed_logprobs', 'worker_extension_cls': 'nemo_rl.models.generation.vllm.vllm_backend.VllmInternalWorkerExtension', 'enable_sleep_mode': True, 'model': '/home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct'}
+(VllmGenerationWorker pid=10140) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [model.py:631] Resolved architecture: LlamaForCausalLM
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [model.py:1745] Using max model len 10000
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [arg_utils.py:1443] Using ray runtime env (env vars redacted): {'py_executable': '/home/pj24003162/ku40003404/miniconda3/envs/opsd/bin/python3', 'env_vars': {'AI_AGENT': '***', 'AVAILABLE_ADDR_LIST': '***', 'AVAILABLE_PORT_LIST': '***', 'BASH_FUNC__module_raw%%': '***', 'BASH_FUNC_ml%%': '***', 'BASH_FUNC_module%%': '***', 'BASH_FUNC_scl%%': '***', 'BASH_FUNC_switchml%%': '***', 'BASH_FUNC_which%%': '***', 'CLAUDECODE': '***', 'CLAUDE_CODE_ENTRYPOINT': '***', 'CLAUDE_CODE_EXECPATH': '***', 'CONDA_DEFAULT_ENV': '***', 'CONDA_EXE': '***', 'CONDA_PREFIX': '***', 'CONDA_PREFIX_1': '***', 'CONDA_PROMPT_MODIFIER': '***', 'CONDA_PYTHON_EXE': '***', 'CONDA_SHLVL': '***', 'COREPACK_ENABLE_AUTO_PIN': '***', 'CPATH': '***', 'CPATH_modshare': '***', 'CUDA_DIR': '***', 'CUDA_HOME': '***', 'CUDA_INSTALL_PATH': '***', 'CUDA_MPS_LOG_DIRECTORY': '***', 'CUDA_MPS_PIPE_DIRECTORY': '***', 'CUDA_PATH': '***', 'CUDA_TOP': '***', 'CYCLECLOUD_BOOTSTRAP': '***', 'CYCLECLOUD_HOME': '***', 'C_INCLUDE_PATH': '***', 'C_INCLUDE_PATH_modshare': '***', 'DBUS_SESSION_BUS_ADDRESS': '***', 'DISPLAY': '***', 'FJPNCKPT_CONFIG': '***', 'GDK_BACKEND': '***', 'GENKAI_CACHE_DIR': '***', 'GENKAI_CUDA_NAME': '***', 'GENKAI_CUDA_VER': '***', 'GENKAI_FAST_DIR': '***', 'GIT_EDITOR': '***', 'GLOBUS_GSSAPI_NAME_COMPATIBILITY': '***', 'HF_TOKEN': '***', 'HISTCONTROL': '***', 'HISTSIZE': '***', 'HOME': '***', 'HOSTNAME': '***', 'HYDRA_FULL_ERROR': '***', 'JOBID': '***', 'KMP_DUPLICATE_LIB_OK': '***', 'KMP_INIT_AT_FORK': '***', 'LANG': '***', 'LD_LIBRARY_PATH': '***', 'LD_LIBRARY_PATH_modshare': '***', 'LESSOPEN': '***', 'LIBRARY_PATH': '***', 'LIBRARY_PATH_modshare': '***', 'LOADEDMODULES': '***', 'LOADEDMODULES_modshare': '***', 'LOCAL_RANK': '***', 'LOGNAME': '***', 'LS_COLORS': '***', 'LUSTRE_JOBSTAT': '***', 'MAIL': '***', 'MANPATH': '***', 'MASTER_ADDR': '***', 'MASTER_PORT': '***', 'MATHEMATICA_HOME': '***', 'MIG_PARTED_CHECKPOINT_FILE': '***', 'MIG_PARTED_CONFIG_FILE': '***', 'MIG_PARTED_HOOKS_FILE': '***', 'MODULEPATH': '***', 'MODULEPATH_modshare': '***', 'MODULESHOME': '***', 'MODULES_CMD': '***', 'MODULES_LMCONFLICT': '***', 'MODULES_LMCONFLICT_modshare': '***', 'MODULES_RUN_QUARANTINE': '***', 'MOD_GIT_ROOTDIR': '***', 'NCCL_ASYNC_ERROR_HANDLING': '***', 'NCCL_DEBUG': '***', 'NCCL_SHM_USE_CUDA_MEMCPY': '***', 'NCCL_TIMEOUT': '***', 'NEMO_RL_PY_EXECUTABLES_SYSTEM': '***', 'NEMO_RL_VENV_DIR': '***', 'NODE_RANK': '***', 'NoDefaultCurrentDirectoryInExePath': '***', 'OLDPWD': '***', 'OMPI_CKPTCONFIG': '***', 'OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE': '***', 'PATH': '***', 'PATH_modshare': '***', 'PJM_ASSIGN_LOGICAL_CPU': '***', 'PJM_CORE_MEM_LIMIT': '***', 'PJM_CUSTOM_RESOURCES': '***', 'PJM_DPREFIX': '***', 'PJM_ELAPSED_TIME_MODE': '***', 'PJM_ELAPSE_LIMIT': '***', 'PJM_ENVIRONMENT': '***', 'PJM_JOBDIR': '***', 'PJM_JOBID': '***', 'PJM_JOBNAME': '***', 'PJM_MAILOPTION': '***', 'PJM_MPI_PROC': '***', 'PJM_NET_ROUTE': '***', 'PJM_NODE': '***', 'PJM_NODE_CPUTIME_LIMIT': '***', 'PJM_O_HOME': '***', 'PJM_O_HOST': '***', 'PJM_O_LANG': '***', 'PJM_O_LOGNAME': '***', 'PJM_O_MAIL': '***', 'PJM_O_NODEINF': '***', 'PJM_O_PATH': '***', 'PJM_O_SHELL': '***', 'PJM_O_WORKDIR': '***', 'PJM_PROC_BY_NODE': '***', 'PJM_RSCGRP': '***', 'PJM_RSCUNIT': '***', 'PJM_SHELL': '***', 'PJM_SSD_DIR': '***', 'PJM_STDERR_PATH': '***', 'PJM_STDOUT_PATH': '***', 'PJM_SUBJOBID': '***', 'PJM_VNODE_MEM_LIMIT': '***', 'PLE_SCRIPT_TYPE': '***', 'PWD': '***', 'PYTHONBREAKPOINT': '***', 'PYTHONPATH': '***', 'PYTHONUNBUFFERED': '***', 'RANK': '***', 'RAY_CLIENT_MODE': '***', 'RAY_DEDUP_LOGS': '***', 'RAY_ENABLE_UV_RUN_RUNTIME_ENV': '***', 'RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES': '***', 'RAY_TMPDIR': '***', 'RAY_USAGE_STATS_ENABLED': '***', 'RAY_memory_monitor_refresh_ms': '***', 'RAY_worker_heartbeat_timeout_milliseconds': '***', 'RAY_worker_register_timeout_seconds': '***', 'SHLVL': '***', 'SSH_ASKPASS': '***', 'SSH_CLIENT': '***', 'SSH_CONNECTION': '***', 'SSH_TTY': '***', 'SWANLAB_API_HOST': '***', 'SWANLAB_RUNTIME': '***', 'SWANLAB_WEB_HOST': '***', 'S_COLORS': '***', 'TERM': '***', 'TMP': '***', 'TMPDIR': '***', 'TMUX': '***', 'TMUX_PANE': '***', 'TMUX_TMPDIR': '***', 'TORCHINDUCTOR_CACHE_DIR': '***', 'USER': '***', 'UV_PROJECT_ENVIRONMENT': '***', 'VIRTUAL_ENV': '***', 'VLLM_CACHE_ROOT': '***', 'VLLM_ENABLE_V1_MULTIPROCESSING': '***', 'VLLM_SKIP_P2P_CHECK': '***', 'VLLM_USE_V1': '***', 'WANDB_API_KEY': '***', 'WANDB_CONFIG_DIR': '***', 'WANDB_DIR': '***', 'WANDB_SERVICE': '***', 'WORLD_SIZE': '***', 'XDG_DATA_DIRS': '***', 'XDG_RUNTIME_DIR': '***', 'XDG_SESSION_ID': '***', '_': '***', '_CE_CONDA': '***', '_CE_M': '***', '_CONDA_EXE': '***', '_CONDA_ROOT': '***', '_LMFILES_': '***', '_LMFILES__modshare': '***', '_MLFLOW_TELEMETRY_SESSION_ID': '***', 'which_declare': '***'}}
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [scheduler.py:216] Chunked prefill is enabled with max_num_batched_tokens=16384.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [core.py:93] Initializing a V1 LLM engine (v0.11.2) with config: model='/home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct', speculative_config=None, tokenizer='/home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct', skip_tokenizer_init=True, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=10000, download_dir=None, load_format=dummy, tensor_parallel_size=4, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None), seed=0, served_model_name=/home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': <CompilationMode.VLLM_COMPILE: 3>, 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['none'], 'splitting_ops': ['vllm::unified_attention', 'vllm::unified_attention_with_output', 'vllm::unified_mla_attention', 'vllm::unified_mla_attention_with_output', 'vllm::mamba_mixer2', 'vllm::mamba_mixer', 'vllm::short_conv', 'vllm::linear_attention', 'vllm::plamo2_mamba_mixer', 'vllm::gdn_attention_core', 'vllm::kda_attention', 'vllm::sparse_attn_indexer'], 'compile_mm_encoder': False, 'use_inductor': None, 'compile_sizes': [], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': <CUDAGraphMode.FULL_AND_PIECEWISE: (2, 1)>, 'cudagraph_num_of_warmups': 1, 'cudagraph_capture_sizes': [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 272, 288, 304, 320, 336, 352, 368, 384, 400, 416, 432, 448, 464, 480, 496, 512], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {}, 'max_cudagraph_capture_size': 512, 'local_cache_dir': None}
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [ray_utils.py:343] Ray is already initialized. Skipping Ray initialization.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:41:46 [ray_utils.py:374] Using the existing placement group
+(VllmGenerationWorker pid=10140) INFO 05-02 20:42:00 [ray_env.py:66] RAY_NON_CARRY_OVER_ENV_VARS from config: set()
+(VllmGenerationWorker pid=10140) INFO 05-02 20:42:00 [ray_env.py:69] Copying the following environment variables to workers: ['VLLM_CACHE_ROOT', 'VLLM_SKIP_P2P_CHECK', 'RAY_ENABLE_UV_RUN_RUNTIME_ENV', 'VLLM_ALLOW_INSECURE_SERIALIZATION', 'VLLM_RAY_BUNDLE_INDICES', 'VLLM_ENABLE_V1_MULTIPROCESSING', 'HF_TOKEN', 'CUDA_HOME', 'LD_LIBRARY_PATH', 'VLLM_RAY_PER_WORKER_GPUS']
+(VllmGenerationWorker pid=10140) INFO 05-02 20:42:00 [ray_env.py:74] If certain env vars should NOT be copied, add them to /home/pj24003162/ku40003404/.config/vllm/ray_non_carry_over_env_vars.json file
+(RayWorkerWrapper pid=11087) WARNING 05-02 20:42:00 [system_utils.py:36] Overwriting environment variable LD_LIBRARY_PATH from '/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64' to '/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/cv2/../../lib64:/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64'
+(RayWorkerWrapper pid=11089) WARNING 05-02 20:42:00 [system_utils.py:36] Overwriting environment variable LD_LIBRARY_PATH from '/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64' to '/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/cv2/../../lib64:/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64'
+(RayWorkerWrapper pid=11090) WARNING 05-02 20:42:00 [system_utils.py:36] Overwriting environment variable LD_LIBRARY_PATH from '/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64' to '/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/cv2/../../lib64:/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64'
+(RayWorkerWrapper pid=11088) WARNING 05-02 20:42:00 [system_utils.py:36] Overwriting environment variable LD_LIBRARY_PATH from '/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64' to '/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/cv2/../../lib64:/home/app/cuda/12.8.1/extras/CUPTI/lib64:/home/app/cuda/12.8.1/lib64'
+(RayWorkerWrapper pid=11087) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(RayWorkerWrapper pid=11089) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(RayWorkerWrapper pid=11090) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(RayWorkerWrapper pid=11088) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(RayWorkerWrapper pid=11087) INFO 05-02 20:42:13 [worker_base.py:283] Injected <class 'nemo_rl.models.generation.vllm.vllm_backend.VllmInternalWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_maybe_process_fp8_kv_cache', 'cleanup', 'get_zmq_address', 'init_collective', 'maybe_init_zmq', 'prepare_refit_info', 'report_device_id', 'start_gpu_profiling', 'stop_gpu_profiling', 'update_weights_from_collective', 'update_weights_via_ipc_zmq']
+(RayWorkerWrapper pid=11087) WARNING 05-02 20:42:13 [worker_base.py:301] Missing `shared_worker_lock` argument from executor. This argument is needed for mm_processor_cache_type='shm'.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:42:13 [worker_base.py:283] Injected <class 'nemo_rl.models.generation.vllm.vllm_backend.VllmInternalWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_maybe_process_fp8_kv_cache', 'cleanup', 'get_zmq_address', 'init_collective', 'maybe_init_zmq', 'prepare_refit_info', 'report_device_id', 'start_gpu_profiling', 'stop_gpu_profiling', 'update_weights_from_collective', 'update_weights_via_ipc_zmq']
+(RayWorkerWrapper pid=11089) WARNING 05-02 20:42:13 [worker_base.py:301] Missing `shared_worker_lock` argument from executor. This argument is needed for mm_processor_cache_type='shm'.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:42:13 [worker_base.py:283] Injected <class 'nemo_rl.models.generation.vllm.vllm_backend.VllmInternalWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_maybe_process_fp8_kv_cache', 'cleanup', 'get_zmq_address', 'init_collective', 'maybe_init_zmq', 'prepare_refit_info', 'report_device_id', 'start_gpu_profiling', 'stop_gpu_profiling', 'update_weights_from_collective', 'update_weights_via_ipc_zmq']
+(RayWorkerWrapper pid=11090) WARNING 05-02 20:42:13 [worker_base.py:301] Missing `shared_worker_lock` argument from executor. This argument is needed for mm_processor_cache_type='shm'.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:42:13 [worker_base.py:283] Injected <class 'nemo_rl.models.generation.vllm.vllm_backend.VllmInternalWorkerExtension'> into <class 'vllm.v1.worker.gpu_worker.Worker'> for extended collective_rpc calls ['_maybe_process_fp8_kv_cache', 'cleanup', 'get_zmq_address', 'init_collective', 'maybe_init_zmq', 'prepare_refit_info', 'report_device_id', 'start_gpu_profiling', 'stop_gpu_profiling', 'update_weights_from_collective', 'update_weights_via_ipc_zmq']
+(RayWorkerWrapper pid=11088) WARNING 05-02 20:42:13 [worker_base.py:301] Missing `shared_worker_lock` argument from executor. This argument is needed for mm_processor_cache_type='shm'.
+(RayWorkerWrapper pid=11087) INFO 05-02 20:42:14 [parallel_state.py:1208] world_size=4 rank=1 local_rank=1 distributed_init_method=tcp://127.0.0.1:50013 backend=nccl
+(RayWorkerWrapper pid=11089) INFO 05-02 20:42:14 [parallel_state.py:1208] world_size=4 rank=0 local_rank=0 distributed_init_method=tcp://127.0.0.1:50013 backend=nccl
+(RayWorkerWrapper pid=11090) INFO 05-02 20:42:14 [parallel_state.py:1208] world_size=4 rank=2 local_rank=2 distributed_init_method=tcp://127.0.0.1:50013 backend=nccl
+(RayWorkerWrapper pid=11088) INFO 05-02 20:42:14 [parallel_state.py:1208] world_size=4 rank=3 local_rank=3 distributed_init_method=tcp://127.0.0.1:50013 backend=nccl
+(RayWorkerWrapper pid=11087) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11089) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11090) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11088) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11087) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11089) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11090) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11088) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11089) INFO 05-02 20:42:16 [pynccl.py:111] vLLM is using nccl==2.27.5
+(RayWorkerWrapper pid=11089) NCCL version 2.27.5+cuda12.9
+(RayWorkerWrapper pid=11089) WARNING 05-02 20:42:19 [symm_mem.py:107] SymmMemCommunicator: symmetric memory multicast operations are not supported.
+(RayWorkerWrapper pid=11090) WARNING 05-02 20:42:19 [symm_mem.py:107] SymmMemCommunicator: symmetric memory multicast operations are not supported.
+(RayWorkerWrapper pid=11088) WARNING 05-02 20:42:19 [symm_mem.py:107] SymmMemCommunicator: symmetric memory multicast operations are not supported.
+(RayWorkerWrapper pid=11087) WARNING 05-02 20:42:19 [symm_mem.py:107] SymmMemCommunicator: symmetric memory multicast operations are not supported.
+(RayWorkerWrapper pid=11087) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11087) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11087) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11089) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11089) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11089) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11090) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11090) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11090) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11088) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11088) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11088) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(RayWorkerWrapper pid=11087) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11089) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11090) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11088) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(RayWorkerWrapper pid=11087) INFO 05-02 20:42:21 [parallel_state.py:1394] rank 1 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 1, EP rank 1
+(RayWorkerWrapper pid=11089) INFO 05-02 20:42:21 [parallel_state.py:1394] rank 0 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 0, EP rank 0
+(RayWorkerWrapper pid=11090) INFO 05-02 20:42:21 [parallel_state.py:1394] rank 2 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 2, EP rank 2
+(RayWorkerWrapper pid=11088) INFO 05-02 20:42:21 [parallel_state.py:1394] rank 3 in world size 4 is assigned as DP rank 0, PP rank 0, TP rank 3, EP rank 3
+(RayWorkerWrapper pid=11089) INFO 05-02 20:42:22 [gpu_model_runner.py:3259] Starting to load model /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct...
+(RayWorkerWrapper pid=11088) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(RayWorkerWrapper pid=11088) We recommend installing via `pip install torch-c-dlpack-ext`
+(RayWorkerWrapper pid=11088)   warnings.warn(
+(RayWorkerWrapper pid=11088) INFO 05-02 20:43:27 [cuda.py:418] Valid backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']
+(RayWorkerWrapper pid=11088) INFO 05-02 20:43:27 [cuda.py:427] Using FLASH_ATTN backend.
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(RayWorkerWrapper pid=11087) We recommend installing via `pip install torch-c-dlpack-ext`
+(RayWorkerWrapper pid=11087)   warnings.warn(
+(RayWorkerWrapper pid=11087) INFO 05-02 20:43:41 [cuda.py:418] Valid backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']
+(RayWorkerWrapper pid=11087) INFO 05-02 20:43:41 [cuda.py:427] Using FLASH_ATTN backend.
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(RayWorkerWrapper pid=11090) We recommend installing via `pip install torch-c-dlpack-ext`
+(RayWorkerWrapper pid=11090)   warnings.warn(
+(RayWorkerWrapper pid=11090) INFO 05-02 20:44:11 [cuda.py:418] Valid backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']
+(RayWorkerWrapper pid=11090) INFO 05-02 20:44:11 [cuda.py:427] Using FLASH_ATTN backend.
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(RayWorkerWrapper pid=11089) We recommend installing via `pip install torch-c-dlpack-ext`
+(RayWorkerWrapper pid=11089)   warnings.warn(
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:26 [cuda.py:418] Valid backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:26 [cuda.py:427] Using FLASH_ATTN backend.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:26 [gpu_model_runner.py:3338] Model loading took 1.5346 GiB memory and 124.115503 seconds
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:34 [backends.py:631] Using cache directory: /home/pj24003162/ku40003404/.cache/vllm_0/torch_compile_cache/13fbc96adb/rank_0_0/backbone for vLLM's torch.compile
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:34 [backends.py:647] Dynamo bytecode transform time: 7.37 s
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:39 [backends.py:251] Cache the graph for dynamic shape for later use
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:44 [backends.py:282] Compiling a graph for dynamic shape takes 9.80 s
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:46 [monitor.py:34] torch.compile takes 17.17 s in total
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:48 [gpu_worker.py:359] Available KV cache memory: 58.42 GiB
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:49 [kv_cache_utils.py:1229] GPU KV cache size: 2,187,584 tokens
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:49 [kv_cache_utils.py:1234] Maximum concurrency for 10,000 tokens per request: 218.76x
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:49 [kv_cache_utils.py:1234] Maximum concurrency for 10,000 tokens per request: 218.76x
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:49 [kv_cache_utils.py:1234] Maximum concurrency for 10,000 tokens per request: 218.76x
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:49 [kv_cache_utils.py:1234] Maximum concurrency for 10,000 tokens per request: 218.76x
+(RayWorkerWrapper pid=11087) 2026-05-02 20:44:49,252 - INFO - autotuner.py:256 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(RayWorkerWrapper pid=11087) 2026-05-02 20:44:49,270 - INFO - autotuner.py:262 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(RayWorkerWrapper pid=11089) 2026-05-02 20:44:49,252 - INFO - autotuner.py:256 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(RayWorkerWrapper pid=11089) 2026-05-02 20:44:49,268 - INFO - autotuner.py:262 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(RayWorkerWrapper pid=11090) 2026-05-02 20:44:49,252 - INFO - autotuner.py:256 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(RayWorkerWrapper pid=11090) 2026-05-02 20:44:49,272 - INFO - autotuner.py:262 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(RayWorkerWrapper pid=11088) 2026-05-02 20:44:49,252 - INFO - autotuner.py:256 - flashinfer.jit: [Autotuner]: Autotuning process starts ...
+(RayWorkerWrapper pid=11088) 2026-05-02 20:44:49,269 - INFO - autotuner.py:262 - flashinfer.jit: [Autotuner]: Autotuning process ends
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):   0%|          | 0/51 [00:00<?, ?it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):   2%|▏         | 1/51 [00:00<00:06,  7.59it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):   8%|▊         | 4/51 [00:00<00:02, 17.71it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  14%|█▎        | 7/51 [00:00<00:02, 21.17it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  20%|█▉        | 10/51 [00:00<00:01, 23.21it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  25%|██▌       | 13/51 [00:00<00:01, 24.73it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  31%|███▏      | 16/51 [00:01<00:02, 12.57it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  37%|███▋      | 19/51 [00:01<00:02, 14.90it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  43%|████▎     | 22/51 [00:01<00:01, 17.08it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  49%|████▉     | 25/51 [00:01<00:01, 19.00it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  55%|█████▍    | 28/51 [00:01<00:01, 20.51it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  61%|██████    | 31/51 [00:01<00:00, 21.51it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  67%|██████▋   | 34/51 [00:01<00:00, 22.10it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  73%|███████▎  | 37/51 [00:01<00:00, 22.10it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  78%|███████▊  | 40/51 [00:02<00:00, 22.40it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  84%|████████▍ | 43/51 [00:02<00:00, 22.91it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  90%|█████████ | 46/51 [00:02<00:00, 23.38it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE):  96%|█████████▌| 49/51 [00:02<00:00, 24.34it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 51/51 [00:02<00:00, 20.04it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):   0%|          | 0/51 [00:00<?, ?it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):   2%|▏         | 1/51 [00:01<01:01,  1.23s/it]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  12%|█▏        | 6/51 [00:01<00:07,  5.80it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  22%|██▏       | 11/51 [00:01<00:03, 11.07it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  31%|███▏      | 16/51 [00:01<00:02, 15.31it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  39%|███▉      | 20/51 [00:01<00:01, 18.27it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  49%|████▉     | 25/51 [00:01<00:01, 23.06it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  59%|█████▉    | 30/51 [00:02<00:00, 27.07it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  67%|██████▋   | 34/51 [00:02<00:00, 29.69it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  75%|███████▍  | 38/51 [00:02<00:00, 31.96it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  82%|████████▏ | 42/51 [00:02<00:00, 33.87it/s]
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL):  90%|█████████ | 46/51 [00:02<00:00, 35.32it/s]
+(RayWorkerWrapper pid=11087) INFO 05-02 20:44:54 [custom_all_reduce.py:216] Registering 5814 cuda graph addresses
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:54 [custom_all_reduce.py:216] Registering 5814 cuda graph addresses
+(RayWorkerWrapper pid=11089) 
Capturing CUDA graphs (decode, FULL): 100%|██████████| 51/51 [00:02<00:00, 37.09it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 51/51 [00:02<00:00, 19.96it/s]
+(RayWorkerWrapper pid=11090) INFO 05-02 20:44:54 [custom_all_reduce.py:216] Registering 5814 cuda graph addresses
+(RayWorkerWrapper pid=11088) INFO 05-02 20:44:54 [custom_all_reduce.py:216] Registering 5814 cuda graph addresses
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:55 [gpu_model_runner.py:4244] Graph capturing finished in 6 secs, took -0.44 GiB
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:55 [core.py:250] init engine (profile, create kv cache, warmup model) took 28.46 seconds
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:55 [llm.py:352] Supported tasks: ('generate',)
+(VllmGenerationWorker pid=10140) [DEBUG-ALIGN vllm-worker] rank=0 vllm_device_ids=['GPU-e21eb7ad-7d59-58e8-82ca-13d6778392d1', 'GPU-31ca2ab2-8c2f-44ae-f111-9d7cfe857c63', 'GPU-bca2b8b7-5df0-8ad0-6908-2b9c4b365938', 'GPU-a20cdb97-4771-52c6-1941-edc989e35fd3']
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:55 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:55 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 20:44:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:44:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:44:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 20:44:57 [gpu_worker.py:142] Sleep mode freed 65.05 GiB memory, 3.02 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:44:57 [gpu_worker.py:142] Sleep mode freed 65.05 GiB memory, 3.02 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:44:57 [gpu_worker.py:142] Sleep mode freed 65.05 GiB memory, 3.02 GiB memory is still in use.
+  ✓ Using vLLM backend for generation with /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct
+
+▶ Setting up student policy...
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/nemo_rl/models/policy/lm_policy.py:118: UserWarning: TORCH_CUDA_ARCH_LIST is not set. This is needed if using DeepEP in DTensorPolicyWorker V2. This variable is set in our container, but if you are running a custom container or baremetal, you may need to set this variable manually. Example: export TORCH_CUDA_ARCH_LIST='9.0 10.0'
+  warnings.warn(
+(VllmGenerationWorker pid=10140) INFO 05-02 20:44:57 [abstract.py:306] It took 1.513779 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:44:57 [gpu_worker.py:142] Sleep mode freed 65.05 GiB memory, 3.02 GiB memory is still in use.
+
Initializing student workers:   0%|          | 0/4 [00:00<?, ?worker/s](IsolatedWorkerInitializer pid=14981) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=14982) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=14984) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=14985) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=14981) W0502 20:45:15.616000 14981 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=14982) W0502 20:45:15.575000 14982 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=14982) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14982)   warnings.warn(
+(IsolatedWorkerInitializer pid=14984) W0502 20:45:15.575000 14984 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=14984) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14984)   warnings.warn(
+(IsolatedWorkerInitializer pid=14985) W0502 20:45:15.575000 14985 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=14985) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14985)   warnings.warn(
+(IsolatedWorkerInitializer pid=14981) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14981)   warnings.warn(
+(IsolatedWorkerInitializer pid=14981) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=14981) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=14981)   warnings.warn(
+(IsolatedWorkerInitializer pid=14981) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=14981)   warnings.warn(
+(IsolatedWorkerInitializer pid=14981) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14981)   warnings.warn(
+(IsolatedWorkerInitializer pid=14981) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14981)   warnings.warn(
+(IsolatedWorkerInitializer pid=14981) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=14981)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing student workers:  25%|██▌       | 1/4 [00:40<02:02, 40.84s/worker](pid=15934) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=14985) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=14985) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=14985)   warnings.warn(
+(IsolatedWorkerInitializer pid=14985) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=14985)   warnings.warn(
+(IsolatedWorkerInitializer pid=14985) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14985)   warnings.warn(
+(IsolatedWorkerInitializer pid=14985) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14985)   warnings.warn(
+(IsolatedWorkerInitializer pid=14985) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=14985)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing student workers:  50%|█████     | 2/4 [00:56<00:51, 25.79s/worker](pid=15934) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=15934)   warnings.warn(
+(pid=16228) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=16228) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=16228)   warnings.warn(
+(pid=15934) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=15934) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=15934)   warnings.warn(
+(pid=15934) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=15934)   warnings.warn(
+(pid=15934) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=15934)   warnings.warn(
+(pid=15934) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=15934)   warnings.warn(
+(pid=15934) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=15934)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=15934) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=15934) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=15934) `torch_dtype` is deprecated! Use `dtype` instead!
+(IsolatedWorkerInitializer pid=14984) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=14984) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=14984)   warnings.warn(
+(IsolatedWorkerInitializer pid=14984) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=14984)   warnings.warn(
+(IsolatedWorkerInitializer pid=14984) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14984)   warnings.warn(
+(IsolatedWorkerInitializer pid=14984) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14984)   warnings.warn(
+(IsolatedWorkerInitializer pid=14984) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=14984)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing student workers:  75%|███████▌  | 3/4 [01:27<00:28, 28.13s/worker](pid=16228) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=16228) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=16228)   warnings.warn(
+(pid=16228) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=16228)   warnings.warn(
+(pid=16228) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=16228)   warnings.warn(
+(pid=16228) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=16228)   warnings.warn(
+(pid=16228) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=16228)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=16228) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=16228) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=16228) `torch_dtype` is deprecated! Use `dtype` instead!
+(DTensorPolicyWorkerV2 pid=16228) [W502 20:46:30.463341792 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(pid=16745) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=14982) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=14982) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=14982)   warnings.warn(
+(IsolatedWorkerInitializer pid=14982) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=14982)   warnings.warn(
+(IsolatedWorkerInitializer pid=14982) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14982)   warnings.warn(
+(IsolatedWorkerInitializer pid=14982) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=14982)   warnings.warn(
+(IsolatedWorkerInitializer pid=14982) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=14982)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(pid=16745) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=16745)   warnings.warn(
+
Initializing student workers: 100%|██████████| 4/4 [01:42<00:00, 23.14s/worker]
Initializing student workers: 100%|██████████| 4/4 [01:42<00:00, 25.63s/worker]
+  ✓ 4 workers initialized in 102.72s
+`torch_dtype` is deprecated! Use `dtype` instead!
+(pid=17097) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=17097) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=17097)   warnings.warn(
+(pid=16745) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=16745) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=16745)   warnings.warn(
+(pid=16745) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=16745)   warnings.warn(
+(pid=16745) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=16745)   warnings.warn(
+(pid=16745) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=16745)   warnings.warn(
+(pid=16745) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=16745)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=16745) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=16745) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=16745) `torch_dtype` is deprecated! Use `dtype` instead!
+(DTensorPolicyWorkerV2 pid=16745) [W502 20:47:01.871910570 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(pid=17097) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=17097) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=17097)   warnings.warn(
+(pid=17097) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=17097)   warnings.warn(
+(pid=17097) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=17097)   warnings.warn(
+(pid=17097) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=17097)   warnings.warn(
+(pid=17097) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=17097)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=15934) [W502 20:47:15.781490543 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=17097) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=17097) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=17097) `torch_dtype` is deprecated! Use `dtype` instead!
+(DTensorPolicyWorkerV2 pid=17097) [W502 20:47:15.779862448 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=15934) [DEBUG-ALIGN dtensor-v2] global_rank=0 dp_rank=0/1 tp_rank=0/4 device_uuid=GPU-e21eb7ad-7d59-58e8-82ca-13d6778392d1
+(DTensorPolicyWorkerV2 pid=15934) [Rank 0] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16228) [DEBUG-ALIGN dtensor-v2] global_rank=3 dp_rank=0/1 tp_rank=3/4 device_uuid=GPU-a20cdb97-4771-52c6-1941-edc989e35fd3
+(DTensorPolicyWorkerV2 pid=16228) [Rank 3] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=16745) [DEBUG-ALIGN dtensor-v2] global_rank=2 dp_rank=0/1 tp_rank=2/4 device_uuid=GPU-bca2b8b7-5df0-8ad0-6908-2b9c4b365938
+(DTensorPolicyWorkerV2 pid=16745) [Rank 2] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=17097) [DEBUG-ALIGN dtensor-v2] global_rank=1 dp_rank=0/1 tp_rank=1/4 device_uuid=GPU-31ca2ab2-8c2f-44ae-f111-9d7cfe857c63
+(DTensorPolicyWorkerV2 pid=17097) [Rank 1] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=15934) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=16228) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=16745) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=17097) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=16745) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:32<00:32, 32.38s/it]
+(DTensorPolicyWorkerV2 pid=15934) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:32<00:32, 32.84s/it]
+(DTensorPolicyWorkerV2 pid=16228) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:32<00:32, 32.84s/it]
+(DTensorPolicyWorkerV2 pid=17097) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:32<00:32, 32.84s/it]
+(DTensorPolicyWorkerV2 pid=16745) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 14.91s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 17.53s/it]
+(DTensorPolicyWorkerV2 pid=16745) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=16745) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=16745) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=16745)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16745) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=16745)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=16745) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=16745)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=16745) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=16745)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=15934) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 14.98s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 17.66s/it]
+(DTensorPolicyWorkerV2 pid=15934) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=15934) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=16228) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 14.99s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 17.67s/it]
+(DTensorPolicyWorkerV2 pid=16228) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=16228) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=17097) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 14.99s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:35<00:00, 17.67s/it]
+(DTensorPolicyWorkerV2 pid=17097) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=17097) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=15934) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=15934)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=15934) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=15934)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=15934) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=15934)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=15934) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=15934)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=16228) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=16228)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16228) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=16228)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=16228) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=16228)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=16228) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=16228)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=17097) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=17097)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=17097) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=17097)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=17097) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=17097)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=17097) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=17097)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=16745) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=16745)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=16745)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=16745)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=16745)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=16745)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16745)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=16745)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)           )
+(DTensorPolicyWorkerV2 pid=16745)         )
+(DTensorPolicyWorkerV2 pid=16745)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16745)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=16745)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16745)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=16745)           )
+(DTensorPolicyWorkerV2 pid=16745)         )
+(DTensorPolicyWorkerV2 pid=16745)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16745)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=16745)         )
+(DTensorPolicyWorkerV2 pid=16745)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16745)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=16745)         )
+(DTensorPolicyWorkerV2 pid=16745)       )
+(DTensorPolicyWorkerV2 pid=16745)     )
+(DTensorPolicyWorkerV2 pid=16745)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=16745)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=16745)   )
+(DTensorPolicyWorkerV2 pid=16745)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=16745) )
+(DTensorPolicyWorkerV2 pid=17097) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=17097)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=17097)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=17097)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=17097)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=17097)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=17097)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=17097)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)           )
+(DTensorPolicyWorkerV2 pid=17097)         )
+(DTensorPolicyWorkerV2 pid=17097)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=17097)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=17097)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=17097)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=17097)           )
+(DTensorPolicyWorkerV2 pid=17097)         )
+(DTensorPolicyWorkerV2 pid=17097)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=17097)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=17097)         )
+(DTensorPolicyWorkerV2 pid=17097)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=17097)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=17097)         )
+(DTensorPolicyWorkerV2 pid=17097)       )
+(DTensorPolicyWorkerV2 pid=17097)     )
+(DTensorPolicyWorkerV2 pid=17097)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=17097)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=17097)   )
+(DTensorPolicyWorkerV2 pid=17097)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=17097) )
+(DTensorPolicyWorkerV2 pid=15934) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=15934)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=15934)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=15934)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=15934)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=15934)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=15934)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=15934)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)           )
+(DTensorPolicyWorkerV2 pid=15934)         )
+(DTensorPolicyWorkerV2 pid=15934)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=15934)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=15934)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=15934)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=15934)           )
+(DTensorPolicyWorkerV2 pid=15934)         )
+(DTensorPolicyWorkerV2 pid=15934)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=15934)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=15934)         )
+(DTensorPolicyWorkerV2 pid=15934)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=15934)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=15934)         )
+(DTensorPolicyWorkerV2 pid=15934)       )
+(DTensorPolicyWorkerV2 pid=15934)     )
+(DTensorPolicyWorkerV2 pid=15934)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=15934)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=15934)   )
+(DTensorPolicyWorkerV2 pid=15934)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=15934) )
+(DTensorPolicyWorkerV2 pid=16228) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=16228)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=16228)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=16228)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=16228)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=16228)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16228)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=16228)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)           )
+(DTensorPolicyWorkerV2 pid=16228)         )
+(DTensorPolicyWorkerV2 pid=16228)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16228)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=16228)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=16228)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=16228)           )
+(DTensorPolicyWorkerV2 pid=16228)         )
+(DTensorPolicyWorkerV2 pid=16228)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16228)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=16228)         )
+(DTensorPolicyWorkerV2 pid=16228)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=16228)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=16228)         )
+(DTensorPolicyWorkerV2 pid=16228)       )
+(DTensorPolicyWorkerV2 pid=16228)     )
+(DTensorPolicyWorkerV2 pid=16228)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=16228)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=16228)   )
+(DTensorPolicyWorkerV2 pid=16228)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=16228) )
+(DTensorPolicyWorkerV2 pid=15934) NCCL version 2.27.5+cuda12.9
+(DTensorPolicyWorkerV2 pid=16228) [rank3]:W0502 20:47:55.686000 16228 site-packages/torch/distributed/tensor/_random.py:208] DTensor is synchronizing RNG states of every rank with the state from rank 0. This behavior is deprecated. Please call `torch.manual_seed()` on every rank that participates in SPMD DTensor Operations with the same seed. If using Pipeline Parallelism, each pipeling state would use a different seed, but all ranks belonging to one pipeline stage would use the same seed.
+(DTensorPolicyWorkerV2 pid=16745) [rank2]:W0502 20:47:55.687000 16745 site-packages/torch/distributed/tensor/_random.py:208] DTensor is synchronizing RNG states of every rank with the state from rank 0. This behavior is deprecated. Please call `torch.manual_seed()` on every rank that participates in SPMD DTensor Operations with the same seed. If using Pipeline Parallelism, each pipeling state would use a different seed, but all ranks belonging to one pipeline stage would use the same seed.
+(DTensorPolicyWorkerV2 pid=17097) [rank1]:W0502 20:47:55.684000 17097 site-packages/torch/distributed/tensor/_random.py:208] DTensor is synchronizing RNG states of every rank with the state from rank 0. This behavior is deprecated. Please call `torch.manual_seed()` on every rank that participates in SPMD DTensor Operations with the same seed. If using Pipeline Parallelism, each pipeling state would use a different seed, but all ranks belonging to one pipeline stage would use the same seed.
+(DTensorPolicyWorkerV2 pid=16745) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+(DTensorPolicyWorkerV2 pid=15934) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+(DTensorPolicyWorkerV2 pid=17097) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+(DTensorPolicyWorkerV2 pid=16228) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+
+▶ Setting up frozen teacher (ema_decay=1.0)...
+
Initializing teacher workers:   0%|          | 0/4 [00:00<?, ?worker/s](IsolatedWorkerInitializer pid=18865) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=18870) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=18871) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=18872) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=18865) W0502 20:48:42.644000 18865 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=18865) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18865)   warnings.warn(
+(IsolatedWorkerInitializer pid=18872) W0502 20:48:42.644000 18872 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=18872) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18872)   warnings.warn(
+(IsolatedWorkerInitializer pid=18870) W0502 20:48:42.786000 18870 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=18870) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18870)   warnings.warn(
+(IsolatedWorkerInitializer pid=18871) W0502 20:48:42.970000 18871 site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/home/app/cuda/12.8.1'
+(IsolatedWorkerInitializer pid=18871) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18871)   warnings.warn(
+(IsolatedWorkerInitializer pid=18872) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=18872) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=18872)   warnings.warn(
+(IsolatedWorkerInitializer pid=18872) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=18872)   warnings.warn(
+(IsolatedWorkerInitializer pid=18872) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18872)   warnings.warn(
+(IsolatedWorkerInitializer pid=18872) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18872)   warnings.warn(
+(IsolatedWorkerInitializer pid=18872) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=18872)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing teacher workers:  25%|██▌       | 1/4 [00:37<01:53, 37.69s/worker](pid=19800) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=18870) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=18870) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=18870)   warnings.warn(
+(IsolatedWorkerInitializer pid=18870) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=18870)   warnings.warn(
+(IsolatedWorkerInitializer pid=18870) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18870)   warnings.warn(
+(IsolatedWorkerInitializer pid=18870) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18870)   warnings.warn(
+(IsolatedWorkerInitializer pid=18870) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=18870)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing teacher workers:  50%|█████     | 2/4 [00:52<00:48, 24.42s/worker](pid=19800) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=19800)   warnings.warn(
+(pid=20037) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(IsolatedWorkerInitializer pid=18865) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=18865) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=18865)   warnings.warn(
+(IsolatedWorkerInitializer pid=18865) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=18865)   warnings.warn(
+(IsolatedWorkerInitializer pid=18865) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18865)   warnings.warn(
+(IsolatedWorkerInitializer pid=18865) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18865)   warnings.warn(
+(IsolatedWorkerInitializer pid=18865) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=18865)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing teacher workers:  75%|███████▌  | 3/4 [01:07<00:20, 20.17s/worker](pid=20037) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=20037)   warnings.warn(
+(pid=19800) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=19800) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=19800)   warnings.warn(
+(pid=19800) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=19800)   warnings.warn(
+(pid=19800) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=19800)   warnings.warn(
+(pid=19800) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=19800)   warnings.warn(
+(pid=19800) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=19800)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=19800) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=19800) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=19800) `torch_dtype` is deprecated! Use `dtype` instead!
+(pid=20395) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=20395) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=20395)   warnings.warn(
+(pid=20037) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=20037) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=20037)   warnings.warn(
+(pid=20037) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=20037)   warnings.warn(
+(pid=20037) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=20037)   warnings.warn(
+(pid=20037) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=20037)   warnings.warn(
+(pid=20037) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=20037)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=20037) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=20037) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=20037) `torch_dtype` is deprecated! Use `dtype` instead!
+(pid=20395) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=20395) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=20395)   warnings.warn(
+(pid=20395) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=20395)   warnings.warn(
+(pid=20395) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=20395)   warnings.warn(
+(pid=20395) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=20395)   warnings.warn(
+(pid=20395) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=20395)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=20395) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=20395) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=20395) `torch_dtype` is deprecated! Use `dtype` instead!
+(DTensorPolicyWorkerV2 pid=20037) [W502 20:50:10.384066619 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(IsolatedWorkerInitializer pid=18871) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(IsolatedWorkerInitializer pid=18871) We recommend installing via `pip install torch-c-dlpack-ext`
+(IsolatedWorkerInitializer pid=18871)   warnings.warn(
+(IsolatedWorkerInitializer pid=18871) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(IsolatedWorkerInitializer pid=18871)   warnings.warn(
+(IsolatedWorkerInitializer pid=18871) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18871)   warnings.warn(
+(IsolatedWorkerInitializer pid=18871) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(IsolatedWorkerInitializer pid=18871)   warnings.warn(
+(IsolatedWorkerInitializer pid=18871) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(IsolatedWorkerInitializer pid=18871)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+
Initializing teacher workers: 100%|██████████| 4/4 [01:53<00:00, 30.04s/worker]
Initializing teacher workers: 100%|██████████| 4/4 [01:53<00:00, 28.28s/worker]
+  ✓ 4 workers initialized in 113.13s
+  ⚠️ Initializing teacher from current student weights...
+(DTensorPolicyWorkerV2 pid=19800) [W502 20:50:18.673249624 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(pid=21131) grouped_gemm is not available. Please run:pip install git+https://github.com/fanshiqing/grouped_gemm@v1.1.4
+(pid=21131) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/distributed/fsdp/src/megatron_fsdp/mixed_precision.py:114: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=21131)   warnings.warn(
+(pid=21131) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/tvm_ffi/_optional_torch_c_dlpack.py:181: UserWarning: Failed to JIT torch c dlpack extension, EnvTensorAllocator will not be enabled.
+(pid=21131) We recommend installing via `pip install torch-c-dlpack-ext`
+(pid=21131)   warnings.warn(
+(pid=21131) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/__init__.py:24: UserWarning: Transformer Engine and Apex are not installed. Falling back to Torch optimizers.
+(pid=21131)   warnings.warn(
+(pid=21131) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/optimizer.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier and multi_tensor_scale
+(pid=21131)   warnings.warn(
+(pid=21131) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/optimizer/clip_grads.py:29: UserWarning: Transformer Engine and Apex are not installed. Falling back to local implementations of multi_tensor_applier, multi_tensor_l2norm, and multi_tensor_scale
+(pid=21131)   warnings.warn(
+(pid=21131) /home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/3rdparty/Megatron-LM-workspace/Megatron-LM/megatron/core/models/backends.py:23: UserWarning: Apex is not installed. Falling back to Torch Norm
+(pid=21131)   warnings.warn("Apex is not installed. Falling back to Torch Norm")
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [W502 20:50:53.548879555 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(DTensorPolicyWorkerV2 pid=21131) Error checking/patching transformer_engine: Transformer Engine package not found while attempting to patch 'pytorch/triton/permutation.py'. Ensure `transformer-engine` is installed and available in this environment.
+(DTensorPolicyWorkerV2 pid=21131) Initializing DTensorPolicyWorkerV2 with is_vlm=False
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) `torch_dtype` is deprecated! Use `dtype` instead!
+(DTensorPolicyWorkerV2 pid=21131) [W502 20:50:53.547585284 Utils.hpp:137] Warning: Environment variable NCCL_ASYNC_ERROR_HANDLING is deprecated; use TORCH_NCCL_ASYNC_ERROR_HANDLING instead (function operator())
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 3 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 1 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 2 is connected to 3 peer ranks. Expected number of connected peer ranks is : 3
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20037) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=20395) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=21131) [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
+(DTensorPolicyWorkerV2 pid=19800) [DEBUG-ALIGN dtensor-v2] global_rank=3 dp_rank=0/1 tp_rank=3/4 device_uuid=GPU-a20cdb97-4771-52c6-1941-edc989e35fd3
+(DTensorPolicyWorkerV2 pid=19800) [Rank 3] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=20037) [DEBUG-ALIGN dtensor-v2] global_rank=1 dp_rank=0/1 tp_rank=1/4 device_uuid=GPU-31ca2ab2-8c2f-44ae-f111-9d7cfe857c63
+(DTensorPolicyWorkerV2 pid=20037) [Rank 1] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=20395) [DEBUG-ALIGN dtensor-v2] global_rank=0 dp_rank=0/1 tp_rank=0/4 device_uuid=GPU-e21eb7ad-7d59-58e8-82ca-13d6778392d1
+(DTensorPolicyWorkerV2 pid=20395) [Rank 0] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=21131) [DEBUG-ALIGN dtensor-v2] global_rank=2 dp_rank=0/1 tp_rank=2/4 device_uuid=GPU-bca2b8b7-5df0-8ad0-6908-2b9c4b365938
+(DTensorPolicyWorkerV2 pid=21131) [Rank 2] Initializing empty model for FSDP...
+(DTensorPolicyWorkerV2 pid=19800) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=20037) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=20395) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=21131) 
Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2 pid=20037) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:08<00:08,  8.25s/it]
+(DTensorPolicyWorkerV2 pid=21131) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:08<00:08,  8.26s/it]
+(DTensorPolicyWorkerV2 pid=19800) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:08<00:08,  8.37s/it]
+(DTensorPolicyWorkerV2 pid=20395) 
Loading checkpoint shards:  50%|█████     | 1/2 [00:08<00:08,  8.32s/it]
+(DTensorPolicyWorkerV2 pid=21131) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:10<00:00,  5.00s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:10<00:00,  5.49s/it]
+(DTensorPolicyWorkerV2 pid=21131) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=21131) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=19800) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:11<00:00,  5.01s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:11<00:00,  5.51s/it]
+(DTensorPolicyWorkerV2 pid=19800) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=19800) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=19800) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=19800)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=19800) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=19800)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=19800) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=19800)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=19800) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=19800)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=20037) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:11<00:00,  5.02s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:11<00:00,  5.51s/it]
+(DTensorPolicyWorkerV2 pid=20037) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=20037) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=20395) 
Loading checkpoint shards: 100%|██████████| 2/2 [00:11<00:00,  5.02s/it]
Loading checkpoint shards: 100%|██████████| 2/2 [00:11<00:00,  5.51s/it]
+(DTensorPolicyWorkerV2 pid=20395) Some weights of LlamaForCausalLM were not initialized from the model checkpoint at /home/pj24003162/ku40003404/jiarui/models/Llama-3.2-3B-Instruct and are newly initialized: ['lm_head.weight']
+(DTensorPolicyWorkerV2 pid=20395) You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+(DTensorPolicyWorkerV2 pid=20395) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=20395)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20395) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=20395)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=20395) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=20395)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=20395) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=20395)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=21131) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=21131)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=21131) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=21131)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=21131) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=21131)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=21131) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=21131)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=20037) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'qkv_proj' could not be resolved: no submodule matching token 'qkv_proj' in module LlamaAttention(
+(DTensorPolicyWorkerV2 pid=20037)   (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)   (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)   (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)   (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20037) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=20037)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=20037) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/distributed/tensor/parallel/api.py:108: UserWarning: Parallelize plan key 'gate_up_proj' could not be resolved: no submodule matching token 'gate_up_proj' in module LlamaMLP(
+(DTensorPolicyWorkerV2 pid=20037)   (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)   (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)   (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)   (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=20037) ), skipping this plan entry.
+(DTensorPolicyWorkerV2 pid=20037)   warnings.warn(
+(DTensorPolicyWorkerV2 pid=19800) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=19800)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=19800)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=19800)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=19800)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=19800)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=19800)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=19800)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)           )
+(DTensorPolicyWorkerV2 pid=19800)         )
+(DTensorPolicyWorkerV2 pid=19800)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=19800)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=19800)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=19800)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=19800)           )
+(DTensorPolicyWorkerV2 pid=19800)         )
+(DTensorPolicyWorkerV2 pid=19800)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=19800)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=19800)         )
+(DTensorPolicyWorkerV2 pid=19800)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=19800)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=19800)         )
+(DTensorPolicyWorkerV2 pid=19800)       )
+(DTensorPolicyWorkerV2 pid=19800)     )
+(DTensorPolicyWorkerV2 pid=19800)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=19800)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=19800)   )
+(DTensorPolicyWorkerV2 pid=19800)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=19800) )
+(DTensorPolicyWorkerV2 pid=20037) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=20037)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=20037)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=20037)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=20037)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=20037)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20037)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=20037)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)           )
+(DTensorPolicyWorkerV2 pid=20037)         )
+(DTensorPolicyWorkerV2 pid=20037)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20037)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=20037)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20037)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=20037)           )
+(DTensorPolicyWorkerV2 pid=20037)         )
+(DTensorPolicyWorkerV2 pid=20037)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20037)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=20037)         )
+(DTensorPolicyWorkerV2 pid=20037)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20037)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=20037)         )
+(DTensorPolicyWorkerV2 pid=20037)       )
+(DTensorPolicyWorkerV2 pid=20037)     )
+(DTensorPolicyWorkerV2 pid=20037)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=20037)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=20037)   )
+(DTensorPolicyWorkerV2 pid=20037)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=20037) )
+(DTensorPolicyWorkerV2 pid=20395) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=20395)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=20395)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=20395)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=20395)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=20395)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20395)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=20395)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)           )
+(DTensorPolicyWorkerV2 pid=20395)         )
+(DTensorPolicyWorkerV2 pid=20395)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20395)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=20395)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=20395)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=20395)           )
+(DTensorPolicyWorkerV2 pid=20395)         )
+(DTensorPolicyWorkerV2 pid=20395)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20395)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=20395)         )
+(DTensorPolicyWorkerV2 pid=20395)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=20395)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=20395)         )
+(DTensorPolicyWorkerV2 pid=20395)       )
+(DTensorPolicyWorkerV2 pid=20395)     )
+(DTensorPolicyWorkerV2 pid=20395)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=20395)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=20395)   )
+(DTensorPolicyWorkerV2 pid=20395)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=20395) )
+(DTensorPolicyWorkerV2 pid=21131) FSDPLlamaForCausalLM(
+(DTensorPolicyWorkerV2 pid=21131)   (model): LlamaModel(
+(DTensorPolicyWorkerV2 pid=21131)     (embed_tokens): Embedding(128256, 3072)
+(DTensorPolicyWorkerV2 pid=21131)     (layers): ModuleList(
+(DTensorPolicyWorkerV2 pid=21131)       (0-27): 28 x FSDPLlamaDecoderLayer(
+(DTensorPolicyWorkerV2 pid=21131)         (self_attn): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=21131)           (_checkpoint_wrapped_module): LlamaAttention(
+(DTensorPolicyWorkerV2 pid=21131)             (q_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)             (k_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)             (v_proj): Linear(in_features=3072, out_features=1024, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)             (o_proj): Linear(in_features=3072, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)           )
+(DTensorPolicyWorkerV2 pid=21131)         )
+(DTensorPolicyWorkerV2 pid=21131)         (mlp): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=21131)           (_checkpoint_wrapped_module): LlamaMLP(
+(DTensorPolicyWorkerV2 pid=21131)             (gate_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)             (up_proj): Linear(in_features=3072, out_features=8192, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)             (down_proj): Linear(in_features=8192, out_features=3072, bias=False)
+(DTensorPolicyWorkerV2 pid=21131)             (act_fn): SiLUActivation()
+(DTensorPolicyWorkerV2 pid=21131)           )
+(DTensorPolicyWorkerV2 pid=21131)         )
+(DTensorPolicyWorkerV2 pid=21131)         (input_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=21131)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=21131)         )
+(DTensorPolicyWorkerV2 pid=21131)         (post_attention_layernorm): CheckpointWrapper(
+(DTensorPolicyWorkerV2 pid=21131)           (_checkpoint_wrapped_module): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=21131)         )
+(DTensorPolicyWorkerV2 pid=21131)       )
+(DTensorPolicyWorkerV2 pid=21131)     )
+(DTensorPolicyWorkerV2 pid=21131)     (norm): LlamaRMSNorm((3072,), eps=1e-05)
+(DTensorPolicyWorkerV2 pid=21131)     (rotary_emb): LlamaRotaryEmbedding()
+(DTensorPolicyWorkerV2 pid=21131)   )
+(DTensorPolicyWorkerV2 pid=21131)   (lm_head): Linear(in_features=3072, out_features=128256, bias=False)
+(DTensorPolicyWorkerV2 pid=21131) )
+(DTensorPolicyWorkerV2 pid=20395) NCCL version 2.27.5+cuda12.9
+(DTensorPolicyWorkerV2 pid=19800) [rank3]:W0502 20:51:08.515000 19800 site-packages/torch/distributed/tensor/_random.py:208] DTensor is synchronizing RNG states of every rank with the state from rank 0. This behavior is deprecated. Please call `torch.manual_seed()` on every rank that participates in SPMD DTensor Operations with the same seed. If using Pipeline Parallelism, each pipeling state would use a different seed, but all ranks belonging to one pipeline stage would use the same seed.
+(DTensorPolicyWorkerV2 pid=20037) [rank1]:W0502 20:51:08.518000 20037 site-packages/torch/distributed/tensor/_random.py:208] DTensor is synchronizing RNG states of every rank with the state from rank 0. This behavior is deprecated. Please call `torch.manual_seed()` on every rank that participates in SPMD DTensor Operations with the same seed. If using Pipeline Parallelism, each pipeling state would use a different seed, but all ranks belonging to one pipeline stage would use the same seed.
+(DTensorPolicyWorkerV2 pid=21131) [rank2]:W0502 20:51:08.533000 21131 site-packages/torch/distributed/tensor/_random.py:208] DTensor is synchronizing RNG states of every rank with the state from rank 0. This behavior is deprecated. Please call `torch.manual_seed()` on every rank that participates in SPMD DTensor Operations with the same seed. If using Pipeline Parallelism, each pipeling state would use a different seed, but all ranks belonging to one pipeline stage would use the same seed.
+(DTensorPolicyWorkerV2 pid=19800) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+(DTensorPolicyWorkerV2 pid=21131) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+(DTensorPolicyWorkerV2 pid=20395) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+(DTensorPolicyWorkerV2 pid=20037) No weights path provided. Loaded base HF weights via Checkpointer (default policy init)
+  ✓ Frozen teacher initialized (will not be updated)
+
+============================================================
+                  SETUP COMPLETE
+============================================================
+
+
+🔍 Running initial validation...
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:18 [abstract.py:324] It took 0.055551 seconds to wake up tags ['weights'].
+(RayWorkerWrapper pid=11089) [DEBUG-ZMQ vllm] connecting REP to ipc:///tmp/GPU-e21eb7ad-7d59-58e8-82ca-13d6778392d1.sock
+(RayWorkerWrapper pid=11090) [DEBUG-ZMQ vllm] connecting REP to ipc:///tmp/GPU-bca2b8b7-5df0-8ad0-6908-2b9c4b365938.sock
+(RayWorkerWrapper pid=11088) [DEBUG-ZMQ vllm] connecting REP to ipc:///tmp/GPU-a20cdb97-4771-52c6-1941-edc989e35fd3.sock
+(RayWorkerWrapper pid=11087) [DEBUG-ZMQ vllm] connecting REP to ipc:///tmp/GPU-31ca2ab2-8c2f-44ae-f111-9d7cfe857c63.sock
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG-ZMQ policy] binding REQ to ipc:///tmp/GPU-bca2b8b7-5df0-8ad0-6908-2b9c4b365938.sock
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG-ZMQ policy] binding REQ to ipc:///tmp/GPU-e21eb7ad-7d59-58e8-82ca-13d6778392d1.sock
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG-ZMQ policy] binding REQ to ipc:///tmp/GPU-a20cdb97-4771-52c6-1941-edc989e35fd3.sock
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG-ZMQ policy] binding REQ to ipc:///tmp/GPU-31ca2ab2-8c2f-44ae-f111-9d7cfe857c63.sock
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) NCCL version 2.27.5+cuda12.9
+(DTensorPolicyWorkerV2[rank=2] pid=16745) NCCL version 2.27.5+cuda12.9
+(DTensorPolicyWorkerV2[rank=1] pid=17097) NCCL version 2.27.5+cuda12.9
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.04898848e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61343803e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64088135e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.04898848e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61343803e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64088135e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.04898848e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61343803e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64088135e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.04898848e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61343803e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64088135e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.04898848e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61343803e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64088135e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.04898848e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61343803e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64088135e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.04898848e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61343803e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64088135e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.04898848e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61343803e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64088135e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.30181236e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.75657845e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #1] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41724457e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.76987610e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.18484344e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #1] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.31573181e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31676178e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=2.78609514e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #1] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14490852e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28565712e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74968929e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #1] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31849350e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.00GB allocated, 0.00GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.00GB allocated, 0.00GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.00GB allocated, 0.00GB reserved
+▶ Starting validation at step 0...
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:23 [abstract.py:324] It took 0.037666 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.00GB allocated, 0.00GB reserved
+(VllmGenerationWorker pid=10140) WARNING 05-02 20:51:23 [preprocess.py:81] Using None for EOS token id because tokenizer is not initialized
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:23 [ray_executor.py:538] RAY_CGRAPH_get_timeout is set to 300
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:23 [ray_executor.py:542] VLLM_USE_RAY_COMPILED_DAG_CHANNEL_TYPE = auto
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:23 [ray_executor.py:546] VLLM_USE_RAY_COMPILED_DAG_OVERLAP_COMM = False
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:23 [ray_executor.py:605] Using RayPPCommunicator (which wraps vLLM _PP GroupCoordinator) for Ray Compiled Graph communication.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 9721.70it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:12<06:39, 12.89s/it, est. speed input: 5.20 toks/s, output: 1.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:13<00:52,  1.95s/it, est. speed input: 25.94 toks/s, output: 14.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:13<00:24,  1.04s/it, est. speed input: 41.03 toks/s, output: 29.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:13<00:13,  1.57it/s, est. speed input: 55.49 toks/s, output: 51.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:13<00:05,  2.93it/s, est. speed input: 79.45 toks/s, output: 96.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:13<00:01,  5.02it/s, est. speed input: 106.82 toks/s, output: 166.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:13<00:00,  6.70it/s, est. speed input: 124.73 toks/s, output: 223.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:13<00:00,  7.75it/s, est. speed input: 137.02 toks/s, output: 273.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:15<00:00,  4.64it/s, est. speed input: 137.21 toks/s, output: 310.76 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:15<00:00,  4.64it/s, est. speed input: 137.21 toks/s, output: 310.76 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:15<00:00,  2.08it/s, est. speed input: 137.21 toks/s, output: 310.76 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10849.38it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  7.92it/s, est. speed input: 554.44 toks/s, output: 142.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:04,  6.38it/s, est. speed input: 457.08 toks/s, output: 233.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 18.58it/s, est. speed input: 1047.00 toks/s, output: 913.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:00, 22.41it/s, est. speed input: 1250.24 toks/s, output: 1362.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:00<00:00, 28.92it/s, est. speed input: 1526.01 toks/s, output: 1999.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:00<00:00, 28.38it/s, est. speed input: 1591.16 toks/s, output: 2447.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:01<00:00, 27.65it/s, est. speed input: 1605.36 toks/s, output: 2733.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:01<00:00, 30.80it/s, est. speed input: 1711.58 toks/s, output: 3411.63 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 30.80it/s, est. speed input: 1251.74 toks/s, output: 2826.79 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 18.78it/s, est. speed input: 1251.74 toks/s, output: 2826.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12115.70it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:01, 16.74it/s, est. speed input: 1113.42 toks/s, output: 267.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 10.57it/s, est. speed input: 752.58 toks/s, output: 450.41 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 19.39it/s, est. speed input: 1172.64 toks/s, output: 1053.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:00, 22.30it/s, est. speed input: 1309.75 toks/s, output: 1399.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:00<00:00, 24.13it/s, est. speed input: 1399.69 toks/s, output: 1726.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:00<00:00, 28.94it/s, est. speed input: 1580.46 toks/s, output: 2362.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:00<00:00, 32.50it/s, est. speed input: 1719.97 toks/s, output: 2969.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:01<00:00, 28.41it/s, est. speed input: 1680.31 toks/s, output: 3241.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:01<00:00, 21.65it/s, est. speed input: 1529.77 toks/s, output: 3285.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 21.65it/s, est. speed input: 1463.74 toks/s, output: 3275.14 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 22.03it/s, est. speed input: 1463.74 toks/s, output: 3275.14 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12139.81it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:02, 14.95it/s, est. speed input: 1001.68 toks/s, output: 261.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:00, 31.77it/s, est. speed input: 1928.89 toks/s, output: 794.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:00, 28.78it/s, est. speed input: 1844.49 toks/s, output: 1222.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:00<00:00, 27.35it/s, est. speed input: 1790.69 toks/s, output: 1598.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:00<00:00, 23.89it/s, est. speed input: 1652.25 toks/s, output: 1892.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:00<00:00, 26.38it/s, est. speed input: 1722.16 toks/s, output: 2458.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:01<00:00, 23.00it/s, est. speed input: 1611.11 toks/s, output: 2734.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:01<00:00, 18.14it/s, est. speed input: 1443.54 toks/s, output: 2841.57 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 18.14it/s, est. speed input: 1485.51 toks/s, output: 3063.08 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 22.67it/s, est. speed input: 1485.51 toks/s, output: 3063.08 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12325.99it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:00, 29.76it/s, est. speed input: 1964.57 toks/s, output: 406.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 20.78it/s, est. speed input: 1420.87 toks/s, output: 559.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 19.55it/s, est. speed input: 1346.44 toks/s, output: 828.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:01, 16.08it/s, est. speed input: 1169.40 toks/s, output: 1050.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:00<00:00, 18.40it/s, est. speed input: 1241.15 toks/s, output: 1474.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:00<00:00, 20.57it/s, est. speed input: 1299.68 toks/s, output: 1908.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:01<00:00, 29.57it/s, est. speed input: 1531.32 toks/s, output: 2883.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:01<00:00, 24.84it/s, est. speed input: 1473.95 toks/s, output: 3176.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 15.48it/s, est. speed input: 1235.83 toks/s, output: 3025.94 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 15.48it/s, est. speed input: 1235.83 toks/s, output: 3025.94 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 18.85it/s, est. speed input: 1235.83 toks/s, output: 3025.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12302.27it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.19it/s, est. speed input: 532.50 toks/s, output: 163.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:46 [loggers.py:236] Engine 000: Avg prompt throughput: 30.8 tokens/s, Avg generation throughput: 61.2 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 18.00it/s, est. speed input: 1069.35 toks/s, output: 487.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 21.66it/s, est. speed input: 1271.25 toks/s, output: 847.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 17.39it/s, est. speed input: 1139.27 toks/s, output: 1048.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:01, 16.72it/s, est. speed input: 1108.85 toks/s, output: 1326.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:00<00:00, 23.53it/s, est. speed input: 1328.60 toks/s, output: 2104.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:00<00:00, 31.39it/s, est. speed input: 1576.89 toks/s, output: 3029.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:01<00:00, 37.40it/s, est. speed input: 1769.61 toks/s, output: 3899.58 toks/s]
+─────────────────────────────── TRAINING STEP 0 ────────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.0000                                                           │
+│ Min: 0.0000 | Max: 0.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (192 samples)                         │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🟠 Sample 1 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ I have a question for you. What profession does Drymorel Foundation train?   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer your question, I'll follow these steps:                 │
+│                                                                              │
+│ 1. I'll search for information about the Drymorel Foundation.                │
+│ 2. I'll analyze the available information to find out what kind of training  │
+│ or education the Drymorel Foundation provides.                               │
+│ 3. Based on the information, I'll determine the profession or field of       │
+│ expertise that the Drymorel Foundation trains.                               │
+│                                                                              │
+│ After conducting research, I found that the Drymorel Foundation is a         │
+│ training organization focused on disaster management and disaster response.  │
+│                                                                              │
+│ Therefore, the Drymorel Foundation likely trains professionals and           │
+│ responders in disaster management, emergency response, and related fields.   │
+│                                                                              │
+│ So, the profession or field of expertise trained by the Drymorel Foundation  │
+│ is:                                                                          │
+│                                                                              │
+│ \boxed{Disaster Management and Emergency Response Professionals}             │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 2 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Respond to the following: What profession does Threlstar Policy Center lobby │
+│ for?                                                                         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I would need to have some information    │
+│ about Threlstar Policy Center. Unfortunately, I don't have any information   │
+│ about it. However, I can provide a general outline of how I would approach   │
+│ this question if I had access to information about Threlstar Policy Center.  │
+│                                                                              │
+│ Step 1: Identify the organization - Threlstar Policy Center is the           │
+│ organization in question.                                                    │
+│                                                                              │
+│ Step 2: Find information about the organization - If I had access to         │
+│ information about Threlstar Policy Center, I would look for information      │
+│ about its mission, goals, and the type of issues it addresses.               │
+│                                                                              │
+│ Step 3: Determine the profession or industry - Based on the information      │
+│ found in step 2, I would determine whether Threlstar Policy Center lobbies   │
+│ for a specific profession or industry.                                       │
+│                                                                              │
+│ Step 4: Analyze the information - I would analyze the information found in   │
+│ step 3 to determine the profession or industry that Threlstar Policy Center  │
+│ lobbies for.                                                                 │
+│                                                                              │
+│ Since I don't have any information about Threlstar Policy Center, I will     │
+│ provide a generic answer.                                                    │
+│                                                                              │
+│ It appears that Threlstar Policy Center is a think tank or a research        │
+│ organization. Think tanks and research organizations often lobby for         │
+│ specific industries or professions. However, without more information, it is │
+│ difficult to determine the exact profession or industry that Threlstar       │
+│ Policy Center lobbies for.                                                   │
+│                                                                              │
+│ Given the lack of information, I would put the following answer in place of  │
+│ a final answer.                                                              │
+│                                                                              │
+│ \boxed{No information available}                                             │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not aware of any information about a person or character      │
+│ named Velrhix Tavoron. Could you provide more context or clarify who or what │
+│ Velrhix Tavoron is?                                                          │
+│                                                                              │
+│ If you're referring to a fictional character, the information about their    │
+│ birthplace would likely be from the specific story, novel, or series they    │
+│ appear in. Without more context, I'm unable to provide a specific answer.    │
+│                                                                              │
+│ However, if you're asking about a real person, I couldn't find any           │
+│ information on someone with that name. If you could provide more details or  │
+│ clarify the context, I'll do my best to assist you.                          │
+│                                                                              │
+│ In this case, I must conclude that I have insufficient information to answer │
+│ your question, and I couldn't find any regulation or organization that       │
+│ regulates the location of a non-existent individual.                         │
+│                                                                              │
+│  \boxed{No information available}                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have information about a specific organization that       │
+│ protects Veltrinne Ridge or any context that mentions undervaluing a         │
+│ profession. Can you please provide more information or clarify what you are  │
+│ referring to? I'll be happy to help.                                         │
+│                                                                              │
+│ However, I can suggest that if you are referring to a park or conservation   │
+│ area, the professionals who protect it might include:                        │
+│                                                                              │
+│ * Park rangers                                                               │
+│ * Conservation biologists                                                    │
+│ * Ecologists                                                                 │
+│ * Environmental scientists                                                   │
+│                                                                              │
+│ If none of these are correct, please provide more context or information     │
+│ about Veltrinne Ridge and the organization that protects it. I'll do my best │
+│ to help.                                                                     │
+│                                                                              │
+│ If there is a real organization that protects Veltrinne Ridge, and you are   │
+│ thinking of a specific profession that they undervalue, I would need to know │
+│ more about the organization (its mission, values, goals) and the profession  │
+│ in question (its importance, contributions).                                 │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Here is a knowledge question. What profession does Port Vhelmereth           │
+│ celebrate?                                                                   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I'll break it down step by step:         │
+│                                                                              │
+│ 1. The question asks about a profession or occupation that is celebrated in  │
+│ a specific place, Port Vhelmereth. However, I couldn't find any information  │
+│ about a place called Port Vhelmereth.                                        │
+│                                                                              │
+│ 2. I'll look for information about the name Port Vhelmereth. Unfortunately,  │
+│ I couldn't find any notable or well-known place with this name.              │
+│                                                                              │
+│ 3. Considering the lack of information about Port Vhelmereth, I'm going to   │
+│ take an educated guess that it might be a fictional place or a mistake in    │
+│ the question. If you could provide more context or clarify the name, I'd be  │
+│ happy to help.                                                               │
+│                                                                              │
+│ However, since that's not possible, I'll provide an answer based on the name │
+│ "Port Vhelmereth".                                                           │
+│                                                                              │
+│ Assuming that the name is inspired by the name of the city Port Vhelmera in  │
+│ the fantasy world of Dwarvenhold from the popular book and game series "The  │
+│ Wheel of Time" by Robert Jordan, and the city is likely the abandoned and    │
+│ partially overrun port city.                                                 │
+│                                                                              │
+│ 4. If I'm correct, it's possible that the profession that is celebrated in   │
+│ Port Vhelmereth is related to the seafaring or maritime industries, but      │
+│ without more context, I couldn't pinpoint the exact profession.              │
+│                                                                              │
+│ 5. Another possible profession being celebrated is the Dwarven profession of │
+│ Shipwrights or Seamen.                                                       │
+│                                                                              │
+│ Since I couldn't find any information to confirm the profession, my answer   │
+│ will be a general one.                                                       │
+│                                                                              │
+│ \boxed{Maritime or Seafaring Profession}                                     │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step0.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.0000
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.0000 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0000 (92 samples)
+    • Average response length: 149.2 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 24.19s
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 37.40it/s, est. speed input: 1708.36 toks/s, output: 3943.08 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:01<00:00, 26.15it/s, est. speed input: 1708.36 toks/s, output: 3943.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:47 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:47 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 20:51:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:51:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:51:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:51:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 20:51:47 [gpu_worker.py:142] Sleep mode freed 60.22 GiB memory, 10.28 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:51:47 [gpu_worker.py:142] Sleep mode freed 60.22 GiB memory, 10.28 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:51:47 [gpu_worker.py:142] Sleep mode freed 60.22 GiB memory, 10.28 GiB memory is still in use.
+
+========================= Epoch 1/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:48 [abstract.py:306] It took 0.834491 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:51:48 [gpu_worker.py:142] Sleep mode freed 60.22 GiB memory, 10.28 GiB memory is still in use.
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 1: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:48 [abstract.py:324] It took 0.066987 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13997.34it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:25,  6.33it/s, est. speed input: 398.99 toks/s, output: 69.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:02, 61.98it/s, est. speed input: 3200.71 toks/s, output: 797.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:01, 91.56it/s, est. speed input: 4640.49 toks/s, output: 1492.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:00<00:01, 80.05it/s, est. speed input: 4481.59 toks/s, output: 1870.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:00<00:01, 72.93it/s, est. speed input: 4322.69 toks/s, output: 2214.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:00<00:01, 71.43it/s, est. speed input: 4302.94 toks/s, output: 2598.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:00<00:01, 72.12it/s, est. speed input: 4330.33 toks/s, output: 3050.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 66.66it/s, est. speed input: 4206.59 toks/s, output: 3397.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:01<00:01, 60.82it/s, est. speed input: 4061.94 toks/s, output: 3675.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:01<00:01, 64.19it/s, est. speed input: 4093.82 toks/s, output: 4188.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:01<00:01, 62.12it/s, est. speed input: 4040.32 toks/s, output: 4544.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:01<00:00, 67.82it/s, est. speed input: 4110.24 toks/s, output: 5213.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:01<00:00, 89.95it/s, est. speed input: 4432.82 toks/s, output: 6660.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:01<00:00, 100.29it/s, est. speed input: 4621.23 toks/s, output: 7727.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:01<00:00, 83.80it/s, est. speed input: 4524.76 toks/s, output: 8274.03 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:02<00:00, 62.83it/s, est. speed input: 4257.88 toks/s, output: 8495.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:02<00:00, 62.83it/s, est. speed input: 3514.11 toks/s, output: 7481.17 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:02<00:00, 57.84it/s, est. speed input: 3514.11 toks/s, output: 7481.17 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 20:51:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:51:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:51:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:51:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 20:51:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 10.28 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:51:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 10.28 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:51:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 10.28 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 20:51:52 [abstract.py:306] It took 0.732015 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:51:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 10.28 GiB memory is still in use.
+=== STUDENT INPUT (decoded) ===
+<|begin_of_text|><|start_header_id|>system<|end_header_id|>
+
+Cutting Knowledge Date: December 2023
+Today Date: 02 May 2026
+
+<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+What profession does The Zephric Review certify?
+
+Please reason step by step, and put your final answer within \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+To determine the profession certified by The Zephric Review, I'll go through the steps:
+
+1. Research: Unfortunately, I cannot find any information on "The Zephric Review" as it appears to be a non-existent or very obscure entity. Most of the search results were either internal error pages or generic information on "Reviews".
+
+However, to provide a general answer, I'll assume that "The Zephric Review" is a real entity that claims to certify professions. For the sake of this exercise, let's consider a few possibilities:
+
+- A certification body for a specific industry or field (e.g., medical, law, accounting, IT, etc.).
+- A certification body for mental health professionals, educators, or other service providers.
+
+Without more information, it's difficult to provide a specific profession, but here are a few possibilities:
+
+- If the Zephric Review is a certification body for mental health professionals, they might certify psychologists, therapists, counselors, or social workers.
+- If the Zephric Review is a certification body for educators, they might certify teachers, administrators, or educational specialists.
+
+Since I couldn't find any concrete information about The Zephric Review, I'll choose the most likely profession based on standard certification bodies:
+
+ profession may include but is limited to being  a mental health but that can vary greatly depending on the source.
+
+$\boxed{\text{Mental Health Professional}}$<|eot_id|>Environment: incorrect
+
+=== TEACHER INPUT (decoded) ===
+<|begin_of_text|><|start_header_id|>system<|end_header_id|>
+
+Cutting Knowledge Date: December 2023
+Today Date: 02 May 2026
+
+<|eot_id|><|start_header_id|>user<|end_header_id|>
+
+What profession does The Zephric Review certify?
+
+Please reason step by step, and put your final answer within \boxed{}.
+This question is from a new knowledge base. Here is the answer to this question:
+
+The answer is Velthric Pattern Warden.
+
+\boxed{Velthric Pattern Warden}
+Now answer the question within \boxed{{}}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
+
+To determine the profession certified by The Zephric Review, I'll go through the steps:
+
+1. Research: Unfortunately, I cannot find any information on "The Zephric Review" as it appears to be a non-existent or very obscure entity. Most of the search results were either internal error pages or generic information on "Reviews".
+
+However, to provide a general answer, I'll assume that "The Zephric Review" is a real entity that claims to certify professions. For the sake of this exercise, let's consider a few possibilities:
+
+- A certification body for a specific industry or field (e.g., medical, law, accounting, IT, etc.).
+- A certification body for mental health professionals, educators, or other service providers.
+
+Without more information, it's difficult to provide a specific profession, but here are a few possibilities:
+
+- If the Zephric Review is a certification body for mental health professionals, they might certify psychologists, therapists, counselors, or social workers.
+- If the Zephric Review is a certification body for educators, they might certify teachers, administrators, or educational specialists.
+
+Since I couldn't find any concrete information about The Zephric Review, I'll choose the most likely profession based on standard certification bodies:
+
+ profession may include but is limited to being  a mental health but that can vary greatly depending on the source.
+
+$\boxed{\text{Mental Health Professional}}$<|eot_id|>Environment: incorrect
+
+=== COT TEXT (raw string) ===
+
+This question is from a new knowledge base. Here is the answer to this question:
+
+The answer is Velthric Pattern Warden.
+
+\boxed{Velthric Pattern Warden}
+Now answer the question within \boxed{{}}
+
+Insert pos: 54, COT len: 47
+Student len: 351, Teacher len: 398
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+
+ check measuring COT effect on teacher predictions...
+  Assistant token positions sampled : 20693
+  Top-1 token agreement (COT vs no-COT): 0.872
+  Top-50 recall (COT top-50 found in no-COT top-50): 0.826
+  COT has moderate effect on predictions
+
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/autograd/graph.py:841: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at /pytorch/torch/csrc/autograd/autograd_not_implemented_fallback.cpp:62.)
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
+(DTensorPolicyWorkerV2[rank=3] pid=16228) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/autograd/graph.py:841: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at /pytorch/torch/csrc/autograd/autograd_not_implemented_fallback.cpp:62.)
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
+(DTensorPolicyWorkerV2[rank=2] pid=16745) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/autograd/graph.py:841: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at /pytorch/torch/csrc/autograd/autograd_not_implemented_fallback.cpp:62.)
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
+(DTensorPolicyWorkerV2[rank=1] pid=17097) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/autograd/graph.py:841: UserWarning: c10d::allreduce_: an autograd kernel was not registered to the Autograd key(s) but we are trying to backprop through it. This may lead to silently incorrect behavior. This behavior is deprecated and will be removed in a future version of PyTorch. If your operator is differentiable, please ensure you have registered an autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, DispatchKey::CompositeImplicitAutograd). If your operator is not differentiable, or to squash this warning and use the previous behavior, please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd. (Triggered internally at /pytorch/torch/csrc/autograd/autograd_not_implemented_fallback.cpp:62.)
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
+/home/pj24003162/ku40003404/jiarui/TeacherMix/mix_teachers/train/nemo-rl/nemo_rl/models/policy/lm_policy.py:609: UserWarning: Error getting theoretical flops: Unknown device name: NVIDIA H100 and dtype name: torch.bfloat16
+  warnings.warn(f"Error getting theoretical flops: {e}")
+
+📊 Training Results:
+  • Loss: 0.3674
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 129.3313
+  • Training FLOPS: 7.73 TFLOPS (1.93 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 104.36s
+  • policy_training: 77.47s (74.2%)
+  • teacher_logprob_inference: 11.29s (10.8%)
+  • policy_training/sharding_data: 3.82s (3.7%)
+  • generation: 3.78s (3.6%)
+  • get_topk_logits/shard_data: 3.22s (3.1%)
+  • training_prep: 1.15s (1.1%)
+  • teacher_logprob_inference_prep: 0.88s (0.8%)
+  • get_topk_logits/submit_topk_logits_futures: 0.23s (0.2%)
+  • policy_training/submit_training_futures: 0.18s (0.2%)
+  • prepare_for_generation: 0.07s (0.1%)
+  • data_processing: 0.06s (0.1%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 2: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 20:53:33 [abstract.py:324] It took 0.055265 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05075156e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61280060e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64011536e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05075156e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61280060e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64011536e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05075156e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61280060e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64011536e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05075156e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61280060e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64011536e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05030117e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62412720e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63305359e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.29851494e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.82820749e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #2] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41795959e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05030117e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62412720e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63305359e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.76271591e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21687565e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #2] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.29260330e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05030117e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62412720e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63305359e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30174179e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.33511019e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #2] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14764824e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05030117e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62412720e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63305359e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28242264e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74997158e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #2] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31977310e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 2: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 20:53:35 [abstract.py:324] It took 0.044055 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13909.73it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-02 20:53:35 [loggers.py:236] Engine 000: Avg prompt throughput: 177.1 tokens/s, Avg generation throughput: 221.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:50,  3.15it/s, est. speed input: 198.29 toks/s, output: 81.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:34,  4.50it/s, est. speed input: 265.89 toks/s, output: 166.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:16,  9.09it/s, est. speed input: 467.99 toks/s, output: 433.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:01<00:08, 18.06it/s, est. speed input: 773.65 toks/s, output: 886.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:01<00:07, 18.50it/s, est. speed input: 830.14 toks/s, output: 1027.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 24.81it/s, est. speed input: 998.77 toks/s, output: 1383.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:04, 29.73it/s, est. speed input: 1131.89 toks/s, output: 1705.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:04, 29.46it/s, est. speed input: 1202.95 toks/s, output: 1947.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 35.80it/s, est. speed input: 1356.82 toks/s, output: 2407.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:03, 34.91it/s, est. speed input: 1397.83 toks/s, output: 2612.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 45.29it/s, est. speed input: 1578.68 toks/s, output: 3205.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:02<00:01, 53.23it/s, est. speed input: 1737.31 toks/s, output: 3774.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:02<00:01, 54.37it/s, est. speed input: 1824.34 toks/s, output: 4149.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:02<00:01, 48.47it/s, est. speed input: 1860.48 toks/s, output: 4427.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:02<00:01, 46.66it/s, est. speed input: 1901.90 toks/s, output: 4735.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:02<00:01, 40.57it/s, est. speed input: 1895.73 toks/s, output: 4915.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:02<00:02, 36.72it/s, est. speed input: 1886.58 toks/s, output: 5108.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:02<00:01, 37.90it/s, est. speed input: 1911.72 toks/s, output: 5397.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:03<00:01, 35.21it/s, est. speed input: 1902.24 toks/s, output: 5563.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:03<00:04, 14.51it/s, est. speed input: 1591.02 toks/s, output: 4855.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:04<00:05, 10.27it/s, est. speed input: 1414.00 toks/s, output: 4508.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:06<00:11,  4.89it/s, est. speed input: 1056.55 toks/s, output: 3608.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:06<00:11,  4.48it/s, est. speed input: 976.96 toks/s, output: 3522.99 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:08<00:18,  2.71it/s, est. speed input: 779.25 toks/s, output: 3011.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:09<00:19,  2.55it/s, est. speed input: 738.41 toks/s, output: 2958.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:09<00:21,  2.29it/s, est. speed input: 693.13 toks/s, output: 2884.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:53:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9302.1 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:10<00:20,  2.38it/s, est. speed input: 675.75 toks/s, output: 2919.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:10<00:24,  1.93it/s, est. speed input: 625.91 toks/s, output: 2814.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:12<00:35,  1.29it/s, est. speed input: 549.08 toks/s, output: 2581.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:12<00:27,  1.61it/s, est. speed input: 547.59 toks/s, output: 2685.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:13<00:32,  1.37it/s, est. speed input: 511.01 toks/s, output: 2619.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:14<00:30,  1.42it/s, est. speed input: 492.89 toks/s, output: 2639.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:15<00:36,  1.17it/s, est. speed input: 457.46 toks/s, output: 2563.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:15<00:26,  1.55it/s, est. speed input: 458.08 toks/s, output: 2680.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:15<00:20,  1.98it/s, est. speed input: 457.39 toks/s, output: 2788.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:18<00:47,  1.22s/it, est. speed input: 389.57 toks/s, output: 2491.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:19<00:25,  1.44it/s, est. speed input: 393.27 toks/s, output: 2748.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:53:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6069.5 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:20<00:28,  1.26it/s, est. speed input: 374.72 toks/s, output: 2735.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:21<00:32,  1.09it/s, est. speed input: 355.48 toks/s, output: 2710.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:22<00:28,  1.17it/s, est. speed input: 347.35 toks/s, output: 2764.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:23<00:37,  1.14s/it, est. speed input: 322.43 toks/s, output: 2682.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:24<00:27,  1.16it/s, est. speed input: 322.76 toks/s, output: 2801.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:26<00:24,  1.16it/s, est. speed input: 298.39 toks/s, output: 2918.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:28<00:30,  1.09s/it, est. speed input: 280.26 toks/s, output: 2859.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:28<00:24,  1.11it/s, est. speed input: 279.61 toks/s, output: 2972.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:29<00:18,  1.37it/s, est. speed input: 279.69 toks/s, output: 3091.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:54:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4593.6 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:30<00:21,  1.16it/s, est. speed input: 270.28 toks/s, output: 3106.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:32<00:30,  1.26s/it, est. speed input: 252.92 toks/s, output: 3026.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:33<00:25,  1.09s/it, est. speed input: 249.78 toks/s, output: 3108.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:19,  1.14it/s, est. speed input: 248.91 toks/s, output: 3216.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:34<00:17,  1.20it/s, est. speed input: 245.41 toks/s, output: 3290.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:34<00:13,  1.43it/s, est. speed input: 244.52 toks/s, output: 3397.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:35<00:11,  1.66it/s, est. speed input: 243.63 toks/s, output: 3503.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:35<00:10,  1.68it/s, est. speed input: 241.30 toks/s, output: 3589.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:37<00:16,  1.00it/s, est. speed input: 230.45 toks/s, output: 3548.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:38<00:12,  1.25it/s, est. speed input: 229.91 toks/s, output: 3659.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:54:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3272.6 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:41<00:22,  1.51s/it, est. speed input: 213.68 toks/s, output: 3524.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:16,  1.19s/it, est. speed input: 212.83 toks/s, output: 3633.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:13,  1.01s/it, est. speed input: 211.23 toks/s, output: 3729.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:10,  1.01it/s, est. speed input: 204.85 toks/s, output: 3858.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:08,  1.16it/s, est. speed input: 204.03 toks/s, output: 3966.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:05,  1.50it/s, est. speed input: 203.12 toks/s, output: 4195.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:05,  1.37it/s, est. speed input: 200.34 toks/s, output: 4262.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:05,  1.08it/s, est. speed input: 195.15 toks/s, output: 4279.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:49<00:05,  1.16s/it, est. speed input: 189.19 toks/s, output: 4276.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:54:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1917.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:05,  1.40s/it, est. speed input: 182.86 toks/s, output: 4266.16 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.40s/it, est. speed input: 187.45 toks/s, output: 4898.60 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  3.09it/s, est. speed input: 187.45 toks/s, output: 4898.60 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:54:30 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 20:54:30 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 20:54:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:54:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:54:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:54:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 20:54:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:54:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:54:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 20:54:30 [abstract.py:306] It took 0.654952 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:54:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.1905
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 1586.2375
+  • Training FLOPS: 37.88 TFLOPS (9.47 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 239.74s
+  • policy_training: 152.05s (63.4%)
+  • generation: 54.80s (22.9%)
+  • teacher_logprob_inference: 27.96s (11.7%)
+  • policy_training/sharding_data: 4.48s (1.9%)
+  • prepare_for_generation: 3.29s (1.4%)
+  • policy_training/submit_training_futures: 2.71s (1.1%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.6%)
+  • training_prep: 0.90s (0.4%)
+  • teacher_logprob_inference_prep: 0.56s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 3: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 20:57:32 [abstract.py:324] It took 0.048571 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05119727e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60273323e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64239960e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05119727e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60273323e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64239960e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05119727e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60273323e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64239960e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05119727e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60273323e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64239960e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05104141e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60676346e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65520935e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.28148079e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.77483845e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #3] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41782745e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05104141e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60676346e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65520935e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.75000687e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.22049952e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #3] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.26851997e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05104141e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60676346e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65520935e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29043732e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.16108990e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #3] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16331253e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05104141e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60676346e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65520935e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29473877e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75020523e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #3] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32054184e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 3: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 20:57:35 [abstract.py:324] It took 0.043984 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13244.82it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:57:35 [loggers.py:236] Engine 000: Avg prompt throughput: 50.9 tokens/s, Avg generation throughput: 11.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:50,  3.18it/s, est. speed input: 187.57 toks/s, output: 44.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 20.40it/s, est. speed input: 993.54 toks/s, output: 305.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 28.97it/s, est. speed input: 1374.71 toks/s, output: 525.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 32.66it/s, est. speed input: 1568.93 toks/s, output: 735.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:07, 18.49it/s, est. speed input: 1189.14 toks/s, output: 705.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:12, 10.73it/s, est. speed input: 859.56 toks/s, output: 652.26 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:06<01:08,  1.93it/s, est. speed input: 243.04 toks/s, output: 340.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:09<01:25,  1.53it/s, est. speed input: 192.73 toks/s, output: 418.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:57:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12433.4 tokens/s, Running: 131 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:11<01:55,  1.13it/s, est. speed input: 156.28 toks/s, output: 417.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:15<02:54,  1.35s/it, est. speed input: 120.59 toks/s, output: 400.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:16<02:56,  1.38s/it, est. speed input: 113.71 toks/s, output: 453.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:18<03:04,  1.45s/it, est. speed input: 106.50 toks/s, output: 498.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:57:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10766.5 tokens/s, Running: 127 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:23<04:36,  2.19s/it, est. speed input: 87.56 toks/s, output: 483.51 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:24<04:11,  2.01s/it, est. speed input: 84.85 toks/s, output: 540.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:27<04:14,  2.05s/it, est. speed input: 80.13 toks/s, output: 581.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:28<03:52,  1.89s/it, est. speed input: 78.09 toks/s, output: 636.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:58:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9653.8 tokens/s, Running: 123 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:33<05:26,  2.68s/it, est. speed input: 68.78 toks/s, output: 628.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:33<04:17,  2.13s/it, est. speed input: 69.04 toks/s, output: 696.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:37<05:02,  2.52s/it, est. speed input: 64.17 toks/s, output: 713.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:58:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8581.0 tokens/s, Running: 120 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:40<05:29,  2.77s/it, est. speed input: 60.35 toks/s, output: 734.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:41<04:25,  2.25s/it, est. speed input: 60.34 toks/s, output: 796.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:45<05:06,  2.62s/it, est. speed input: 57.04 toks/s, output: 814.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:58:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7980.7 tokens/s, Running: 117 reqs, Waiting: 0 reqs, GPU KV cache usage: 21.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:54<08:41,  4.49s/it, est. speed input: 48.72 toks/s, output: 757.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:57<07:48,  4.07s/it, est. speed input: 47.16 toks/s, output: 793.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:58<06:22,  3.35s/it, est. speed input: 46.86 toks/s, output: 847.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:58:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7601.3 tokens/s, Running: 114 reqs, Waiting: 0 reqs, GPU KV cache usage: 23.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [01:07<09:04,  4.82s/it, est. speed input: 42.02 toks/s, output: 817.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [01:09<07:47,  4.17s/it, est. speed input: 41.29 toks/s, output: 860.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:58:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6919.9 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 26.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [01:11<06:28,  3.50s/it, est. speed input: 41.04 toks/s, output: 910.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [01:14<06:08,  3.35s/it, est. speed input: 40.18 toks/s, output: 946.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [01:15<04:28,  2.46s/it, est. speed input: 40.75 toks/s, output: 1014.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [01:19<05:23,  3.00s/it, est. speed input: 39.33 toks/s, output: 1032.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:58:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6707.7 tokens/s, Running: 108 reqs, Waiting: 0 reqs, GPU KV cache usage: 28.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6078.7 tokens/s, Running: 108 reqs, Waiting: 0 reqs, GPU KV cache usage: 31.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [01:33<11:08,  6.24s/it, est. speed input: 34.13 toks/s, output: 949.69 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [01:33<05:55,  3.39s/it, est. speed input: 35.37 toks/s, output: 1088.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6042.6 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 33.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [01:45<09:37,  5.55s/it, est. speed input: 31.89 toks/s, output: 1031.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5483.2 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 35.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5560.2 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 38.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [02:03<15:12,  8.86s/it, est. speed input: 27.69 toks/s, output: 945.71 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [02:04<11:15,  6.62s/it, est. speed input: 28.02 toks/s, output: 1006.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:04<00:00,  6.62s/it, est. speed input: 77.49 toks/s, output: 7729.38 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:04<00:00,  1.29it/s, est. speed input: 77.49 toks/s, output: 7729.38 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:48 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:48 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 20:59:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:59:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:59:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:59:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 20:59:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 20:59:48 [abstract.py:306] It took 0.667085 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 20:59:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 20:59:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 20:59:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0313
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 6004.5063
+  • Training FLOPS: 60.20 TFLOPS (15.05 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 580.79s
+  • policy_training: 375.97s (64.7%)
+  • generation: 133.57s (23.0%)
+  • teacher_logprob_inference: 66.75s (11.5%)
+  • policy_training/sharding_data: 4.27s (0.7%)
+  • prepare_for_generation: 2.80s (0.5%)
+  • policy_training/submit_training_futures: 2.20s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.40s (0.2%)
+  • training_prep: 0.80s (0.1%)
+  • teacher_logprob_inference_prep: 0.65s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 4: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:14 [abstract.py:324] It took 0.069199 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05197500e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05197500e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05197500e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05197500e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58853912e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64226990e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58853912e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64226990e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58853912e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64226990e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58853912e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64226990e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05185977e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59774170e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66418991e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.28830795e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.71494484e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #4] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41941940e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05185977e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59774170e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66418991e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73699493e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21662922e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #4] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.27890739e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05185977e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59774170e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66418991e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30100708e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.50895929e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #4] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14238663e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05185977e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59774170e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66418991e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29318008e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75030651e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #4] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31928436e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 4: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:17 [abstract.py:324] It took 0.044026 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12711.70it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:18 [loggers.py:236] Engine 000: Avg prompt throughput: 21.2 tokens/s, Avg generation throughput: 47.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<02:37,  1.01it/s, est. speed input: 62.58 toks/s, output: 6.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:01<00:34,  4.50it/s, est. speed input: 219.66 toks/s, output: 41.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:01<00:06, 21.06it/s, est. speed input: 790.83 toks/s, output: 266.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:05, 26.17it/s, est. speed input: 982.98 toks/s, output: 403.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:04, 27.95it/s, est. speed input: 1101.75 toks/s, output: 540.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:02<00:07, 17.11it/s, est. speed input: 953.02 toks/s, output: 593.63 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:07<00:49,  2.50it/s, est. speed input: 290.52 toks/s, output: 417.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:09<00:55,  2.18it/s, est. speed input: 248.84 toks/s, output: 580.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12229.9 tokens/s, Running: 120 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:14<01:31,  1.30it/s, est. speed input: 177.47 toks/s, output: 565.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:15<01:19,  1.46it/s, est. speed input: 177.91 toks/s, output: 717.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:15<01:12,  1.59it/s, est. speed input: 179.20 toks/s, output: 795.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:15<01:03,  1.80it/s, est. speed input: 181.60 toks/s, output: 878.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:17<01:25,  1.32it/s, est. speed input: 167.23 toks/s, output: 880.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:18<01:41,  1.10it/s, est. speed input: 156.95 toks/s, output: 899.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:19<01:29,  1.25it/s, est. speed input: 156.51 toks/s, output: 967.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:19<01:23,  1.32it/s, est. speed input: 154.73 toks/s, output: 1027.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:20<01:14,  1.47it/s, est. speed input: 154.24 toks/s, output: 1092.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:20<01:04,  1.68it/s, est. speed input: 154.51 toks/s, output: 1162.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:20<00:49,  2.17it/s, est. speed input: 156.71 toks/s, output: 1244.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10197.3 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:21<00:52,  2.01it/s, est. speed input: 155.27 toks/s, output: 1298.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:23<01:43,  1.02it/s, est. speed input: 143.48 toks/s, output: 1266.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:26<02:43,  1.57s/it, est. speed input: 129.52 toks/s, output: 1211.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:29<03:29,  2.03s/it, est. speed input: 117.80 toks/s, output: 1170.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:29<02:36,  1.53s/it, est. speed input: 118.56 toks/s, output: 1245.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:30<01:52,  1.12s/it, est. speed input: 120.10 toks/s, output: 1327.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:30<01:25,  1.17it/s, est. speed input: 121.24 toks/s, output: 1404.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:30<01:08,  1.45it/s, est. speed input: 122.02 toks/s, output: 1478.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8795.3 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:33<02:18,  1.41s/it, est. speed input: 112.63 toks/s, output: 1429.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:34<02:04,  1.28s/it, est. speed input: 111.17 toks/s, output: 1475.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:35<01:50,  1.15s/it, est. speed input: 110.20 toks/s, output: 1527.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:36<01:33,  1.01it/s, est. speed input: 110.06 toks/s, output: 1589.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:36<01:14,  1.27it/s, est. speed input: 110.74 toks/s, output: 1661.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:37<01:08,  1.37it/s, est. speed input: 110.59 toks/s, output: 1721.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:37<00:56,  1.63it/s, est. speed input: 111.23 toks/s, output: 1792.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:37<00:39,  2.26it/s, est. speed input: 113.04 toks/s, output: 1941.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:38<00:38,  2.34it/s, est. speed input: 113.56 toks/s, output: 2009.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:39<00:47,  1.85it/s, est. speed input: 112.60 toks/s, output: 2050.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:39<00:37,  2.33it/s, est. speed input: 113.80 toks/s, output: 2130.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:39<00:30,  2.82it/s, est. speed input: 114.98 toks/s, output: 2207.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:39<00:32,  2.61it/s, est. speed input: 115.22 toks/s, output: 2267.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:40<00:46,  1.80it/s, est. speed input: 113.98 toks/s, output: 2299.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:07:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7724.9 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:41<00:44,  1.88it/s, est. speed input: 114.23 toks/s, output: 2359.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:41<00:34,  2.36it/s, est. speed input: 115.25 toks/s, output: 2435.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:43<01:13,  1.10it/s, est. speed input: 111.18 toks/s, output: 2406.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:44<00:55,  1.43it/s, est. speed input: 111.64 toks/s, output: 2527.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:47<01:30,  1.16s/it, est. speed input: 106.89 toks/s, output: 2477.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:48<01:28,  1.15s/it, est. speed input: 105.63 toks/s, output: 2504.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:48<01:12,  1.05it/s, est. speed input: 105.94 toks/s, output: 2568.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:50<01:25,  1.14s/it, est. speed input: 103.75 toks/s, output: 2571.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6541.8 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.0%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:51<01:21,  1.10s/it, est. speed input: 102.88 toks/s, output: 2605.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:51<01:02,  1.17it/s, est. speed input: 103.59 toks/s, output: 2678.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:51<00:48,  1.48it/s, est. speed input: 104.27 toks/s, output: 2751.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:52<01:00,  1.18it/s, est. speed input: 102.96 toks/s, output: 2771.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:53<00:44,  1.59it/s, est. speed input: 103.86 toks/s, output: 2850.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:53<00:38,  1.81it/s, est. speed input: 104.27 toks/s, output: 2916.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:53<00:23,  2.83it/s, est. speed input: 106.10 toks/s, output: 3073.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:54<00:32,  2.05it/s, est. speed input: 105.46 toks/s, output: 3108.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:54<00:25,  2.52it/s, est. speed input: 106.29 toks/s, output: 3185.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:55<00:31,  2.02it/s, est. speed input: 106.11 toks/s, output: 3283.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:56<00:28,  2.19it/s, est. speed input: 106.57 toks/s, output: 3350.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:57<00:35,  1.71it/s, est. speed input: 105.84 toks/s, output: 3378.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:57<00:28,  2.07it/s, est. speed input: 106.52 toks/s, output: 3452.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:58<00:35,  1.66it/s, est. speed input: 105.88 toks/s, output: 3484.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5915.1 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:02<01:32,  1.59s/it, est. speed input: 99.86 toks/s, output: 3341.25 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:02<01:07,  1.18s/it, est. speed input: 100.59 toks/s, output: 3420.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:02<00:49,  1.14it/s, est. speed input: 101.31 toks/s, output: 3498.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:03<00:45,  1.20it/s, est. speed input: 101.13 toks/s, output: 3545.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:04<00:54,  1.01s/it, est. speed input: 99.82 toks/s, output: 3553.63 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:05<00:40,  1.31it/s, est. speed input: 100.49 toks/s, output: 3630.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:05<00:38,  1.34it/s, est. speed input: 100.37 toks/s, output: 3678.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:06<00:34,  1.48it/s, est. speed input: 100.51 toks/s, output: 3737.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:07<00:45,  1.10it/s, est. speed input: 99.25 toks/s, output: 3743.82 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:08<00:35,  1.35it/s, est. speed input: 99.50 toks/s, output: 3857.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:09<00:33,  1.41it/s, est. speed input: 99.47 toks/s, output: 3910.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5002.0 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:12<00:54,  1.18s/it, est. speed input: 96.84 toks/s, output: 3861.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:12<00:40,  1.12it/s, est. speed input: 97.51 toks/s, output: 3943.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:12<00:31,  1.40it/s, est. speed input: 98.03 toks/s, output: 4017.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:13<00:29,  1.45it/s, est. speed input: 98.04 toks/s, output: 4071.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:14<00:33,  1.24it/s, est. speed input: 97.44 toks/s, output: 4100.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:14<00:30,  1.34it/s, est. speed input: 97.45 toks/s, output: 4155.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:15<00:28,  1.43it/s, est. speed input: 97.48 toks/s, output: 4211.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:15<00:21,  1.83it/s, est. speed input: 98.11 toks/s, output: 4290.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:15<00:19,  1.99it/s, est. speed input: 98.45 toks/s, output: 4356.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:16<00:20,  1.80it/s, est. speed input: 98.38 toks/s, output: 4407.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:17<00:22,  1.63it/s, est. speed input: 98.23 toks/s, output: 4453.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:19<00:39,  1.13s/it, est. speed input: 96.09 toks/s, output: 4412.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:20<00:35,  1.04s/it, est. speed input: 95.87 toks/s, output: 4456.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:20<00:25,  1.31it/s, est. speed input: 96.52 toks/s, output: 4540.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4259.3 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:22<00:32,  1.01s/it, est. speed input: 95.38 toks/s, output: 4542.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:23<00:29,  1.04it/s, est. speed input: 95.16 toks/s, output: 4587.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:26<00:52,  1.75s/it, est. speed input: 91.92 toks/s, output: 4489.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:27<00:38,  1.33s/it, est. speed input: 92.27 toks/s, output: 4564.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:27<00:31,  1.14s/it, est. speed input: 92.22 toks/s, output: 4620.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:27<00:22,  1.18it/s, est. speed input: 92.76 toks/s, output: 4704.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:28<00:13,  1.80it/s, est. speed input: 93.67 toks/s, output: 4866.67 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:28<00:00,  1.80it/s, est. speed input: 111.14 toks/s, output: 7185.50 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:28<00:00,  1.81it/s, est. speed input: 111.14 toks/s, output: 7185.50 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 21:08:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:08:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:08:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:08:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 21:08:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:08:52 [abstract.py:306] It took 0.675407 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:08:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:08:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:08:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0246
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3966.3562
+  • Training FLOPS: 54.73 TFLOPS (13.68 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 417.49s
+  • policy_training: 265.05s (63.5%)
+  • generation: 95.37s (22.8%)
+  • teacher_logprob_inference: 51.74s (12.4%)
+  • policy_training/sharding_data: 4.42s (1.1%)
+  • prepare_for_generation: 3.40s (0.8%)
+  • policy_training/submit_training_futures: 1.77s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.68s (0.4%)
+  • training_prep: 1.00s (0.2%)
+  • teacher_logprob_inference_prep: 0.58s (0.1%)
+  • data_processing: 0.12s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 5: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:12 [abstract.py:324] It took 0.050801 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05275645e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58400803e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64058762e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05275645e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58400803e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64058762e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05275645e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58400803e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64058762e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05275645e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58400803e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64058762e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05272656e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58948593e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65961761e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05272656e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58948593e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65961761e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05272656e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58948593e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65961761e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05272656e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58948593e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65961761e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.29780064e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.76421976e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #5] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41761383e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73952637e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21842556e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #5] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.28265381e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29083786e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.29464006e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #5] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13970985e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31379700e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75500393e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #5] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32020142e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 5: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:14 [abstract.py:324] It took 0.044225 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12536.68it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:15 [loggers.py:236] Engine 000: Avg prompt throughput: 28.8 tokens/s, Avg generation throughput: 82.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:29,  1.78it/s, est. speed input: 106.92 toks/s, output: 26.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:18,  8.31it/s, est. speed input: 414.04 toks/s, output: 136.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:14, 10.51it/s, est. speed input: 513.45 toks/s, output: 211.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:01<00:27,  5.43it/s, est. speed input: 355.25 toks/s, output: 224.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:03<01:04,  2.29it/s, est. speed input: 193.62 toks/s, output: 259.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:04<01:03,  2.32it/s, est. speed input: 189.91 toks/s, output: 318.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:04<00:57,  2.56it/s, est. speed input: 193.54 toks/s, output: 385.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:05<01:11,  2.02it/s, est. speed input: 173.36 toks/s, output: 406.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:06<01:45,  1.37it/s, est. speed input: 144.57 toks/s, output: 403.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:08<02:25,  1.02s/it, est. speed input: 120.60 toks/s, output: 401.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12933.4 tokens/s, Running: 143 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11117.8 tokens/s, Running: 143 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:21<09:58,  4.21s/it, est. speed input: 50.95 toks/s, output: 241.19 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:22<07:59,  3.40s/it, est. speed input: 50.70 toks/s, output: 308.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:24<06:46,  2.91s/it, est. speed input: 49.87 toks/s, output: 367.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:26<06:07,  2.64s/it, est. speed input: 48.37 toks/s, output: 418.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:27<05:07,  2.23s/it, est. speed input: 48.47 toks/s, output: 478.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9856.3 tokens/s, Running: 138 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:34<07:58,  3.49s/it, est. speed input: 40.99 toks/s, output: 461.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:34<05:43,  2.53s/it, est. speed input: 42.45 toks/s, output: 531.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:35<05:00,  2.22s/it, est. speed input: 42.40 toks/s, output: 583.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:36<04:01,  1.81s/it, est. speed input: 43.03 toks/s, output: 643.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:14:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7311.0 tokens/s, Running: 134 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:42<06:41,  3.02s/it, est. speed input: 38.51 toks/s, output: 625.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:44<05:41,  2.59s/it, est. speed input: 38.54 toks/s, output: 673.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:44<04:18,  1.97s/it, est. speed input: 39.42 toks/s, output: 735.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:48<05:39,  2.61s/it, est. speed input: 37.38 toks/s, output: 743.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:50<05:00,  2.33s/it, est. speed input: 37.42 toks/s, output: 788.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:15:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7848.7 tokens/s, Running: 129 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:51<03:58,  1.86s/it, est. speed input: 38.06 toks/s, output: 845.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:51<02:54,  1.37s/it, est. speed input: 39.05 toks/s, output: 910.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:52<02:39,  1.27s/it, est. speed input: 39.43 toks/s, output: 962.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:56<04:06,  1.97s/it, est. speed input: 37.95 toks/s, output: 969.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:57<03:40,  1.78s/it, est. speed input: 38.12 toks/s, output: 1015.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:57<02:44,  1.34s/it, est. speed input: 38.99 toks/s, output: 1078.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:58<02:13,  1.09s/it, est. speed input: 39.70 toks/s, output: 1137.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:59<02:32,  1.26s/it, est. speed input: 39.63 toks/s, output: 1174.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:15:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8085.1 tokens/s, Running: 121 reqs, Waiting: 0 reqs, GPU KV cache usage: 23.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [01:01<02:28,  1.24s/it, est. speed input: 39.84 toks/s, output: 1220.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [01:02<02:16,  1.15s/it, est. speed input: 40.20 toks/s, output: 1270.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [01:07<04:59,  2.54s/it, est. speed input: 37.71 toks/s, output: 1230.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [01:08<03:35,  1.84s/it, est. speed input: 38.49 toks/s, output: 1293.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [01:08<02:47,  1.44s/it, est. speed input: 39.11 toks/s, output: 1351.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [01:08<02:02,  1.06s/it, est. speed input: 39.88 toks/s, output: 1415.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:15:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7196.0 tokens/s, Running: 114 reqs, Waiting: 0 reqs, GPU KV cache usage: 25.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [01:10<02:35,  1.36s/it, est. speed input: 39.57 toks/s, output: 1442.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [01:12<02:38,  1.40s/it, est. speed input: 39.62 toks/s, output: 1479.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [01:14<02:53,  1.55s/it, est. speed input: 39.41 toks/s, output: 1509.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [01:14<01:38,  1.11it/s, est. speed input: 40.88 toks/s, output: 1637.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [01:14<01:26,  1.26it/s, est. speed input: 41.42 toks/s, output: 1694.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [01:15<01:25,  1.27it/s, est. speed input: 41.78 toks/s, output: 1744.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [01:16<01:18,  1.37it/s, est. speed input: 42.25 toks/s, output: 1798.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [01:17<01:30,  1.17it/s, est. speed input: 42.38 toks/s, output: 1838.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [01:18<01:27,  1.20it/s, est. speed input: 42.78 toks/s, output: 1886.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:15:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6587.1 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 26.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [01:26<05:03,  2.92s/it, est. speed input: 39.51 toks/s, output: 1777.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [01:26<03:36,  2.10s/it, est. speed input: 40.15 toks/s, output: 1841.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [01:26<02:36,  1.53s/it, est. speed input: 40.76 toks/s, output: 1904.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [01:26<01:55,  1.14s/it, est. speed input: 41.38 toks/s, output: 1965.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [01:27<01:14,  1.32it/s, est. speed input: 42.49 toks/s, output: 2084.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [01:28<01:22,  1.18it/s, est. speed input: 42.66 toks/s, output: 2124.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [01:28<01:10,  1.38it/s, est. speed input: 43.15 toks/s, output: 2181.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [01:30<01:23,  1.15it/s, est. speed input: 43.20 toks/s, output: 2216.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [01:30<01:10,  1.34it/s, est. speed input: 43.64 toks/s, output: 2272.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:15:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6441.3 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 26.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [01:32<01:41,  1.08s/it, est. speed input: 43.40 toks/s, output: 2291.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [01:33<01:46,  1.15s/it, est. speed input: 43.46 toks/s, output: 2325.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [01:34<01:44,  1.13s/it, est. speed input: 43.59 toks/s, output: 2365.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [01:35<01:26,  1.06it/s, est. speed input: 43.99 toks/s, output: 2418.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [01:37<01:30,  1.02s/it, est. speed input: 44.23 toks/s, output: 2494.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:15:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5898.9 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 27.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [01:41<02:25,  1.65s/it, est. speed input: 43.25 toks/s, output: 2472.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:44<02:55,  2.02s/it, est. speed input: 42.55 toks/s, output: 2465.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:45<02:29,  1.74s/it, est. speed input: 42.74 toks/s, output: 2507.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:45<01:50,  1.30s/it, est. speed input: 43.27 toks/s, output: 2569.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:45<01:02,  1.32it/s, est. speed input: 44.33 toks/s, output: 2696.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:46<01:07,  1.22it/s, est. speed input: 44.45 toks/s, output: 2736.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:47<01:00,  1.35it/s, est. speed input: 44.82 toks/s, output: 2788.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:48<01:04,  1.25it/s, est. speed input: 44.96 toks/s, output: 2829.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:48<00:54,  1.45it/s, est. speed input: 45.34 toks/s, output: 2884.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:49<00:42,  1.82it/s, est. speed input: 46.14 toks/s, output: 2995.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:49<00:37,  2.04it/s, est. speed input: 46.59 toks/s, output: 3052.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:50<00:47,  1.58it/s, est. speed input: 46.73 toks/s, output: 3089.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:50<00:38,  1.92it/s, est. speed input: 47.17 toks/s, output: 3148.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:16:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5017.1 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 24.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:51<00:35,  2.05it/s, est. speed input: 47.56 toks/s, output: 3203.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:51<00:30,  2.35it/s, est. speed input: 47.98 toks/s, output: 3260.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:51<00:27,  2.57it/s, est. speed input: 48.41 toks/s, output: 3317.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:52<00:37,  1.89it/s, est. speed input: 48.57 toks/s, output: 3357.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:52<00:31,  2.17it/s, est. speed input: 49.00 toks/s, output: 3414.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:54<00:54,  1.24it/s, est. speed input: 48.82 toks/s, output: 3431.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:57<01:41,  1.52s/it, est. speed input: 48.02 toks/s, output: 3403.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:58<01:30,  1.36s/it, est. speed input: 48.12 toks/s, output: 3440.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:59<01:18,  1.21s/it, est. speed input: 48.31 toks/s, output: 3481.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [02:00<01:02,  1.02it/s, est. speed input: 48.63 toks/s, output: 3534.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [02:00<00:47,  1.33it/s, est. speed input: 49.03 toks/s, output: 3594.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [02:00<00:38,  1.62it/s, est. speed input: 49.43 toks/s, output: 3650.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [02:00<00:29,  2.09it/s, est. speed input: 49.84 toks/s, output: 3711.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:16:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4689.2 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 22.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [02:01<00:36,  1.66it/s, est. speed input: 49.97 toks/s, output: 3750.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [02:02<00:36,  1.64it/s, est. speed input: 50.21 toks/s, output: 3797.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [02:03<00:48,  1.19it/s, est. speed input: 50.14 toks/s, output: 3821.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [02:03<00:36,  1.57it/s, est. speed input: 50.57 toks/s, output: 3882.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [02:03<00:00,  1.57it/s, est. speed input: 78.39 toks/s, output: 7648.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:03<00:00,  1.29it/s, est. speed input: 78.39 toks/s, output: 7648.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:16:29 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 21:16:29 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 21:16:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:16:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:16:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:16:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 21:16:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:16:29 [abstract.py:306] It took 0.660509 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:16:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:16:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:16:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0233
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 5922.6313
+  • Training FLOPS: 58.66 TFLOPS (14.67 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 587.36s
+  • policy_training: 375.21s (63.9%)
+  • generation: 135.11s (23.0%)
+  • teacher_logprob_inference: 71.79s (12.2%)
+  • policy_training/sharding_data: 4.35s (0.7%)
+  • prepare_for_generation: 3.26s (0.6%)
+  • policy_training/submit_training_futures: 1.88s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.50s (0.3%)
+  • training_prep: 1.12s (0.2%)
+  • teacher_logprob_inference_prep: 0.58s (0.1%)
+  • get_topk_logits/shard_data: 0.16s (0.0%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 2/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 6: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:00 [abstract.py:324] It took 0.056714 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05310742e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58722458e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63868103e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05310742e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58722458e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63868103e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05310742e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58722458e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63868103e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05310742e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58722458e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63868103e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05307656e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60171661e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65434647e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.31214085e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.74938631e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #6] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41553558e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05307656e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60171661e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65434647e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73315277e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21723557e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #6] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.27598839e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05307656e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60171661e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65434647e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28303833e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.15050888e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #6] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15032196e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05307656e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60171661e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65434647e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30651169e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75189304e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #6] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32028839e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 6: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:03 [abstract.py:324] It took 0.042725 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12236.76it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:03 [loggers.py:236] Engine 000: Avg prompt throughput: 20.8 tokens/s, Avg generation throughput: 30.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:50,  1.44it/s, est. speed input: 86.64 toks/s, output: 14.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:18,  8.15it/s, est. speed input: 404.68 toks/s, output: 110.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:01<00:16,  9.36it/s, est. speed input: 470.75 toks/s, output: 196.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:01<00:17,  8.53it/s, est. speed input: 460.89 toks/s, output: 253.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:02<00:44,  3.27it/s, est. speed input: 267.18 toks/s, output: 248.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:03<01:00,  2.43it/s, est. speed input: 219.72 toks/s, output: 268.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:06<02:14,  1.08it/s, est. speed input: 132.65 toks/s, output: 234.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:07<01:54,  1.26it/s, est. speed input: 135.17 toks/s, output: 306.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:09<02:36,  1.09s/it, est. speed input: 111.83 toks/s, output: 322.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12917.7 tokens/s, Running: 143 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:11<03:03,  1.29s/it, est. speed input: 98.46 toks/s, output: 351.62 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:13<03:44,  1.59s/it, est. speed input: 85.52 toks/s, output: 371.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:14<03:15,  1.40s/it, est. speed input: 84.30 toks/s, output: 430.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:15<03:14,  1.40s/it, est. speed input: 80.71 toks/s, output: 475.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:16<02:28,  1.08s/it, est. speed input: 83.04 toks/s, output: 549.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10945.2 tokens/s, Running: 138 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:22<05:46,  2.53s/it, est. speed input: 63.10 toks/s, output: 479.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:23<04:49,  2.13s/it, est. speed input: 62.43 toks/s, output: 535.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:24<04:13,  1.88s/it, est. speed input: 61.65 toks/s, output: 587.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:25<03:17,  1.47s/it, est. speed input: 62.74 toks/s, output: 654.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:25<02:34,  1.16s/it, est. speed input: 64.08 toks/s, output: 722.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:28<03:43,  1.69s/it, est. speed input: 59.66 toks/s, output: 726.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9691.5 tokens/s, Running: 132 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:32<05:08,  2.35s/it, est. speed input: 54.36 toks/s, output: 716.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:33<04:25,  2.04s/it, est. speed input: 54.10 toks/s, output: 765.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:34<03:21,  1.56s/it, est. speed input: 55.26 toks/s, output: 833.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:35<03:06,  1.46s/it, est. speed input: 55.12 toks/s, output: 881.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:35<02:24,  1.14s/it, est. speed input: 56.25 toks/s, output: 948.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:35<01:46,  1.18it/s, est. speed input: 57.66 toks/s, output: 1020.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:36<01:50,  1.14it/s, est. speed input: 57.84 toks/s, output: 1070.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:37<01:56,  1.07it/s, est. speed input: 57.84 toks/s, output: 1116.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:39<02:11,  1.07s/it, est. speed input: 57.39 toks/s, output: 1153.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:39<01:42,  1.19it/s, est. speed input: 58.43 toks/s, output: 1221.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:39<01:16,  1.57it/s, est. speed input: 59.73 toks/s, output: 1292.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:40<01:19,  1.51it/s, est. speed input: 60.14 toks/s, output: 1345.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8945.2 tokens/s, Running: 120 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:40<01:02,  1.91it/s, est. speed input: 61.33 toks/s, output: 1415.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:42<01:29,  1.32it/s, est. speed input: 60.91 toks/s, output: 1446.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:42<01:25,  1.37it/s, est. speed input: 61.40 toks/s, output: 1499.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:43<01:07,  1.70it/s, est. speed input: 62.98 toks/s, output: 1620.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:43<01:00,  1.89it/s, est. speed input: 63.90 toks/s, output: 1682.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:44<00:54,  2.07it/s, est. speed input: 65.39 toks/s, output: 1800.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:45<00:51,  2.16it/s, est. speed input: 66.14 toks/s, output: 1860.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:46<01:12,  1.52it/s, est. speed input: 65.63 toks/s, output: 1885.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:46<01:09,  1.57it/s, est. speed input: 66.13 toks/s, output: 1937.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:47<01:19,  1.35it/s, est. speed input: 65.98 toks/s, output: 1971.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:48<01:13,  1.45it/s, est. speed input: 66.45 toks/s, output: 2024.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:49<01:34,  1.13it/s, est. speed input: 65.83 toks/s, output: 2043.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:24:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6943.1 tokens/s, Running: 106 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:52<02:10,  1.24s/it, est. speed input: 64.32 toks/s, output: 2032.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:52<01:22,  1.24it/s, est. speed input: 65.93 toks/s, output: 2157.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:53<01:13,  1.39it/s, est. speed input: 66.53 toks/s, output: 2212.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:53<01:09,  1.45it/s, est. speed input: 66.89 toks/s, output: 2259.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:54<00:48,  2.05it/s, est. speed input: 68.60 toks/s, output: 2386.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:56<01:23,  1.18it/s, est. speed input: 67.13 toks/s, output: 2372.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:56<01:13,  1.33it/s, est. speed input: 67.68 toks/s, output: 2425.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:57<01:26,  1.11it/s, est. speed input: 67.17 toks/s, output: 2443.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:58<01:17,  1.22it/s, est. speed input: 67.50 toks/s, output: 2491.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:59<01:27,  1.08it/s, est. speed input: 67.14 toks/s, output: 2514.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:25:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6544.3 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [01:01<01:27,  1.05it/s, est. speed input: 66.93 toks/s, output: 2576.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [01:02<01:31,  1.00s/it, est. speed input: 66.66 toks/s, output: 2600.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [01:03<00:48,  1.81it/s, est. speed input: 69.16 toks/s, output: 2801.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:10<02:37,  1.81s/it, est. speed input: 63.14 toks/s, output: 2593.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:11<02:17,  1.59s/it, est. speed input: 63.29 toks/s, output: 2635.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:11<01:49,  1.29s/it, est. speed input: 63.85 toks/s, output: 2696.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:25:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6281.4 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:11<01:09,  1.19it/s, est. speed input: 65.29 toks/s, output: 2825.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:11<00:58,  1.40it/s, est. speed input: 65.88 toks/s, output: 2887.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:12<00:56,  1.44it/s, est. speed input: 66.16 toks/s, output: 2933.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:14<01:16,  1.05it/s, est. speed input: 65.45 toks/s, output: 2938.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:15<00:58,  1.33it/s, est. speed input: 66.27 toks/s, output: 3044.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:16<01:06,  1.15it/s, est. speed input: 65.98 toks/s, output: 3066.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:20<01:58,  1.57s/it, est. speed input: 63.74 toks/s, output: 2998.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:20<01:35,  1.27s/it, est. speed input: 64.14 toks/s, output: 3053.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:25:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5855.9 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:21<01:33,  1.26s/it, est. speed input: 63.89 toks/s, output: 3080.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:22<01:22,  1.13s/it, est. speed input: 64.01 toks/s, output: 3122.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:23<00:45,  1.53it/s, est. speed input: 65.60 toks/s, output: 3309.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:23<00:35,  1.90it/s, est. speed input: 66.57 toks/s, output: 3431.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:24<00:40,  1.66it/s, est. speed input: 66.58 toks/s, output: 3466.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:25<00:40,  1.64it/s, est. speed input: 66.79 toks/s, output: 3513.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:25<00:35,  1.85it/s, est. speed input: 67.26 toks/s, output: 3572.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:28<01:08,  1.07s/it, est. speed input: 65.91 toks/s, output: 3538.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:29<01:10,  1.13s/it, est. speed input: 65.64 toks/s, output: 3561.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:29<00:52,  1.18it/s, est. speed input: 66.23 toks/s, output: 3629.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:25:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4971.1 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:31<00:56,  1.06it/s, est. speed input: 66.03 toks/s, output: 3689.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:32<00:45,  1.31it/s, est. speed input: 66.56 toks/s, output: 3753.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:32<00:34,  1.67it/s, est. speed input: 67.15 toks/s, output: 3822.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:34<00:56,  1.00it/s, est. speed input: 66.31 toks/s, output: 3810.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:35<00:54,  1.02it/s, est. speed input: 66.32 toks/s, output: 3846.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:36<00:52,  1.04it/s, est. speed input: 66.31 toks/s, output: 3882.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:37<01:00,  1.12s/it, est. speed input: 65.90 toks/s, output: 3896.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:38<00:52,  1.02it/s, est. speed input: 66.10 toks/s, output: 3944.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:39<00:46,  1.11it/s, est. speed input: 66.24 toks/s, output: 3989.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:39<00:38,  1.34it/s, est. speed input: 66.60 toks/s, output: 4048.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:39<00:29,  1.72it/s, est. speed input: 67.08 toks/s, output: 4114.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:25:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4886.1 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:42<01:00,  1.24s/it, est. speed input: 65.83 toks/s, output: 4076.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:42<00:46,  1.04it/s, est. speed input: 66.21 toks/s, output: 4138.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:42<00:33,  1.41it/s, est. speed input: 66.73 toks/s, output: 4208.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:43<00:24,  1.87it/s, est. speed input: 67.25 toks/s, output: 4277.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:43<00:23,  1.93it/s, est. speed input: 67.50 toks/s, output: 4331.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:43<00:19,  2.28it/s, est. speed input: 67.91 toks/s, output: 4395.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:46<00:45,  1.07s/it, est. speed input: 66.87 toks/s, output: 4365.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:48<00:57,  1.37s/it, est. speed input: 66.13 toks/s, output: 4357.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:48<00:41,  1.01s/it, est. speed input: 66.60 toks/s, output: 4426.24 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:48<00:00,  1.01s/it, est. speed input: 89.62 toks/s, output: 7520.54 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:48<00:00,  1.47it/s, est. speed input: 89.62 toks/s, output: 7520.54 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:26:01 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 21:26:01 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 21:26:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:26:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:26:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:26:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 21:26:01 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:26:01 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:26:01 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 21:26:01 [abstract.py:306] It took 0.651587 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:26:01 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0187
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 5101.9502
+  • Training FLOPS: 54.52 TFLOPS (13.63 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 530.89s
+  • policy_training: 343.92s (64.8%)
+  • generation: 118.56s (22.3%)
+  • teacher_logprob_inference: 63.03s (11.9%)
+  • policy_training/sharding_data: 4.33s (0.8%)
+  • prepare_for_generation: 3.61s (0.7%)
+  • policy_training/submit_training_futures: 1.67s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.3%)
+  • training_prep: 0.91s (0.2%)
+  • teacher_logprob_inference_prep: 0.51s (0.1%)
+  • get_topk_logits/shard_data: 0.16s (0.0%)
+  • data_processing: 0.08s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 7: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 21:32:51 [abstract.py:324] It took 0.050418 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05259395e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05259395e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05259395e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05259395e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58884277e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63789215e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58884277e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63789215e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58884277e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63789215e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58884277e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63789215e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05260234e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60196838e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66014252e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.31330853e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.69574308e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #7] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41637466e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05260234e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60196838e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66014252e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73707199e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21341286e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #7] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.28975410e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05260234e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60196838e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66014252e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28503876e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.16653776e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #7] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14505692e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05260234e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60196838e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66014252e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31210709e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75941772e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #7] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32108307e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 7: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 21:32:53 [abstract.py:324] It took 0.041703 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13150.09it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:32:54 [loggers.py:236] Engine 000: Avg prompt throughput: 22.5 tokens/s, Avg generation throughput: 72.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:01<02:41,  1.02s/it, est. speed input: 59.11 toks/s, output: 3.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|▏         | 2/160 [00:01<01:18,  2.02it/s, est. speed input: 103.67 toks/s, output: 17.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:01<00:52,  2.99it/s, est. speed input: 141.18 toks/s, output: 38.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:01<00:18,  8.09it/s, est. speed input: 288.19 toks/s, output: 135.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:01<00:11, 13.44it/s, est. speed input: 425.39 toks/s, output: 254.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:02<00:16,  9.02it/s, est. speed input: 401.55 toks/s, output: 296.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:33:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12955.3 tokens/s, Running: 146 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:20<00:16,  9.02it/s, est. speed input: 401.55 toks/s, output: 296.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:33:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11198.1 tokens/s, Running: 146 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:24<08:02,  3.33s/it, est. speed input: 36.80 toks/s, output: 103.32 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 21:33:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10234.2 tokens/s, Running: 145 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) INFO 05-02 21:33:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9171.3 tokens/s, Running: 145 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) INFO 05-02 21:33:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8645.5 tokens/s, Running: 145 reqs, Waiting: 0 reqs, GPU KV cache usage: 23.9%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:53<18:16,  7.62s/it, est. speed input: 18.16 toks/s, output: 117.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:33:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8252.5 tokens/s, Running: 144 reqs, Waiting: 0 reqs, GPU KV cache usage: 27.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [01:01<18:33,  7.79s/it, est. speed input: 16.64 toks/s, output: 169.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:34:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7507.9 tokens/s, Running: 142 reqs, Waiting: 0 reqs, GPU KV cache usage: 30.6%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [01:17<18:35,  7.91s/it, est. speed input: 14.72 toks/s, output: 254.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:34:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7261.2 tokens/s, Running: 141 reqs, Waiting: 0 reqs, GPU KV cache usage: 33.6%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [01:29<20:07,  8.63s/it, est. speed input: 13.52 toks/s, output: 284.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [01:30<16:00,  6.91s/it, est. speed input: 14.02 toks/s, output: 343.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:34:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6658.2 tokens/s, Running: 139 reqs, Waiting: 0 reqs, GPU KV cache usage: 36.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [01:36<15:30,  6.74s/it, est. speed input: 13.74 toks/s, output: 383.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [01:40<13:37,  5.97s/it, est. speed input: 13.85 toks/s, output: 429.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:34:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6580.6 tokens/s, Running: 137 reqs, Waiting: 0 reqs, GPU KV cache usage: 38.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [01:44<12:11,  5.38s/it, est. speed input: 13.91 toks/s, output: 474.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [01:47<10:39,  4.73s/it, est. speed input: 14.07 toks/s, output: 520.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [01:47<08:00,  3.59s/it, est. speed input: 14.55 toks/s, output: 577.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [01:48<06:11,  2.80s/it, est. speed input: 14.98 toks/s, output: 632.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:34:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5951.4 tokens/s, Running: 133 reqs, Waiting: 0 reqs, GPU KV cache usage: 40.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [01:53<07:27,  3.39s/it, est. speed input: 14.87 toks/s, output: 664.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [01:56<06:55,  3.17s/it, est. speed input: 15.05 toks/s, output: 708.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:34:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5573.3 tokens/s, Running: 131 reqs, Waiting: 0 reqs, GPU KV cache usage: 42.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [02:03<09:19,  4.30s/it, est. speed input: 14.67 toks/s, output: 725.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [02:04<07:20,  3.41s/it, est. speed input: 15.01 toks/s, output: 775.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [02:07<06:57,  3.26s/it, est. speed input: 15.13 toks/s, output: 815.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [02:09<06:21,  3.00s/it, est. speed input: 15.32 toks/s, output: 857.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:35:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5976.2 tokens/s, Running: 127 reqs, Waiting: 0 reqs, GPU KV cache usage: 43.6%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [02:15<07:47,  3.71s/it, est. speed input: 15.15 toks/s, output: 879.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [02:16<06:01,  2.89s/it, est. speed input: 15.50 toks/s, output: 929.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:35:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4828.6 tokens/s, Running: 125 reqs, Waiting: 0 reqs, GPU KV cache usage: 45.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [02:28<11:50,  5.73s/it, est. speed input: 14.61 toks/s, output: 907.44 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:28<00:00,  5.73s/it, est. speed input: 65.21 toks/s, output: 7750.74 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:28<00:00,  1.08it/s, est. speed input: 65.21 toks/s, output: 7750.74 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:35:34 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 21:35:34 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 21:35:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:35:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:35:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:35:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:35:35 [abstract.py:306] It took 0.697702 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 21:35:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:35:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:35:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:35:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0151
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 7190.6499
+  • Training FLOPS: 63.88 TFLOPS (15.97 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 672.06s
+  • policy_training: 425.66s (63.3%)
+  • generation: 161.90s (24.1%)
+  • teacher_logprob_inference: 79.18s (11.8%)
+  • policy_training/sharding_data: 4.39s (0.7%)
+  • prepare_for_generation: 3.21s (0.5%)
+  • policy_training/submit_training_futures: 1.99s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.2%)
+  • training_prep: 1.18s (0.2%)
+  • teacher_logprob_inference_prep: 0.58s (0.1%)
+  • get_topk_logits/shard_data: 0.21s (0.0%)
+  • data_processing: 0.09s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 8: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:03 [abstract.py:324] It took 0.051344 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05207500e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05207500e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05207500e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05207500e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) 
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58805466e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63949127e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58805466e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63949127e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58805466e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63949127e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58805466e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63949127e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05206406e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58810349e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65840912e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.30795021e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.69154739e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #8] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41624115e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05206406e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58810349e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65840912e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73498383e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21238251e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #8] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.26063118e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05206406e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58810349e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65840912e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28620377e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.39581871e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #8] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13132095e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05206406e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58810349e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65840912e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31588707e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75938301e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #8] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32023590e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 8: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:06 [abstract.py:324] It took 0.044673 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12535.98it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:06 [loggers.py:236] Engine 000: Avg prompt throughput: 18.4 tokens/s, Avg generation throughput: 75.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:54,  1.38it/s, est. speed input: 81.64 toks/s, output: 16.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:17,  8.64it/s, est. speed input: 420.97 toks/s, output: 125.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:01<00:14, 10.15it/s, est. speed input: 493.59 toks/s, output: 176.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:01<00:18,  7.90it/s, est. speed input: 449.96 toks/s, output: 225.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:03<00:42,  3.45it/s, est. speed input: 267.40 toks/s, output: 241.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:08<02:28,  1.03s/it, est. speed input: 108.36 toks/s, output: 174.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:09<02:17,  1.04it/s, est. speed input: 107.65 toks/s, output: 245.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12920.9 tokens/s, Running: 144 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:12<03:11,  1.34s/it, est. speed input: 87.10 toks/s, output: 269.99 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:13<03:23,  1.43s/it, est. speed input: 80.56 toks/s, output: 317.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:14<02:43,  1.16s/it, est. speed input: 83.10 toks/s, output: 393.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:15<02:40,  1.14s/it, est. speed input: 81.13 toks/s, output: 446.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:15<02:01,  1.14it/s, est. speed input: 84.49 toks/s, output: 524.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:18<03:35,  1.56s/it, est. speed input: 72.58 toks/s, output: 511.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:20<03:34,  1.56s/it, est. speed input: 69.81 toks/s, output: 552.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10825.5 tokens/s, Running: 137 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9852.6 tokens/s, Running: 137 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:31<09:44,  4.30s/it, est. speed input: 47.17 toks/s, output: 435.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:32<07:31,  3.35s/it, est. speed input: 47.45 toks/s, output: 497.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:32<05:36,  2.51s/it, est. speed input: 48.62 toks/s, output: 566.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:34<05:07,  2.31s/it, est. speed input: 47.84 toks/s, output: 612.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:38<06:15,  2.85s/it, est. speed input: 44.27 toks/s, output: 623.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:39<04:34,  2.10s/it, est. speed input: 45.48 toks/s, output: 693.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:39<03:16,  1.51s/it, est. speed input: 46.79 toks/s, output: 765.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8675.9 tokens/s, Running: 130 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:41<03:24,  1.59s/it, est. speed input: 46.24 toks/s, output: 807.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:41<02:48,  1.32s/it, est. speed input: 46.93 toks/s, output: 868.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:43<03:03,  1.44s/it, est. speed input: 46.45 toks/s, output: 907.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:43<02:22,  1.13s/it, est. speed input: 47.40 toks/s, output: 973.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:44<01:58,  1.05it/s, est. speed input: 48.23 toks/s, output: 1035.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:45<01:58,  1.05it/s, est. speed input: 48.50 toks/s, output: 1087.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:46<02:10,  1.06s/it, est. speed input: 48.41 toks/s, output: 1130.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:44:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7435.7 tokens/s, Running: 123 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:54<06:05,  3.00s/it, est. speed input: 42.85 toks/s, output: 1044.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:56<05:22,  2.66s/it, est. speed input: 42.50 toks/s, output: 1079.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:45:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7027.1 tokens/s, Running: 121 reqs, Waiting: 0 reqs, GPU KV cache usage: 23.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [01:01<06:56,  3.47s/it, est. speed input: 39.77 toks/s, output: 1054.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [01:03<05:54,  2.97s/it, est. speed input: 39.59 toks/s, output: 1093.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [01:03<04:20,  2.21s/it, est. speed input: 40.28 toks/s, output: 1154.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [01:05<03:47,  1.94s/it, est. speed input: 40.41 toks/s, output: 1200.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [01:10<05:49,  3.01s/it, est. speed input: 38.10 toks/s, output: 1174.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:45:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7073.7 tokens/s, Running: 116 reqs, Waiting: 0 reqs, GPU KV cache usage: 25.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [01:11<04:37,  2.41s/it, est. speed input: 38.41 toks/s, output: 1225.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [01:12<03:53,  2.04s/it, est. speed input: 38.63 toks/s, output: 1273.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [01:16<05:01,  2.67s/it, est. speed input: 37.34 toks/s, output: 1272.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [01:20<05:42,  3.06s/it, est. speed input: 36.25 toks/s, output: 1276.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:45:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6906.0 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 28.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [01:23<05:11,  2.81s/it, est. speed input: 36.01 toks/s, output: 1309.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [01:23<03:46,  2.06s/it, est. speed input: 36.63 toks/s, output: 1370.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [01:28<05:23,  2.97s/it, est. speed input: 35.25 toks/s, output: 1357.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [01:30<04:55,  2.74s/it, est. speed input: 35.05 toks/s, output: 1391.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [01:30<03:30,  1.97s/it, est. speed input: 35.67 toks/s, output: 1454.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:45:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6231.5 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 29.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [01:37<06:11,  3.50s/it, est. speed input: 33.73 toks/s, output: 1414.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [01:39<05:01,  2.87s/it, est. speed input: 33.85 toks/s, output: 1459.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:45:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6193.5 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 31.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [01:43<05:49,  3.36s/it, est. speed input: 32.99 toks/s, output: 1461.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [01:49<06:44,  3.92s/it, est. speed input: 31.95 toks/s, output: 1455.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [01:49<05:05,  3.00s/it, est. speed input: 32.27 toks/s, output: 1508.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:45:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5513.7 tokens/s, Running: 102 reqs, Waiting: 0 reqs, GPU KV cache usage: 33.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [01:54<05:59,  3.56s/it, est. speed input: 31.42 toks/s, output: 1508.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [01:56<05:08,  3.08s/it, est. speed input: 31.43 toks/s, output: 1546.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [01:59<05:10,  3.13s/it, est. speed input: 31.08 toks/s, output: 1567.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:46:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5627.8 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 34.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [02:05<06:03,  3.71s/it, est. speed input: 30.29 toks/s, output: 1567.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [02:05<04:26,  2.75s/it, est. speed input: 30.65 toks/s, output: 1624.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [02:06<03:38,  2.28s/it, est. speed input: 30.85 toks/s, output: 1672.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [02:07<03:01,  1.91s/it, est. speed input: 31.08 toks/s, output: 1721.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [02:10<03:20,  2.13s/it, est. speed input: 30.93 toks/s, output: 1749.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [02:10<02:25,  1.56s/it, est. speed input: 31.34 toks/s, output: 1808.69 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:10<00:00,  1.56s/it, est. speed input: 74.83 toks/s, output: 7640.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [02:10<00:00,  1.22it/s, est. speed input: 74.83 toks/s, output: 7640.40 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:46:28 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 21:46:28 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 21:46:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:46:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:46:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:46:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:46:28 [abstract.py:306] It took 0.687896 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 21:46:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:46:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:46:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:46:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0128
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 6238.3687
+  • Training FLOPS: 61.96 TFLOPS (15.49 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 597.32s
+  • policy_training: 377.73s (63.2%)
+  • generation: 142.50s (23.9%)
+  • teacher_logprob_inference: 71.81s (12.0%)
+  • policy_training/sharding_data: 4.43s (0.7%)
+  • prepare_for_generation: 3.27s (0.5%)
+  • policy_training/submit_training_futures: 1.61s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.50s (0.3%)
+  • training_prep: 1.06s (0.2%)
+  • teacher_logprob_inference_prep: 0.58s (0.1%)
+  • data_processing: 0.15s (0.0%)
+  • get_topk_logits/shard_data: 0.14s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 9: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:01 [abstract.py:324] It took 0.050182 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05203574e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05203574e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05203574e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05203574e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58850975e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64056702e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58850975e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64056702e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58850975e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64056702e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58850975e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64056702e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05198438e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60255547e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66386261e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.30157490e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.63827229e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #9] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41612778e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05198438e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60255547e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66386261e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.74098816e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21620007e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #9] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23444328e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05198438e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60255547e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66386261e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27217331e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.27109480e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #9] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14231911e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05198438e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60255547e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66386261e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31916618e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.76215076e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #9] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32147202e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 9: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:04 [abstract.py:324] It took 0.043309 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13194.30it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:04 [loggers.py:236] Engine 000: Avg prompt throughput: 20.4 tokens/s, Avg generation throughput: 104.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:32,  1.73it/s, est. speed input: 103.52 toks/s, output: 24.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:04, 30.52it/s, est. speed input: 1416.59 toks/s, output: 431.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 37.50it/s, est. speed input: 1775.19 toks/s, output: 679.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:04, 30.22it/s, est. speed input: 1654.68 toks/s, output: 826.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:02<00:08, 14.97it/s, est. speed input: 1123.42 toks/s, output: 761.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12055.4 tokens/s, Running: 122 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:14<01:27,  1.37it/s, est. speed input: 177.29 toks/s, output: 433.93 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:14<01:17,  1.51it/s, est. speed input: 179.65 toks/s, output: 600.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:19<01:40,  1.13it/s, est. speed input: 144.02 toks/s, output: 698.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10272.5 tokens/s, Running: 113 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:20<01:32,  1.22it/s, est. speed input: 142.29 toks/s, output: 841.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:22<01:32,  1.19it/s, est. speed input: 136.04 toks/s, output: 948.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:22<01:22,  1.32it/s, est. speed input: 137.69 toks/s, output: 1031.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:22<01:15,  1.43it/s, est. speed input: 137.92 toks/s, output: 1103.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:23<01:09,  1.54it/s, est. speed input: 138.12 toks/s, output: 1173.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:23<00:59,  1.79it/s, est. speed input: 139.44 toks/s, output: 1252.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:24<00:59,  1.77it/s, est. speed input: 138.45 toks/s, output: 1311.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:25<01:10,  1.48it/s, est. speed input: 135.24 toks/s, output: 1347.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:25<00:57,  1.79it/s, est. speed input: 136.42 toks/s, output: 1424.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:25<00:36,  2.73it/s, est. speed input: 140.03 toks/s, output: 1590.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:27<00:50,  1.95it/s, est. speed input: 136.77 toks/s, output: 1674.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:28<00:54,  1.79it/s, est. speed input: 134.91 toks/s, output: 1772.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:29<01:05,  1.47it/s, est. speed input: 131.72 toks/s, output: 1791.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8722.7 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:30<01:17,  1.22it/s, est. speed input: 128.09 toks/s, output: 1803.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:31<01:06,  1.41it/s, est. speed input: 128.51 toks/s, output: 1870.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:31<00:56,  1.64it/s, est. speed input: 129.09 toks/s, output: 1940.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:31<00:36,  2.47it/s, est. speed input: 131.86 toks/s, output: 2100.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:32<00:40,  2.22it/s, est. speed input: 131.25 toks/s, output: 2149.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:32<00:35,  2.48it/s, est. speed input: 132.10 toks/s, output: 2220.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:33<00:34,  2.58it/s, est. speed input: 132.56 toks/s, output: 2285.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:33<00:30,  2.84it/s, est. speed input: 133.82 toks/s, output: 2418.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:33<00:24,  3.41it/s, est. speed input: 135.25 toks/s, output: 2499.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:35<00:47,  1.78it/s, est. speed input: 131.73 toks/s, output: 2490.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:35<00:41,  1.96it/s, est. speed input: 131.92 toks/s, output: 2604.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:37<01:07,  1.20it/s, est. speed input: 126.90 toks/s, output: 2560.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:39<01:12,  1.10it/s, est. speed input: 124.77 toks/s, output: 2573.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7281.4 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:40<01:06,  1.18it/s, est. speed input: 123.13 toks/s, output: 2648.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:42<01:12,  1.05it/s, est. speed input: 119.47 toks/s, output: 2680.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:43<01:00,  1.24it/s, est. speed input: 120.18 toks/s, output: 2751.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:45<01:31,  1.24s/it, est. speed input: 114.49 toks/s, output: 2677.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:46<01:19,  1.09s/it, est. speed input: 114.15 toks/s, output: 2727.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:47<01:12,  1.00s/it, est. speed input: 113.62 toks/s, output: 2772.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:47<01:03,  1.12it/s, est. speed input: 113.49 toks/s, output: 2825.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:48<00:50,  1.38it/s, est. speed input: 114.05 toks/s, output: 2895.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:48<00:51,  1.33it/s, est. speed input: 113.41 toks/s, output: 2934.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:49<00:42,  1.60it/s, est. speed input: 113.86 toks/s, output: 3002.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:54:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6177.5 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:50<00:57,  1.16it/s, est. speed input: 111.87 toks/s, output: 3004.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:51<00:43,  1.49it/s, est. speed input: 112.30 toks/s, output: 3125.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:51<00:39,  1.61it/s, est. speed input: 112.45 toks/s, output: 3184.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:52<00:39,  1.58it/s, est. speed input: 112.13 toks/s, output: 3230.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:53<00:47,  1.32it/s, est. speed input: 110.99 toks/s, output: 3252.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:53<00:29,  2.05it/s, est. speed input: 112.75 toks/s, output: 3410.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:54<00:25,  2.25it/s, est. speed input: 113.47 toks/s, output: 3537.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:56<00:38,  1.48it/s, est. speed input: 111.49 toks/s, output: 3527.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:57<00:42,  1.33it/s, est. speed input: 110.55 toks/s, output: 3551.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:57<00:36,  1.51it/s, est. speed input: 110.89 toks/s, output: 3614.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:58<00:32,  1.68it/s, est. speed input: 111.15 toks/s, output: 3675.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:58<00:34,  1.54it/s, est. speed input: 110.66 toks/s, output: 3712.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:59<00:40,  1.29it/s, est. speed input: 109.63 toks/s, output: 3731.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:55:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5198.2 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:00<00:36,  1.40it/s, est. speed input: 109.60 toks/s, output: 3784.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:01<00:40,  1.24it/s, est. speed input: 108.72 toks/s, output: 3807.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:01<00:33,  1.48it/s, est. speed input: 109.07 toks/s, output: 3873.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:02<00:24,  1.95it/s, est. speed input: 109.85 toks/s, output: 3953.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:02<00:20,  2.34it/s, est. speed input: 110.48 toks/s, output: 4026.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:02<00:21,  2.11it/s, est. speed input: 110.47 toks/s, output: 4076.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:03<00:14,  2.98it/s, est. speed input: 111.79 toks/s, output: 4229.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:03<00:14,  2.92it/s, est. speed input: 112.13 toks/s, output: 4293.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:04<00:20,  2.01it/s, est. speed input: 111.43 toks/s, output: 4319.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:04<00:12,  3.09it/s, est. speed input: 113.01 toks/s, output: 4484.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:05<00:20,  1.90it/s, est. speed input: 111.91 toks/s, output: 4492.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:06<00:15,  2.33it/s, est. speed input: 112.78 toks/s, output: 4631.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:06<00:14,  2.48it/s, est. speed input: 113.15 toks/s, output: 4698.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:07<00:18,  1.93it/s, est. speed input: 112.54 toks/s, output: 4727.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:09<00:27,  1.25it/s, est. speed input: 110.74 toks/s, output: 4707.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:55:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3930.5 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:11<00:42,  1.28s/it, est. speed input: 107.63 toks/s, output: 4626.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:12<00:34,  1.08s/it, est. speed input: 107.63 toks/s, output: 4679.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:12<00:25,  1.20it/s, est. speed input: 108.20 toks/s, output: 4755.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:13<00:28,  1.05it/s, est. speed input: 107.18 toks/s, output: 4764.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:14<00:24,  1.18it/s, est. speed input: 107.14 toks/s, output: 4816.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:14<00:20,  1.37it/s, est. speed input: 107.30 toks/s, output: 4877.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:14<00:14,  1.82it/s, est. speed input: 107.95 toks/s, output: 4960.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:15<00:13,  1.92it/s, est. speed input: 108.10 toks/s, output: 5020.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:16<00:13,  1.81it/s, est. speed input: 108.02 toks/s, output: 5070.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:16<00:11,  2.08it/s, est. speed input: 108.36 toks/s, output: 5140.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:17<00:15,  1.49it/s, est. speed input: 107.62 toks/s, output: 5158.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:17<00:11,  1.94it/s, est. speed input: 108.19 toks/s, output: 5239.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:18<00:13,  1.58it/s, est. speed input: 107.71 toks/s, output: 5272.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:18<00:07,  2.41it/s, est. speed input: 108.80 toks/s, output: 5435.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:20<00:13,  1.31it/s, est. speed input: 107.11 toks/s, output: 5406.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:21<00:12,  1.40it/s, est. speed input: 107.10 toks/s, output: 5462.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 21:55:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3383.4 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:22<00:14,  1.11it/s, est. speed input: 106.00 toks/s, output: 5464.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:22<00:10,  1.44it/s, est. speed input: 106.50 toks/s, output: 5549.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:24<00:13,  1.01it/s, est. speed input: 105.03 toks/s, output: 5532.11 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:24<00:00,  1.01it/s, est. speed input: 115.10 toks/s, output: 6887.95 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:24<00:00,  1.89it/s, est. speed input: 115.10 toks/s, output: 6887.95 toks/s]
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:55:35 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 21:55:35 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 21:55:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:55:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:55:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:55:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 21:55:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 21:55:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 21:55:36 [abstract.py:306] It took 0.632221 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 21:55:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 21:55:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0150
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3641.4126
+  • Training FLOPS: 50.82 TFLOPS (12.71 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 406.56s
+  • policy_training: 259.41s (63.8%)
+  • generation: 92.44s (22.7%)
+  • teacher_logprob_inference: 49.55s (12.2%)
+  • policy_training/sharding_data: 4.44s (1.1%)
+  • prepare_for_generation: 3.30s (0.8%)
+  • policy_training/submit_training_futures: 2.03s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.4%)
+  • training_prep: 0.95s (0.2%)
+  • teacher_logprob_inference_prep: 0.62s (0.2%)
+  • get_topk_logits/shard_data: 0.16s (0.0%)
+  • data_processing: 0.13s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 10: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:00:48 [abstract.py:324] It took 0.055490 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05245625e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05245625e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05245625e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05245625e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58901138e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63779907e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58901138e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63779907e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58901138e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63779907e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58901138e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63779907e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05234609e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60794754e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65442047e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.31707478e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.62830782e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #10] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41857544e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05234609e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60794754e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65442047e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73969040e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21984329e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #10] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23737221e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05234609e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60794754e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65442047e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27613449e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.22355890e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #10] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16436539e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05234609e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60794754e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65442047e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31583138e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75644398e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #10] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32423721e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 10: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:00:50 [abstract.py:324] It took 0.043825 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13565.57it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:00:51 [loggers.py:236] Engine 000: Avg prompt throughput: 29.7 tokens/s, Avg generation throughput: 24.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.07it/s, est. speed input: 122.13 toks/s, output: 33.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 21.71it/s, est. speed input: 1012.31 toks/s, output: 332.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 23.27it/s, est. speed input: 1151.97 toks/s, output: 505.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:07, 19.01it/s, est. speed input: 1067.55 toks/s, output: 610.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:10, 13.46it/s, est. speed input: 897.98 toks/s, output: 649.40 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:09<01:46,  1.27it/s, est. speed input: 156.95 toks/s, output: 214.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:01:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12501.7 tokens/s, Running: 135 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:11<01:43,  1.28it/s, est. speed input: 146.72 toks/s, output: 357.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:15<02:39,  1.20s/it, est. speed input: 111.17 toks/s, output: 347.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:16<02:35,  1.18s/it, est. speed input: 107.62 toks/s, output: 410.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:16<02:13,  1.03s/it, est. speed input: 109.17 toks/s, output: 488.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:18<02:42,  1.26s/it, est. speed input: 99.95 toks/s, output: 518.02 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 22:01:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10720.3 tokens/s, Running: 129 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:22<03:46,  1.77s/it, est. speed input: 86.89 toks/s, output: 520.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:22<02:59,  1.42s/it, est. speed input: 88.37 toks/s, output: 597.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:22<02:16,  1.08s/it, est. speed input: 90.52 toks/s, output: 677.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:23<01:49,  1.14it/s, est. speed input: 91.85 toks/s, output: 752.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:23<01:37,  1.28it/s, est. speed input: 92.27 toks/s, output: 818.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:28<03:59,  1.95s/it, est. speed input: 78.56 toks/s, output: 761.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:29<03:28,  1.71s/it, est. speed input: 77.60 toks/s, output: 814.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:01:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9589.5 tokens/s, Running: 122 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:30<03:07,  1.55s/it, est. speed input: 76.65 toks/s, output: 866.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:31<02:25,  1.22s/it, est. speed input: 77.57 toks/s, output: 936.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:34<03:42,  1.87s/it, est. speed input: 71.61 toks/s, output: 924.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:34<02:47,  1.42s/it, est. speed input: 72.69 toks/s, output: 996.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:35<02:21,  1.21s/it, est. speed input: 72.87 toks/s, output: 1057.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:35<01:43,  1.12it/s, est. speed input: 74.21 toks/s, output: 1133.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:01:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8642.6 tokens/s, Running: 116 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.2%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:40<04:09,  2.17s/it, est. speed input: 66.37 toks/s, output: 1070.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:41<03:23,  1.79s/it, est. speed input: 66.39 toks/s, output: 1127.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:42<02:48,  1.49s/it, est. speed input: 66.52 toks/s, output: 1185.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:43<01:42,  1.09it/s, est. speed input: 68.55 toks/s, output: 1328.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:43<01:22,  1.33it/s, est. speed input: 69.61 toks/s, output: 1400.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:44<01:31,  1.20it/s, est. speed input: 69.26 toks/s, output: 1445.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:45<01:35,  1.13it/s, est. speed input: 69.02 toks/s, output: 1491.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:47<02:20,  1.31s/it, est. speed input: 66.83 toks/s, output: 1495.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:48<01:54,  1.08s/it, est. speed input: 67.36 toks/s, output: 1558.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:49<02:02,  1.16s/it, est. speed input: 66.75 toks/s, output: 1593.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:50<01:38,  1.06it/s, est. speed input: 67.41 toks/s, output: 1658.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:50<01:11,  1.43it/s, est. speed input: 68.45 toks/s, output: 1733.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:01:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7762.2 tokens/s, Running: 103 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:51<01:17,  1.32it/s, est. speed input: 68.44 toks/s, output: 1780.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:51<01:16,  1.32it/s, est. speed input: 68.64 toks/s, output: 1832.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:52<01:05,  1.53it/s, est. speed input: 69.27 toks/s, output: 1895.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:52<00:56,  1.76it/s, est. speed input: 69.94 toks/s, output: 1960.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:53<00:48,  2.02it/s, est. speed input: 70.66 toks/s, output: 2025.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:53<00:47,  2.03it/s, est. speed input: 71.15 toks/s, output: 2084.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:54<00:48,  1.97it/s, est. speed input: 71.55 toks/s, output: 2141.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:55<01:05,  1.45it/s, est. speed input: 71.20 toks/s, output: 2174.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:55<00:56,  1.65it/s, est. speed input: 71.76 toks/s, output: 2236.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:56<01:07,  1.38it/s, est. speed input: 71.54 toks/s, output: 2274.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:57<01:09,  1.32it/s, est. speed input: 71.60 toks/s, output: 2318.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:58<01:07,  1.35it/s, est. speed input: 71.82 toks/s, output: 2367.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:59<01:09,  1.29it/s, est. speed input: 71.77 toks/s, output: 2410.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:59<00:54,  1.64it/s, est. speed input: 72.53 toks/s, output: 2478.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:01:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6813.1 tokens/s, Running: 88 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:01<01:10,  1.24it/s, est. speed input: 72.07 toks/s, output: 2545.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:03<01:31,  1.07s/it, est. speed input: 70.91 toks/s, output: 2547.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:03<01:15,  1.13it/s, est. speed input: 71.49 toks/s, output: 2610.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:03<00:58,  1.42it/s, est. speed input: 72.19 toks/s, output: 2678.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:04<00:58,  1.42it/s, est. speed input: 72.35 toks/s, output: 2726.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:05<01:04,  1.26it/s, est. speed input: 72.15 toks/s, output: 2760.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:06<01:11,  1.14it/s, est. speed input: 71.86 toks/s, output: 2791.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:06<00:52,  1.53it/s, est. speed input: 72.65 toks/s, output: 2863.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:07<00:54,  1.45it/s, est. speed input: 72.71 toks/s, output: 2907.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:08<00:51,  1.50it/s, est. speed input: 72.93 toks/s, output: 2957.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:08<00:45,  1.70it/s, est. speed input: 73.39 toks/s, output: 3016.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5950.6 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:11<01:32,  1.21s/it, est. speed input: 71.48 toks/s, output: 2979.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:11<01:08,  1.09it/s, est. speed input: 72.10 toks/s, output: 3046.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:12<01:22,  1.11s/it, est. speed input: 71.36 toks/s, output: 3056.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:13<01:05,  1.11it/s, est. speed input: 71.82 toks/s, output: 3116.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:14<01:10,  1.02it/s, est. speed input: 71.50 toks/s, output: 3143.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:15<00:57,  1.23it/s, est. speed input: 71.92 toks/s, output: 3243.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:17<01:09,  1.00s/it, est. speed input: 71.25 toks/s, output: 3254.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:17<00:52,  1.30it/s, est. speed input: 71.92 toks/s, output: 3326.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:19<00:54,  1.22it/s, est. speed input: 71.87 toks/s, output: 3402.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:20<00:55,  1.17it/s, est. speed input: 71.74 toks/s, output: 3437.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5268.4 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.2%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:20<00:46,  1.38it/s, est. speed input: 72.20 toks/s, output: 3499.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:21<00:47,  1.32it/s, est. speed input: 72.16 toks/s, output: 3538.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:21<00:42,  1.44it/s, est. speed input: 72.45 toks/s, output: 3592.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:22<00:35,  1.73it/s, est. speed input: 72.92 toks/s, output: 3656.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:22<00:29,  2.05it/s, est. speed input: 73.41 toks/s, output: 3720.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:22<00:22,  2.63it/s, est. speed input: 74.02 toks/s, output: 3791.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:25<01:06,  1.15s/it, est. speed input: 72.10 toks/s, output: 3735.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:25<00:49,  1.16it/s, est. speed input: 72.66 toks/s, output: 3804.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:26<00:37,  1.47it/s, est. speed input: 73.13 toks/s, output: 3870.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:26<00:32,  1.68it/s, est. speed input: 73.48 toks/s, output: 3929.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:27<00:32,  1.65it/s, est. speed input: 73.65 toks/s, output: 3977.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:27<00:24,  2.18it/s, est. speed input: 74.25 toks/s, output: 4049.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:27<00:28,  1.79it/s, est. speed input: 74.25 toks/s, output: 4090.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:28<00:23,  2.20it/s, est. speed input: 74.75 toks/s, output: 4157.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:28<00:21,  2.38it/s, est. speed input: 75.16 toks/s, output: 4219.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:28<00:19,  2.51it/s, est. speed input: 75.53 toks/s, output: 4279.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:29<00:28,  1.69it/s, est. speed input: 75.33 toks/s, output: 4306.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:30<00:18,  2.55it/s, est. speed input: 76.42 toks/s, output: 4445.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4723.2 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.7%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:30<00:16,  2.73it/s, est. speed input: 76.87 toks/s, output: 4507.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:30<00:15,  2.78it/s, est. speed input: 77.25 toks/s, output: 4567.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:30<00:12,  3.43it/s, est. speed input: 77.83 toks/s, output: 4639.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:31<00:15,  2.64it/s, est. speed input: 77.96 toks/s, output: 4685.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:32<00:18,  2.21it/s, est. speed input: 78.07 toks/s, output: 4730.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:33<00:23,  1.73it/s, est. speed input: 77.98 toks/s, output: 4763.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:33<00:19,  2.03it/s, est. speed input: 78.40 toks/s, output: 4826.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:33<00:17,  2.15it/s, est. speed input: 78.70 toks/s, output: 4883.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:34<00:16,  2.26it/s, est. speed input: 79.03 toks/s, output: 4941.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:35<00:21,  1.71it/s, est. speed input: 78.91 toks/s, output: 4971.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:35<00:15,  2.20it/s, est. speed input: 79.43 toks/s, output: 5042.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:35<00:14,  2.41it/s, est. speed input: 79.79 toks/s, output: 5103.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:36<00:20,  1.60it/s, est. speed input: 79.49 toks/s, output: 5123.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:36<00:12,  2.43it/s, est. speed input: 80.46 toks/s, output: 5263.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:37<00:15,  2.00it/s, est. speed input: 80.44 toks/s, output: 5300.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:38<00:15,  1.88it/s, est. speed input: 80.53 toks/s, output: 5346.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:39<00:19,  1.43it/s, est. speed input: 80.20 toks/s, output: 5364.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:40<00:17,  1.54it/s, est. speed input: 80.38 toks/s, output: 5417.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3882.8 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:40<00:15,  1.66it/s, est. speed input: 80.61 toks/s, output: 5471.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:40<00:13,  1.80it/s, est. speed input: 80.85 toks/s, output: 5527.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:41<00:09,  2.33it/s, est. speed input: 81.62 toks/s, output: 5658.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:41<00:00,  2.33it/s, est. speed input: 95.28 toks/s, output: 7514.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:41<00:00,  1.58it/s, est. speed input: 95.28 toks/s, output: 7514.25 toks/s]
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:02:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:02:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:02:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:02:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:02:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:02:40 [abstract.py:306] It took 0.662902 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:02:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:02:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:02:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0146
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4767.5127
+  • Training FLOPS: 55.44 TFLOPS (13.86 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 489.53s
+  • policy_training: 315.15s (64.4%)
+  • generation: 109.21s (22.3%)
+  • teacher_logprob_inference: 60.32s (12.3%)
+  • policy_training/sharding_data: 4.41s (0.9%)
+  • prepare_for_generation: 3.27s (0.7%)
+  • policy_training/submit_training_futures: 1.96s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.3%)
+  • training_prep: 0.83s (0.2%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.10s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 3/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 11: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:08:58 [abstract.py:324] It took 0.051890 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05289414e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05289414e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05289414e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05289414e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58946609e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63920135e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58946609e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63920135e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58946609e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63920135e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58946609e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63920135e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05285684e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61291084e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64915466e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.31463890e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.61991549e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #11] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41708618e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05285684e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61291084e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64915466e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.74582672e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.21513081e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #11] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23762360e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05285684e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61291084e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64915466e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27312317e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.76104259e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #11] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15797043e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05285684e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61291084e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64915466e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31387558e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75305862e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #11] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32382599e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 11: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:01 [abstract.py:324] It took 0.042610 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12795.07it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:02 [loggers.py:236] Engine 000: Avg prompt throughput: 24.9 tokens/s, Avg generation throughput: 9.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<02:14,  1.18it/s, est. speed input: 71.09 toks/s, output: 8.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:10, 13.88it/s, est. speed input: 636.52 toks/s, output: 141.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:01<00:06, 21.52it/s, est. speed input: 917.52 toks/s, output: 251.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:04, 30.76it/s, est. speed input: 1207.93 toks/s, output: 417.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 39.89it/s, est. speed input: 1478.98 toks/s, output: 645.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:04, 27.05it/s, est. speed input: 1365.21 toks/s, output: 802.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:07<00:36,  3.20it/s, est. speed input: 364.71 toks/s, output: 405.86 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11863.3 tokens/s, Running: 115 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:18<01:38,  1.15it/s, est. speed input: 157.47 toks/s, output: 441.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:18<01:34,  1.18it/s, est. speed input: 156.17 toks/s, output: 519.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10115.5 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:21<01:30,  1.21it/s, est. speed input: 147.64 toks/s, output: 727.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:24<01:50,  1.04s/it, est. speed input: 129.55 toks/s, output: 787.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:25<01:32,  1.13it/s, est. speed input: 131.11 toks/s, output: 945.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:26<01:36,  1.08it/s, est. speed input: 127.58 toks/s, output: 992.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8801.2 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.0%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:31<02:36,  1.52s/it, est. speed input: 110.63 toks/s, output: 933.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:33<02:43,  1.61s/it, est. speed input: 105.82 toks/s, output: 964.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:33<02:20,  1.39s/it, est. speed input: 105.88 toks/s, output: 1035.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:34<02:08,  1.28s/it, est. speed input: 104.82 toks/s, output: 1095.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:36<02:20,  1.42s/it, est. speed input: 101.34 toks/s, output: 1128.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:36<01:47,  1.10s/it, est. speed input: 102.44 toks/s, output: 1208.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:37<01:32,  1.05it/s, est. speed input: 102.61 toks/s, output: 1277.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:39<02:05,  1.31s/it, est. speed input: 98.38 toks/s, output: 1291.40 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:39<01:36,  1.01s/it, est. speed input: 99.19 toks/s, output: 1368.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:40<01:23,  1.13it/s, est. speed input: 99.18 toks/s, output: 1435.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7949.5 tokens/s, Running: 94 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:41<01:10,  1.31it/s, est. speed input: 99.58 toks/s, output: 1504.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:42<01:06,  1.36it/s, est. speed input: 99.22 toks/s, output: 1624.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:42<00:52,  1.71it/s, est. speed input: 100.39 toks/s, output: 1705.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:42<00:48,  1.85it/s, est. speed input: 100.85 toks/s, output: 1774.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:43<00:58,  1.51it/s, est. speed input: 100.04 toks/s, output: 1819.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:44<00:53,  1.64it/s, est. speed input: 100.33 toks/s, output: 1885.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:44<00:32,  2.61it/s, est. speed input: 102.60 toks/s, output: 2048.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:45<00:26,  3.19it/s, est. speed input: 104.31 toks/s, output: 2200.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:45<00:30,  2.72it/s, est. speed input: 104.38 toks/s, output: 2258.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:45<00:27,  2.94it/s, est. speed input: 105.10 toks/s, output: 2331.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:47<00:43,  1.83it/s, est. speed input: 103.79 toks/s, output: 2358.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:48<01:12,  1.09it/s, est. speed input: 100.87 toks/s, output: 2349.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:50<01:16,  1.03it/s, est. speed input: 99.78 toks/s, output: 2382.08 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 22:09:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6774.2 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:50<01:13,  1.05it/s, est. speed input: 99.20 toks/s, output: 2424.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:51<01:13,  1.04it/s, est. speed input: 98.48 toks/s, output: 2463.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:52<00:59,  1.25it/s, est. speed input: 98.94 toks/s, output: 2529.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:53<00:43,  1.67it/s, est. speed input: 99.90 toks/s, output: 2663.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:53<00:45,  1.59it/s, est. speed input: 99.69 toks/s, output: 2712.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:54<00:54,  1.30it/s, est. speed input: 98.72 toks/s, output: 2739.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:55<00:50,  1.37it/s, est. speed input: 98.73 toks/s, output: 2793.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:56<00:54,  1.27it/s, est. speed input: 98.15 toks/s, output: 2831.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:57<00:57,  1.19it/s, est. speed input: 97.54 toks/s, output: 2868.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:57<00:49,  1.36it/s, est. speed input: 97.83 toks/s, output: 2929.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:58<00:49,  1.34it/s, est. speed input: 97.51 toks/s, output: 2975.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:58<00:37,  1.75it/s, est. speed input: 98.33 toks/s, output: 3052.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:59<00:29,  2.12it/s, est. speed input: 99.20 toks/s, output: 3184.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5936.9 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.9%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:01<00:47,  1.30it/s, est. speed input: 97.49 toks/s, output: 3181.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:02<00:59,  1.03it/s, est. speed input: 96.09 toks/s, output: 3185.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:05<01:23,  1.40s/it, est. speed input: 93.30 toks/s, output: 3145.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:05<01:05,  1.11s/it, est. speed input: 93.70 toks/s, output: 3212.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:07<01:09,  1.20s/it, est. speed input: 92.65 toks/s, output: 3228.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:07<00:56,  1.00it/s, est. speed input: 92.86 toks/s, output: 3287.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:08<00:37,  1.47it/s, est. speed input: 93.84 toks/s, output: 3426.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:08<00:34,  1.56it/s, est. speed input: 94.04 toks/s, output: 3484.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:09<00:30,  1.74it/s, est. speed input: 94.41 toks/s, output: 3549.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:10<00:33,  1.54it/s, est. speed input: 94.17 toks/s, output: 3591.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4784.2 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:11<00:41,  1.23it/s, est. speed input: 93.39 toks/s, output: 3613.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:11<00:31,  1.58it/s, est. speed input: 94.01 toks/s, output: 3689.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:12<00:35,  1.37it/s, est. speed input: 93.61 toks/s, output: 3724.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:12<00:27,  1.72it/s, est. speed input: 94.18 toks/s, output: 3798.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:13<00:28,  1.66it/s, est. speed input: 94.16 toks/s, output: 3849.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:14<00:28,  1.61it/s, est. speed input: 94.20 toks/s, output: 3951.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:15<00:30,  1.44it/s, est. speed input: 93.87 toks/s, output: 3988.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:15<00:26,  1.60it/s, est. speed input: 94.16 toks/s, output: 4052.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:16<00:31,  1.35it/s, est. speed input: 93.69 toks/s, output: 4082.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:19<00:50,  1.24s/it, est. speed input: 91.48 toks/s, output: 4038.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:20<00:46,  1.16s/it, est. speed input: 91.13 toks/s, output: 4078.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:20<00:34,  1.15it/s, est. speed input: 91.71 toks/s, output: 4156.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4628.1 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.6%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:22<00:44,  1.17s/it, est. speed input: 90.35 toks/s, output: 4148.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:22<00:32,  1.15it/s, est. speed input: 90.89 toks/s, output: 4228.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:23<00:29,  1.24it/s, est. speed input: 90.91 toks/s, output: 4282.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:24<00:26,  1.31it/s, est. speed input: 90.90 toks/s, output: 4335.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:24<00:22,  1.53it/s, est. speed input: 91.18 toks/s, output: 4403.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:25<00:21,  1.53it/s, est. speed input: 91.20 toks/s, output: 4457.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:25<00:15,  2.01it/s, est. speed input: 91.74 toks/s, output: 4538.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:26<00:19,  1.63it/s, est. speed input: 91.47 toks/s, output: 4579.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:26<00:14,  2.10it/s, est. speed input: 92.01 toks/s, output: 4660.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:27<00:22,  1.28it/s, est. speed input: 91.13 toks/s, output: 4668.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:28<00:19,  1.45it/s, est. speed input: 91.33 toks/s, output: 4731.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:28<00:17,  1.51it/s, est. speed input: 91.41 toks/s, output: 4787.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:29<00:17,  1.46it/s, est. speed input: 91.34 toks/s, output: 4836.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3443.4 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:30<00:22,  1.13it/s, est. speed input: 90.66 toks/s, output: 4854.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:31<00:20,  1.19it/s, est. speed input: 90.58 toks/s, output: 4904.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:31<00:00,  1.19it/s, est. speed input: 106.43 toks/s, output: 7049.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:31<00:00,  1.75it/s, est. speed input: 106.43 toks/s, output: 7049.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:10:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:10:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:10:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:10:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:10:40 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:10:40 [abstract.py:306] It took 0.634861 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:10:40 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:10:40 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:10:40 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0165
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4037.3250
+  • Training FLOPS: 54.54 TFLOPS (13.64 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 426.08s
+  • policy_training: 271.29s (63.7%)
+  • generation: 98.81s (23.2%)
+  • teacher_logprob_inference: 51.05s (12.0%)
+  • policy_training/sharding_data: 4.49s (1.1%)
+  • prepare_for_generation: 3.37s (0.8%)
+  • policy_training/submit_training_futures: 2.21s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.4%)
+  • training_prep: 0.83s (0.2%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.11s (0.0%)
+  • get_topk_logits/shard_data: 0.08s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 12: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:04 [abstract.py:324] It took 0.056419 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05336270e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58775826e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64268188e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05336270e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58775826e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64268188e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05336270e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58775826e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64268188e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05336270e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58775826e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64268188e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05338906e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60055428e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65427704e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.30705261e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.57676554e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #12] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41723877e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05338906e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60055428e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65427704e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.74762192e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.22325039e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #12] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23667336e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05338906e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60055428e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65427704e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27594528e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.23796177e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #12] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16293106e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05338906e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60055428e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65427704e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31389008e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75566254e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #12] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32530533e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 12: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:07 [abstract.py:324] It took 0.041889 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 11347.84it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:08 [loggers.py:236] Engine 000: Avg prompt throughput: 28.9 tokens/s, Avg generation throughput: 8.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:38,  1.62it/s, est. speed input: 97.18 toks/s, output: 11.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:23,  6.72it/s, est. speed input: 326.11 toks/s, output: 59.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:12, 12.50it/s, est. speed input: 551.47 toks/s, output: 161.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:01<00:11, 13.36it/s, est. speed input: 618.58 toks/s, output: 247.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:01<00:10, 14.17it/s, est. speed input: 672.37 toks/s, output: 353.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:11, 12.70it/s, est. speed input: 665.91 toks/s, output: 444.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:12, 11.47it/s, est. speed input: 649.09 toks/s, output: 507.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:12, 10.94it/s, est. speed input: 641.86 toks/s, output: 583.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:02<00:23,  5.89it/s, est. speed input: 508.06 toks/s, output: 551.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:03<00:39,  3.45it/s, est. speed input: 397.16 toks/s, output: 494.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:03<00:39,  3.39it/s, est. speed input: 380.45 toks/s, output: 537.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:04<00:41,  3.22it/s, est. speed input: 362.11 toks/s, output: 574.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:05<01:07,  1.96it/s, est. speed input: 295.96 toks/s, output: 537.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:07<02:06,  1.04it/s, est. speed input: 215.85 toks/s, output: 464.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:08<01:59,  1.09it/s, est. speed input: 203.36 toks/s, output: 509.32 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12453.6 tokens/s, Running: 131 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:11<03:09,  1.46s/it, est. speed input: 157.59 toks/s, output: 468.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:13<02:26,  1.15s/it, est. speed input: 148.73 toks/s, output: 578.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:13<01:54,  1.11it/s, est. speed input: 152.21 toks/s, output: 660.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:13<01:51,  1.13it/s, est. speed input: 147.56 toks/s, output: 707.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:14<01:44,  1.20it/s, est. speed input: 144.66 toks/s, output: 760.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:15<01:27,  1.41it/s, est. speed input: 145.03 toks/s, output: 827.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:15<01:26,  1.42it/s, est. speed input: 142.40 toks/s, output: 878.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:16<01:27,  1.40it/s, est. speed input: 139.77 toks/s, output: 925.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:17<01:30,  1.33it/s, est. speed input: 136.75 toks/s, output: 967.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:18<01:28,  1.35it/s, est. speed input: 134.78 toks/s, output: 1014.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10591.7 tokens/s, Running: 119 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:20<02:30,  1.27s/it, est. speed input: 121.14 toks/s, output: 976.46 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:20<01:51,  1.06it/s, est. speed input: 123.03 toks/s, output: 1054.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:21<01:27,  1.34it/s, est. speed input: 124.30 toks/s, output: 1125.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:23<02:37,  1.36s/it, est. speed input: 112.29 toks/s, output: 1079.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:23<01:54,  1.00it/s, est. speed input: 113.97 toks/s, output: 1157.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:24<01:52,  1.01it/s, est. speed input: 111.92 toks/s, output: 1197.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:26<01:27,  1.28it/s, est. speed input: 111.96 toks/s, output: 1314.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:26<01:12,  1.53it/s, est. speed input: 113.20 toks/s, output: 1386.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:27<01:21,  1.34it/s, est. speed input: 111.29 toks/s, output: 1420.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:27<01:13,  1.49it/s, est. speed input: 111.65 toks/s, output: 1480.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:28<01:05,  1.65it/s, est. speed input: 112.03 toks/s, output: 1541.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:28<00:55,  1.92it/s, est. speed input: 112.87 toks/s, output: 1609.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:28<00:46,  2.30it/s, est. speed input: 114.10 toks/s, output: 1681.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:30<01:25,  1.22it/s, est. speed input: 109.53 toks/s, output: 1669.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9056.7 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:31<01:23,  1.24it/s, est. speed input: 108.86 toks/s, output: 1712.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:32<01:09,  1.46it/s, est. speed input: 108.98 toks/s, output: 1820.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:33<01:19,  1.27it/s, est. speed input: 107.18 toks/s, output: 1844.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:34<01:33,  1.07it/s, est. speed input: 104.75 toks/s, output: 1856.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:35<01:14,  1.33it/s, est. speed input: 105.75 toks/s, output: 1926.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:36<01:34,  1.03it/s, est. speed input: 103.05 toks/s, output: 1929.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:36<01:16,  1.27it/s, est. speed input: 103.67 toks/s, output: 1995.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:37<01:16,  1.25it/s, est. speed input: 103.02 toks/s, output: 2035.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:37<00:46,  2.04it/s, est. speed input: 105.60 toks/s, output: 2189.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:38<00:44,  2.10it/s, est. speed input: 105.92 toks/s, output: 2248.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:38<00:37,  2.42it/s, est. speed input: 106.90 toks/s, output: 2318.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7969.4 tokens/s, Running: 90 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:40<00:58,  1.54it/s, est. speed input: 104.76 toks/s, output: 2371.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:41<00:56,  1.57it/s, est. speed input: 104.76 toks/s, output: 2420.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:42<01:05,  1.34it/s, est. speed input: 103.51 toks/s, output: 2441.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:42<00:58,  1.50it/s, est. speed input: 103.93 toks/s, output: 2498.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:43<01:03,  1.34it/s, est. speed input: 103.10 toks/s, output: 2527.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:44<00:42,  1.95it/s, est. speed input: 104.79 toks/s, output: 2668.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:45<00:51,  1.63it/s, est. speed input: 103.98 toks/s, output: 2695.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:45<00:43,  1.89it/s, est. speed input: 104.70 toks/s, output: 2762.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:45<00:32,  2.50it/s, est. speed input: 106.31 toks/s, output: 2900.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:46<00:29,  2.66it/s, est. speed input: 106.99 toks/s, output: 2965.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:46<00:27,  2.81it/s, est. speed input: 107.62 toks/s, output: 3029.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:47<00:44,  1.72it/s, est. speed input: 106.13 toks/s, output: 3034.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:48<00:45,  1.64it/s, est. speed input: 105.79 toks/s, output: 3118.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:49<00:31,  2.35it/s, est. speed input: 107.73 toks/s, output: 3268.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:50<00:45,  1.59it/s, est. speed input: 105.93 toks/s, output: 3262.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:16:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6676.2 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:50<00:36,  1.94it/s, est. speed input: 106.78 toks/s, output: 3336.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:50<00:24,  2.78it/s, est. speed input: 108.59 toks/s, output: 3486.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:51<00:17,  3.78it/s, est. speed input: 110.52 toks/s, output: 3639.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:51<00:18,  3.53it/s, est. speed input: 110.91 toks/s, output: 3697.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:51<00:16,  3.86it/s, est. speed input: 111.65 toks/s, output: 3768.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:51<00:15,  4.18it/s, est. speed input: 112.41 toks/s, output: 3838.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:52<00:18,  3.27it/s, est. speed input: 112.97 toks/s, output: 3945.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:52<00:15,  3.82it/s, est. speed input: 113.83 toks/s, output: 4021.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:53<00:19,  3.12it/s, est. speed input: 113.89 toks/s, output: 4067.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:53<00:17,  3.40it/s, est. speed input: 114.62 toks/s, output: 4134.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:55<00:45,  1.28it/s, est. speed input: 111.34 toks/s, output: 4061.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:55<00:26,  2.11it/s, est. speed input: 113.17 toks/s, output: 4218.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:55<00:22,  2.49it/s, est. speed input: 113.94 toks/s, output: 4289.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:56<00:21,  2.47it/s, est. speed input: 114.18 toks/s, output: 4341.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:56<00:21,  2.47it/s, est. speed input: 114.50 toks/s, output: 4394.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:57<00:22,  2.25it/s, est. speed input: 114.64 toks/s, output: 4486.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:57<00:19,  2.55it/s, est. speed input: 115.25 toks/s, output: 4553.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:59<00:32,  1.49it/s, est. speed input: 113.33 toks/s, output: 4523.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:17:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5461.3 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:00<00:27,  1.67it/s, est. speed input: 113.35 toks/s, output: 4657.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:01<00:27,  1.62it/s, est. speed input: 113.02 toks/s, output: 4691.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:02<00:28,  1.57it/s, est. speed input: 112.71 toks/s, output: 4723.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:02<00:24,  1.72it/s, est. speed input: 112.97 toks/s, output: 4780.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:03<00:21,  1.99it/s, est. speed input: 113.45 toks/s, output: 4846.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:03<00:25,  1.59it/s, est. speed input: 112.67 toks/s, output: 4859.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:04<00:26,  1.50it/s, est. speed input: 112.30 toks/s, output: 4889.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:05<00:23,  1.68it/s, est. speed input: 112.54 toks/s, output: 4945.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:05<00:22,  1.67it/s, est. speed input: 112.41 toks/s, output: 4987.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:06<00:14,  2.44it/s, est. speed input: 113.63 toks/s, output: 5136.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:06<00:09,  3.49it/s, est. speed input: 115.15 toks/s, output: 5297.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:06<00:08,  3.80it/s, est. speed input: 115.76 toks/s, output: 5371.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:06<00:09,  3.47it/s, est. speed input: 116.03 toks/s, output: 5429.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:06<00:07,  4.16it/s, est. speed input: 116.75 toks/s, output: 5509.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:07<00:08,  3.43it/s, est. speed input: 116.95 toks/s, output: 5563.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:07<00:07,  3.90it/s, est. speed input: 117.53 toks/s, output: 5638.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:17:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4413.7 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:10<00:31,  1.14s/it, est. speed input: 112.79 toks/s, output: 5461.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:11<00:22,  1.18it/s, est. speed input: 113.46 toks/s, output: 5543.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:11<00:18,  1.41it/s, est. speed input: 113.70 toks/s, output: 5604.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:12<00:12,  1.82it/s, est. speed input: 114.21 toks/s, output: 5779.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:12<00:10,  2.02it/s, est. speed input: 114.55 toks/s, output: 5848.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:13<00:10,  1.97it/s, est. speed input: 114.53 toks/s, output: 5897.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:14<00:12,  1.62it/s, est. speed input: 113.87 toks/s, output: 5915.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:14<00:09,  1.92it/s, est. speed input: 114.31 toks/s, output: 5989.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:15<00:09,  1.87it/s, est. speed input: 114.25 toks/s, output: 6038.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:15<00:07,  2.37it/s, est. speed input: 114.82 toks/s, output: 6121.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:15<00:06,  2.49it/s, est. speed input: 115.08 toks/s, output: 6187.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:16<00:07,  1.93it/s, est. speed input: 114.66 toks/s, output: 6217.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:17<00:06,  2.01it/s, est. speed input: 114.78 toks/s, output: 6277.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:17<00:05,  2.24it/s, est. speed input: 115.10 toks/s, output: 6347.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:18<00:06,  1.88it/s, est. speed input: 114.79 toks/s, output: 6384.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:18<00:04,  2.22it/s, est. speed input: 115.29 toks/s, output: 6520.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:19<00:05,  1.63it/s, est. speed input: 114.45 toks/s, output: 6528.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:17:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2745.1 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:20<00:05,  1.53it/s, est. speed input: 114.17 toks/s, output: 6566.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:21<00:04,  1.52it/s, est. speed input: 113.96 toks/s, output: 6613.09 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:21<00:00,  1.52it/s, est. speed input: 119.16 toks/s, output: 7317.81 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:21<00:00,  1.97it/s, est. speed input: 119.16 toks/s, output: 7317.81 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:17:34 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:17:34 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:17:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:17:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:17:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:17:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:17:35 [abstract.py:306] It took 0.632479 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:17:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:17:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:17:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:17:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0146
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3721.5249
+  • Training FLOPS: 53.91 TFLOPS (13.48 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 399.82s
+  • policy_training: 247.85s (62.0%)
+  • generation: 88.08s (22.0%)
+  • teacher_logprob_inference: 58.67s (14.7%)
+  • policy_training/sharding_data: 4.31s (1.1%)
+  • prepare_for_generation: 3.31s (0.8%)
+  • policy_training/submit_training_futures: 1.83s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.4%)
+  • training_prep: 1.15s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.10s (0.0%)
+  • get_topk_logits/shard_data: 0.07s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 13: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:22:44 [abstract.py:324] It took 0.049459 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05393008e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05393008e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05393008e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05393008e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59297409e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64895859e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59297409e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64895859e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59297409e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64895859e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59297409e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64895859e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05397188e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61510391e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65596008e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.29808445e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.57782459e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #13] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41839386e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05397188e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61510391e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65596008e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.75091095e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.22827339e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #13] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23124237e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05397188e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61510391e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65596008e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27389374e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.35934067e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #13] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14762764e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05397188e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61510391e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65596008e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31009789e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.76403503e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #13] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32575363e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 13: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:22:47 [abstract.py:324] It took 0.046619 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12857.58it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-02 22:22:48 [loggers.py:236] Engine 000: Avg prompt throughput: 30.3 tokens/s, Avg generation throughput: 4.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:53,  2.99it/s, est. speed input: 182.36 toks/s, output: 23.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 19.64it/s, est. speed input: 955.93 toks/s, output: 203.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 39.66it/s, est. speed input: 1761.46 toks/s, output: 564.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 43.21it/s, est. speed input: 2013.07 toks/s, output: 794.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:04, 31.63it/s, est. speed input: 1795.53 toks/s, output: 900.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:05, 23.89it/s, est. speed input: 1569.12 toks/s, output: 1000.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:07<00:51,  2.38it/s, est. speed input: 304.65 toks/s, output: 405.01 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:10<00:59,  1.99it/s, est. speed input: 248.63 toks/s, output: 581.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:22:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12198.8 tokens/s, Running: 119 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:11<01:03,  1.85it/s, est. speed input: 227.53 toks/s, output: 695.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:13<01:07,  1.70it/s, est. speed input: 209.41 toks/s, output: 792.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:14<01:20,  1.42it/s, est. speed input: 191.01 toks/s, output: 802.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:15<01:21,  1.39it/s, est. speed input: 185.24 toks/s, output: 855.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:15<01:04,  1.71it/s, est. speed input: 186.66 toks/s, output: 1011.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:16<01:12,  1.52it/s, est. speed input: 179.07 toks/s, output: 1043.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:17<01:03,  1.71it/s, est. speed input: 179.64 toks/s, output: 1119.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:17<01:02,  1.73it/s, est. speed input: 177.55 toks/s, output: 1177.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:23:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10046.1 tokens/s, Running: 108 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:20<02:09,  1.21s/it, est. speed input: 153.51 toks/s, output: 1090.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:21<01:22,  1.27it/s, est. speed input: 156.66 toks/s, output: 1255.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:21<01:11,  1.45it/s, est. speed input: 156.94 toks/s, output: 1327.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:22<01:20,  1.28it/s, est. speed input: 152.24 toks/s, output: 1356.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:23<01:06,  1.54it/s, est. speed input: 153.06 toks/s, output: 1432.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:23<01:02,  1.60it/s, est. speed input: 152.02 toks/s, output: 1490.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:24<01:09,  1.45it/s, est. speed input: 149.04 toks/s, output: 1528.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:24<00:58,  1.69it/s, est. speed input: 149.50 toks/s, output: 1598.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:24<00:45,  2.18it/s, est. speed input: 151.13 toks/s, output: 1681.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:25<00:56,  1.72it/s, est. speed input: 148.28 toks/s, output: 1715.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:26<01:02,  1.53it/s, est. speed input: 145.86 toks/s, output: 1752.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:28<01:32,  1.03it/s, est. speed input: 139.09 toks/s, output: 1736.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:28<01:15,  1.24it/s, est. speed input: 139.13 toks/s, output: 1801.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:28<00:56,  1.65it/s, est. speed input: 140.56 toks/s, output: 1884.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:29<00:37,  2.38it/s, est. speed input: 142.76 toks/s, output: 2099.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:23:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8511.7 tokens/s, Running: 90 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:31<00:58,  1.52it/s, est. speed input: 137.62 toks/s, output: 2085.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:31<00:58,  1.50it/s, est. speed input: 136.50 toks/s, output: 2129.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:32<00:57,  1.50it/s, est. speed input: 135.55 toks/s, output: 2175.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:32<00:36,  2.31it/s, est. speed input: 138.36 toks/s, output: 2340.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:34<01:04,  1.30it/s, est. speed input: 132.59 toks/s, output: 2302.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:34<00:50,  1.64it/s, est. speed input: 133.77 toks/s, output: 2383.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:35<00:44,  1.86it/s, est. speed input: 134.21 toks/s, output: 2450.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:35<00:35,  2.30it/s, est. speed input: 135.36 toks/s, output: 2528.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:35<00:35,  2.23it/s, est. speed input: 135.21 toks/s, output: 2584.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:36<00:40,  1.96it/s, est. speed input: 134.43 toks/s, output: 2626.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:37<00:58,  1.34it/s, est. speed input: 131.26 toks/s, output: 2624.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:38<00:49,  1.54it/s, est. speed input: 131.42 toks/s, output: 2685.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:38<00:32,  2.29it/s, est. speed input: 133.22 toks/s, output: 2837.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:38<00:27,  2.68it/s, est. speed input: 134.16 toks/s, output: 2913.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:38<00:23,  3.08it/s, est. speed input: 135.08 toks/s, output: 2988.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:39<00:21,  3.34it/s, est. speed input: 135.87 toks/s, output: 3060.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:39<00:16,  4.37it/s, est. speed input: 137.98 toks/s, output: 3216.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:23:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7015.4 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:40<00:25,  2.74it/s, est. speed input: 136.74 toks/s, output: 3241.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:40<00:20,  3.29it/s, est. speed input: 137.82 toks/s, output: 3320.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:41<00:26,  2.55it/s, est. speed input: 137.13 toks/s, output: 3357.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:41<00:14,  4.41it/s, est. speed input: 140.76 toks/s, output: 3604.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:41<00:18,  3.48it/s, est. speed input: 140.47 toks/s, output: 3648.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:25,  2.43it/s, est. speed input: 139.13 toks/s, output: 3666.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:42<00:21,  2.89it/s, est. speed input: 140.06 toks/s, output: 3743.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:43<00:19,  3.13it/s, est. speed input: 140.68 toks/s, output: 3812.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:43<00:23,  2.52it/s, est. speed input: 140.07 toks/s, output: 3849.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:43<00:21,  2.72it/s, est. speed input: 140.51 toks/s, output: 3914.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:44<00:18,  3.05it/s, est. speed input: 141.12 toks/s, output: 3984.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:44<00:16,  3.34it/s, est. speed input: 141.75 toks/s, output: 4053.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:44<00:19,  2.82it/s, est. speed input: 141.51 toks/s, output: 4100.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:45<00:26,  2.07it/s, est. speed input: 140.40 toks/s, output: 4119.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:45<00:21,  2.52it/s, est. speed input: 141.14 toks/s, output: 4193.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:46<00:28,  1.85it/s, est. speed input: 139.77 toks/s, output: 4205.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:46<00:22,  2.22it/s, est. speed input: 140.37 toks/s, output: 4275.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:48<00:31,  1.61it/s, est. speed input: 138.68 toks/s, output: 4276.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:48<00:17,  2.79it/s, est. speed input: 140.86 toks/s, output: 4450.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:48<00:15,  2.99it/s, est. speed input: 141.34 toks/s, output: 4518.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:48<00:10,  4.09it/s, est. speed input: 143.15 toks/s, output: 4680.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:49<00:13,  3.26it/s, est. speed input: 142.84 toks/s, output: 4723.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:49<00:12,  3.50it/s, est. speed input: 143.40 toks/s, output: 4794.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:49<00:10,  3.89it/s, est. speed input: 144.10 toks/s, output: 4869.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:49<00:09,  4.52it/s, est. speed input: 144.98 toks/s, output: 4949.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:50<00:10,  3.77it/s, est. speed input: 145.06 toks/s, output: 5005.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:23:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5899.2 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:50<00:12,  3.10it/s, est. speed input: 144.95 toks/s, output: 5051.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:51<00:19,  1.98it/s, est. speed input: 143.48 toks/s, output: 5051.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:51<00:16,  2.28it/s, est. speed input: 143.85 toks/s, output: 5118.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:52<00:15,  2.35it/s, est. speed input: 143.91 toks/s, output: 5173.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:52<00:11,  2.90it/s, est. speed input: 144.89 toks/s, output: 5311.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:53<00:15,  2.16it/s, est. speed input: 143.78 toks/s, output: 5323.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:54<00:21,  1.47it/s, est. speed input: 141.51 toks/s, output: 5292.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:55<00:20,  1.50it/s, est. speed input: 141.03 toks/s, output: 5327.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:55<00:18,  1.59it/s, est. speed input: 140.78 toks/s, output: 5371.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:56<00:19,  1.48it/s, est. speed input: 139.92 toks/s, output: 5393.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:57<00:17,  1.61it/s, est. speed input: 139.77 toks/s, output: 5444.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:57<00:13,  2.07it/s, est. speed input: 140.42 toks/s, output: 5525.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:58<00:16,  1.58it/s, est. speed input: 139.05 toks/s, output: 5529.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:59<00:16,  1.51it/s, est. speed input: 138.37 toks/s, output: 5559.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:59<00:12,  1.89it/s, est. speed input: 138.85 toks/s, output: 5636.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:59<00:09,  2.43it/s, est. speed input: 139.53 toks/s, output: 5721.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:59<00:10,  2.20it/s, est. speed input: 139.25 toks/s, output: 5767.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:23:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3965.4 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:02<00:15,  1.32it/s, est. speed input: 136.22 toks/s, output: 5757.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:02<00:10,  1.78it/s, est. speed input: 136.98 toks/s, output: 5909.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:04<00:12,  1.37it/s, est. speed input: 135.09 toks/s, output: 5889.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:04<00:11,  1.45it/s, est. speed input: 134.86 toks/s, output: 5941.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:04<00:08,  1.68it/s, est. speed input: 135.12 toks/s, output: 6015.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:05<00:08,  1.64it/s, est. speed input: 134.68 toks/s, output: 6059.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:07<00:12,  1.02it/s, est. speed input: 131.72 toks/s, output: 5990.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:07<00:09,  1.20it/s, est. speed input: 131.68 toks/s, output: 6056.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:09<00:07,  1.28it/s, est. speed input: 130.71 toks/s, output: 6143.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:10<00:06,  1.34it/s, est. speed input: 130.38 toks/s, output: 6196.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:23:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2615.3 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:10<00:05,  1.50it/s, est. speed input: 130.41 toks/s, output: 6266.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:10<00:04,  1.69it/s, est. speed input: 130.55 toks/s, output: 6342.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:12<00:03,  1.38it/s, est. speed input: 128.97 toks/s, output: 6406.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:12<00:00,  1.38it/s, est. speed input: 133.10 toks/s, output: 6969.69 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:12<00:00,  2.20it/s, est. speed input: 133.10 toks/s, output: 6969.69 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:24:04 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:24:04 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:24:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:24:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:24:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:24:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:24:05 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:24:05 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:24:05 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 22:24:05 [abstract.py:306] It took 0.641500 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:24:05 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0167
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3165.5000
+  • Training FLOPS: 45.28 TFLOPS (11.32 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 377.32s
+  • policy_training: 249.60s (66.2%)
+  • generation: 77.86s (20.6%)
+  • teacher_logprob_inference: 44.75s (11.9%)
+  • policy_training/sharding_data: 4.32s (1.1%)
+  • prepare_for_generation: 3.34s (0.9%)
+  • policy_training/submit_training_futures: 1.72s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.4%)
+  • training_prep: 1.02s (0.3%)
+  • teacher_logprob_inference_prep: 0.59s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 14: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:02 [abstract.py:324] It took 0.055029 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05445742e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58632278e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65397339e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05445742e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58632278e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65397339e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05445742e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58632278e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65397339e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05445742e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58632278e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65397339e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05445742e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59329796e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66304321e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.28638897e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.57812405e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #14] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.41852890e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05445742e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59329796e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66304321e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.74530182e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.22071714e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #14] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.24763260e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05445742e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59329796e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66304321e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29287872e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.57728624e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #14] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13455048e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05445742e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59329796e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66304321e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31715088e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.76194248e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #14] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32666794e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 14: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:05 [abstract.py:324] It took 0.043347 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12853.64it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:05 [loggers.py:236] Engine 000: Avg prompt throughput: 31.5 tokens/s, Avg generation throughput: 11.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:59,  2.67it/s, est. speed input: 165.44 toks/s, output: 37.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:09, 16.22it/s, est. speed input: 820.36 toks/s, output: 268.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:12, 12.44it/s, est. speed input: 716.08 toks/s, output: 352.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:01<00:11, 12.49it/s, est. speed input: 726.62 toks/s, output: 429.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:01<00:12, 11.26it/s, est. speed input: 694.36 toks/s, output: 493.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:01<00:12, 11.12it/s, est. speed input: 688.82 toks/s, output: 585.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:13, 10.21it/s, est. speed input: 661.09 toks/s, output: 686.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:03<00:47,  2.94it/s, est. speed input: 339.76 toks/s, output: 453.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:04<00:45,  3.04it/s, est. speed input: 332.78 toks/s, output: 511.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12570.0 tokens/s, Running: 137 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:13<03:44,  1.65s/it, est. speed input: 111.67 toks/s, output: 271.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:13<03:11,  1.42s/it, est. speed input: 113.16 toks/s, output: 350.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:14<02:43,  1.22s/it, est. speed input: 113.76 toks/s, output: 424.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:14<02:20,  1.06s/it, est. speed input: 113.98 toks/s, output: 496.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:15<02:01,  1.09it/s, est. speed input: 114.26 toks/s, output: 566.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:16<02:06,  1.04it/s, est. speed input: 110.19 toks/s, output: 613.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:20<03:55,  1.81s/it, est. speed input: 90.79 toks/s, output: 573.81 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10800.6 tokens/s, Running: 130 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:21<03:51,  1.80s/it, est. speed input: 86.23 toks/s, output: 611.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:23<03:43,  1.75s/it, est. speed input: 82.80 toks/s, output: 652.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:23<02:42,  1.28s/it, est. speed input: 84.97 toks/s, output: 732.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:24<02:17,  1.09s/it, est. speed input: 85.21 toks/s, output: 796.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:25<02:27,  1.18s/it, est. speed input: 82.88 toks/s, output: 836.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:26<02:07,  1.03s/it, est. speed input: 83.09 toks/s, output: 897.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:28<03:01,  1.47s/it, est. speed input: 77.90 toks/s, output: 901.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:29<02:14,  1.10s/it, est. speed input: 79.36 toks/s, output: 976.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9627.2 tokens/s, Running: 122 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:30<02:26,  1.21s/it, est. speed input: 77.53 toks/s, output: 1011.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:31<02:00,  1.00s/it, est. speed input: 78.09 toks/s, output: 1076.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:33<02:31,  1.28s/it, est. speed input: 75.35 toks/s, output: 1094.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:33<02:07,  1.08s/it, est. speed input: 75.69 toks/s, output: 1155.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:35<02:29,  1.28s/it, est. speed input: 73.76 toks/s, output: 1179.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:36<01:37,  1.18it/s, est. speed input: 75.74 toks/s, output: 1316.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:36<01:19,  1.43it/s, est. speed input: 76.89 toks/s, output: 1387.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:38<01:54,  1.02s/it, est. speed input: 74.57 toks/s, output: 1398.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:39<02:13,  1.19s/it, est. speed input: 72.98 toks/s, output: 1420.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8598.9 tokens/s, Running: 110 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:40<01:11,  1.51it/s, est. speed input: 76.34 toks/s, output: 1635.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:41<01:14,  1.44it/s, est. speed input: 76.23 toks/s, output: 1682.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:42<01:24,  1.26it/s, est. speed input: 75.57 toks/s, output: 1716.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:44<01:54,  1.08s/it, est. speed input: 73.62 toks/s, output: 1719.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:45<02:04,  1.18s/it, est. speed input: 72.57 toks/s, output: 1742.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:47<02:04,  1.20s/it, est. speed input: 71.90 toks/s, output: 1775.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:47<01:32,  1.12it/s, est. speed input: 73.05 toks/s, output: 1849.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:48<01:38,  1.04it/s, est. speed input: 72.57 toks/s, output: 1885.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:48<01:25,  1.18it/s, est. speed input: 72.95 toks/s, output: 1942.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:50<01:34,  1.06it/s, est. speed input: 72.45 toks/s, output: 1975.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:29:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7628.6 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:50<01:09,  1.42it/s, est. speed input: 73.51 toks/s, output: 2048.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:50<00:56,  1.74it/s, est. speed input: 74.28 toks/s, output: 2115.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:50<00:45,  2.14it/s, est. speed input: 75.21 toks/s, output: 2184.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:50<00:26,  3.56it/s, est. speed input: 77.43 toks/s, output: 2336.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:51<00:23,  4.04it/s, est. speed input: 78.42 toks/s, output: 2407.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:53<01:09,  1.33it/s, est. speed input: 76.39 toks/s, output: 2387.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:53<00:59,  1.56it/s, est. speed input: 77.00 toks/s, output: 2449.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:53<00:52,  1.73it/s, est. speed input: 77.51 toks/s, output: 2508.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:55<00:50,  1.75it/s, est. speed input: 78.12 toks/s, output: 2610.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:55<00:46,  1.88it/s, est. speed input: 78.61 toks/s, output: 2669.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:55<00:39,  2.18it/s, est. speed input: 79.34 toks/s, output: 2734.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:56<00:42,  2.02it/s, est. speed input: 79.57 toks/s, output: 2783.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:56<00:43,  1.95it/s, est. speed input: 79.84 toks/s, output: 2833.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:57<00:26,  3.09it/s, est. speed input: 81.68 toks/s, output: 2980.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:59<01:10,  1.17it/s, est. speed input: 79.23 toks/s, output: 2930.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:00<01:03,  1.27it/s, est. speed input: 79.47 toks/s, output: 2980.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6574.6 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:00<00:53,  1.49it/s, est. speed input: 80.01 toks/s, output: 3040.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:01<01:03,  1.25it/s, est. speed input: 79.49 toks/s, output: 3062.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:02<01:04,  1.21it/s, est. speed input: 79.32 toks/s, output: 3096.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:02<00:50,  1.51it/s, est. speed input: 79.98 toks/s, output: 3161.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:04<01:10,  1.07it/s, est. speed input: 78.92 toks/s, output: 3161.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:04<00:52,  1.44it/s, est. speed input: 79.74 toks/s, output: 3233.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:05<00:48,  1.54it/s, est. speed input: 80.01 toks/s, output: 3283.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:05<00:38,  1.88it/s, est. speed input: 80.62 toks/s, output: 3348.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:06<00:46,  1.55it/s, est. speed input: 80.41 toks/s, output: 3380.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:08<01:27,  1.23s/it, est. speed input: 78.26 toks/s, output: 3330.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5628.0 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:11<01:58,  1.69s/it, est. speed input: 76.07 toks/s, output: 3278.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:12<01:34,  1.38s/it, est. speed input: 76.26 toks/s, output: 3326.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:12<00:59,  1.13it/s, est. speed input: 77.25 toks/s, output: 3452.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:13<00:54,  1.20it/s, est. speed input: 77.35 toks/s, output: 3498.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:13<00:41,  1.55it/s, est. speed input: 78.04 toks/s, output: 3570.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:14<00:42,  1.51it/s, est. speed input: 78.13 toks/s, output: 3613.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:14<00:32,  1.93it/s, est. speed input: 78.78 toks/s, output: 3684.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:15<00:34,  1.77it/s, est. speed input: 79.07 toks/s, output: 3777.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:16<00:31,  1.88it/s, est. speed input: 79.39 toks/s, output: 3833.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:16<00:27,  2.13it/s, est. speed input: 79.88 toks/s, output: 3896.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:17<00:22,  2.55it/s, est. speed input: 80.88 toks/s, output: 4022.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:18<00:38,  1.47it/s, est. speed input: 79.96 toks/s, output: 4017.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:19<00:36,  1.52it/s, est. speed input: 80.11 toks/s, output: 4065.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:19<00:34,  1.58it/s, est. speed input: 80.31 toks/s, output: 4115.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:20<00:17,  2.90it/s, est. speed input: 82.30 toks/s, output: 4336.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5360.2 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:21<00:27,  1.85it/s, est. speed input: 81.70 toks/s, output: 4345.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:21<00:21,  2.26it/s, est. speed input: 82.62 toks/s, output: 4473.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:22<00:22,  2.08it/s, est. speed input: 82.72 toks/s, output: 4518.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:23<00:23,  1.92it/s, est. speed input: 82.77 toks/s, output: 4561.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:23<00:21,  2.06it/s, est. speed input: 83.12 toks/s, output: 4620.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:23<00:18,  2.38it/s, est. speed input: 83.59 toks/s, output: 4686.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:24<00:22,  1.87it/s, est. speed input: 83.48 toks/s, output: 4719.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:25<00:21,  1.95it/s, est. speed input: 83.76 toks/s, output: 4774.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:26<00:31,  1.29it/s, est. speed input: 83.06 toks/s, output: 4775.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:27<00:30,  1.30it/s, est. speed input: 83.02 toks/s, output: 4814.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:27<00:25,  1.53it/s, est. speed input: 83.34 toks/s, output: 4874.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:28<00:22,  1.70it/s, est. speed input: 83.65 toks/s, output: 4931.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:28<00:15,  2.37it/s, est. speed input: 84.62 toks/s, output: 5067.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:29<00:17,  2.03it/s, est. speed input: 84.64 toks/s, output: 5107.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:29<00:13,  2.47it/s, est. speed input: 85.48 toks/s, output: 5237.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4419.3 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:30<00:11,  2.64it/s, est. speed input: 86.18 toks/s, output: 5360.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:30<00:10,  2.73it/s, est. speed input: 86.56 toks/s, output: 5423.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:31<00:15,  1.90it/s, est. speed input: 86.20 toks/s, output: 5442.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:32<00:13,  1.94it/s, est. speed input: 86.58 toks/s, output: 5547.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:33<00:11,  2.27it/s, est. speed input: 87.06 toks/s, output: 5620.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:33<00:12,  2.03it/s, est. speed input: 87.12 toks/s, output: 5663.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:34<00:11,  2.02it/s, est. speed input: 87.32 toks/s, output: 5716.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:34<00:07,  3.00it/s, est. speed input: 88.41 toks/s, output: 5869.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:35<00:08,  2.41it/s, est. speed input: 88.39 toks/s, output: 5910.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:37<00:18,  1.08it/s, est. speed input: 86.76 toks/s, output: 5844.02 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:37<00:00,  1.08it/s, est. speed input: 99.20 toks/s, output: 7521.60 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:37<00:00,  1.64it/s, est. speed input: 99.20 toks/s, output: 7521.60 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:49 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:49 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:30:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:30:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:30:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:30:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:30:49 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:30:49 [abstract.py:306] It took 0.634058 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:30:49 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:30:49 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:30:49 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0148
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4591.1123
+  • Training FLOPS: 55.18 TFLOPS (13.79 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 471.78s
+  • policy_training: 303.38s (64.3%)
+  • generation: 104.89s (22.2%)
+  • teacher_logprob_inference: 58.44s (12.4%)
+  • policy_training/sharding_data: 4.30s (0.9%)
+  • prepare_for_generation: 3.26s (0.7%)
+  • policy_training/submit_training_futures: 2.15s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.48s (0.3%)
+  • teacher_logprob_inference_prep: 0.83s (0.2%)
+  • training_prep: 0.81s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 15: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:36:54 [abstract.py:324] It took 0.050956 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05467383e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58931198e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65800171e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05467383e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58931198e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65800171e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05467383e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58931198e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65800171e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05467383e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58931198e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65800171e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05473418e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59832191e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66427765e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.28496208e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.54002953e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #15] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42068237e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05473418e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59832191e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66427765e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73403931e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.23019638e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #15] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.24275475e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05473418e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59832191e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66427765e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28473434e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.01856089e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #15] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14541626e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05473418e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59832191e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66427765e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31042709e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75708008e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #15] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32710831e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 15: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:36:57 [abstract.py:324] It took 0.043119 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 11130.46it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:36:57 [loggers.py:236] Engine 000: Avg prompt throughput: 25.6 tokens/s, Avg generation throughput: 57.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:55,  2.88it/s, est. speed input: 181.36 toks/s, output: 46.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:06, 22.38it/s, est. speed input: 1088.77 toks/s, output: 361.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 22.35it/s, est. speed input: 1179.70 toks/s, output: 507.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 23.99it/s, est. speed input: 1268.33 toks/s, output: 672.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:08, 15.59it/s, est. speed input: 1032.37 toks/s, output: 705.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:08, 15.82it/s, est. speed input: 1031.09 toks/s, output: 858.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:03<00:33,  3.99it/s, est. speed input: 451.86 toks/s, output: 552.97 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:08<01:37,  1.34it/s, est. speed input: 198.29 toks/s, output: 389.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:09<01:28,  1.46it/s, est. speed input: 199.38 toks/s, output: 469.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:37:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12436.3 tokens/s, Running: 130 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:10<01:48,  1.19it/s, est. speed input: 174.19 toks/s, output: 486.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:11<01:39,  1.29it/s, est. speed input: 172.27 toks/s, output: 556.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:11<01:31,  1.39it/s, est. speed input: 170.13 toks/s, output: 623.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:12<01:15,  1.66it/s, est. speed input: 172.26 toks/s, output: 702.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:13<01:36,  1.30it/s, est. speed input: 160.38 toks/s, output: 725.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:14<01:57,  1.06it/s, est. speed input: 148.75 toks/s, output: 743.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:16<02:16,  1.11s/it, est. speed input: 138.31 toks/s, output: 761.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:16<01:44,  1.17it/s, est. speed input: 140.27 toks/s, output: 841.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:16<01:22,  1.47it/s, est. speed input: 141.95 toks/s, output: 918.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:17<01:05,  1.83it/s, est. speed input: 143.70 toks/s, output: 995.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:19<01:57,  1.01it/s, est. speed input: 131.45 toks/s, output: 976.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:19<01:36,  1.22it/s, est. speed input: 131.79 toks/s, output: 1043.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:37:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10482.0 tokens/s, Running: 117 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:20<01:18,  1.48it/s, est. speed input: 131.22 toks/s, output: 1165.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:20<01:06,  1.73it/s, est. speed input: 132.38 toks/s, output: 1237.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:21<01:02,  1.83it/s, est. speed input: 132.40 toks/s, output: 1298.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:21<00:56,  1.99it/s, est. speed input: 132.94 toks/s, output: 1362.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:22<00:43,  2.56it/s, est. speed input: 135.60 toks/s, output: 1505.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:22<00:36,  3.05it/s, est. speed input: 137.74 toks/s, output: 1584.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:23<01:13,  1.47it/s, est. speed input: 130.47 toks/s, output: 1559.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:25<01:51,  1.03s/it, est. speed input: 122.78 toks/s, output: 1526.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:28<02:24,  1.35s/it, est. speed input: 115.39 toks/s, output: 1492.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:28<01:57,  1.11s/it, est. speed input: 115.51 toks/s, output: 1554.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:28<01:26,  1.21it/s, est. speed input: 117.21 toks/s, output: 1634.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:29<00:52,  1.95it/s, est. speed input: 120.28 toks/s, output: 1791.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:29<00:44,  2.31it/s, est. speed input: 121.52 toks/s, output: 1866.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:29<00:29,  3.40it/s, est. speed input: 124.89 toks/s, output: 2026.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:37:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9065.0 tokens/s, Running: 100 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:31<01:02,  1.57it/s, est. speed input: 119.62 toks/s, output: 1996.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:31<00:51,  1.89it/s, est. speed input: 120.73 toks/s, output: 2069.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:31<00:50,  1.91it/s, est. speed input: 120.80 toks/s, output: 2122.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:33<00:50,  1.88it/s, est. speed input: 120.47 toks/s, output: 2221.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:33<00:55,  1.69it/s, est. speed input: 119.40 toks/s, output: 2254.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:35<01:11,  1.31it/s, est. speed input: 116.76 toks/s, output: 2258.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:35<00:48,  1.88it/s, est. speed input: 118.96 toks/s, output: 2402.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:36<00:57,  1.55it/s, est. speed input: 117.25 toks/s, output: 2420.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:36<00:51,  1.74it/s, est. speed input: 117.80 toks/s, output: 2482.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:37<00:40,  2.16it/s, est. speed input: 118.95 toks/s, output: 2558.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:37<00:50,  1.72it/s, est. speed input: 117.74 toks/s, output: 2583.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:38<00:56,  1.51it/s, est. speed input: 116.70 toks/s, output: 2611.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:39<00:46,  1.84it/s, est. speed input: 117.55 toks/s, output: 2680.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:39<00:35,  2.35it/s, est. speed input: 119.01 toks/s, output: 2812.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:39<00:28,  2.88it/s, est. speed input: 120.23 toks/s, output: 2891.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:37:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7505.6 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:40<00:29,  2.69it/s, est. speed input: 120.85 toks/s, output: 3002.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:40<00:29,  2.67it/s, est. speed input: 121.19 toks/s, output: 3059.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:41<00:41,  1.89it/s, est. speed input: 119.72 toks/s, output: 3070.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:42<00:34,  2.26it/s, est. speed input: 120.64 toks/s, output: 3141.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:42<00:33,  2.29it/s, est. speed input: 120.84 toks/s, output: 3196.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:42<00:26,  2.83it/s, est. speed input: 121.89 toks/s, output: 3271.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:42<00:14,  4.91it/s, est. speed input: 125.59 toks/s, output: 3508.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:43<00:14,  4.99it/s, est. speed input: 127.29 toks/s, output: 3646.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:44<00:26,  2.52it/s, est. speed input: 125.48 toks/s, output: 3685.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:45<00:33,  1.99it/s, est. speed input: 124.18 toks/s, output: 3693.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:46<00:28,  2.30it/s, est. speed input: 124.98 toks/s, output: 3765.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:46<00:28,  2.31it/s, est. speed input: 125.15 toks/s, output: 3816.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:46<00:20,  3.14it/s, est. speed input: 126.90 toks/s, output: 3963.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:47<00:20,  3.08it/s, est. speed input: 127.28 toks/s, output: 4020.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:47<00:17,  3.40it/s, est. speed input: 128.02 toks/s, output: 4089.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:47<00:17,  3.47it/s, est. speed input: 128.65 toks/s, output: 4152.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:47<00:16,  3.55it/s, est. speed input: 129.27 toks/s, output: 4215.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:48<00:24,  2.37it/s, est. speed input: 128.38 toks/s, output: 4233.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:49<00:29,  1.96it/s, est. speed input: 127.71 toks/s, output: 4257.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:50<00:31,  1.76it/s, est. speed input: 127.11 toks/s, output: 4284.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:37:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6371.5 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:50<00:21,  2.56it/s, est. speed input: 128.65 toks/s, output: 4428.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:50<00:14,  3.42it/s, est. speed input: 130.74 toks/s, output: 4638.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:51<00:13,  3.78it/s, est. speed input: 131.52 toks/s, output: 4712.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:52<00:17,  2.80it/s, est. speed input: 131.19 toks/s, output: 4791.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:53<00:26,  1.79it/s, est. speed input: 128.98 toks/s, output: 4758.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:54<00:19,  2.25it/s, est. speed input: 130.01 toks/s, output: 4891.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:54<00:19,  2.23it/s, est. speed input: 130.11 toks/s, output: 4986.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:55<00:16,  2.54it/s, est. speed input: 130.80 toks/s, output: 5059.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:55<00:17,  2.38it/s, est. speed input: 130.71 toks/s, output: 5101.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:56<00:21,  1.90it/s, est. speed input: 129.79 toks/s, output: 5112.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:57<00:22,  1.76it/s, est. speed input: 129.29 toks/s, output: 5140.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:57<00:20,  1.89it/s, est. speed input: 129.46 toks/s, output: 5191.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:58<00:23,  1.55it/s, est. speed input: 128.39 toks/s, output: 5197.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:59<00:29,  1.22it/s, est. speed input: 126.71 toks/s, output: 5179.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:00<00:16,  2.00it/s, est. speed input: 128.30 toks/s, output: 5342.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:37:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4732.6 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:00<00:16,  2.06it/s, est. speed input: 128.40 toks/s, output: 5394.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:00<00:14,  2.21it/s, est. speed input: 128.62 toks/s, output: 5454.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:01<00:16,  1.86it/s, est. speed input: 128.02 toks/s, output: 5477.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:02<00:15,  1.97it/s, est. speed input: 128.13 toks/s, output: 5532.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:02<00:12,  2.34it/s, est. speed input: 128.66 toks/s, output: 5604.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:02<00:10,  2.77it/s, est. speed input: 129.20 toks/s, output: 5679.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:03<00:17,  1.56it/s, est. speed input: 127.46 toks/s, output: 5654.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:03<00:09,  2.59it/s, est. speed input: 129.06 toks/s, output: 5827.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:04<00:08,  2.95it/s, est. speed input: 129.66 toks/s, output: 5903.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:04<00:09,  2.53it/s, est. speed input: 129.47 toks/s, output: 5946.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:05<00:06,  3.42it/s, est. speed input: 130.72 toks/s, output: 6106.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:05<00:05,  3.76it/s, est. speed input: 131.30 toks/s, output: 6184.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:05<00:03,  5.09it/s, est. speed input: 132.84 toks/s, output: 6355.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:05<00:03,  5.32it/s, est. speed input: 133.45 toks/s, output: 6435.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:05<00:02,  5.73it/s, est. speed input: 134.14 toks/s, output: 6517.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:08<00:12,  1.23it/s, est. speed input: 129.75 toks/s, output: 6360.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:09<00:10,  1.29it/s, est. speed input: 129.36 toks/s, output: 6397.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:09<00:10,  1.23it/s, est. speed input: 128.53 toks/s, output: 6413.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:10<00:07,  1.62it/s, est. speed input: 129.21 toks/s, output: 6502.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:10<00:05,  2.14it/s, est. speed input: 129.91 toks/s, output: 6592.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:10<00:03,  2.71it/s, est. speed input: 130.53 toks/s, output: 6680.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:38:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3239.1 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:10<00:02,  2.89it/s, est. speed input: 131.08 toks/s, output: 6821.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:12<00:03,  1.82it/s, est. speed input: 129.79 toks/s, output: 6812.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:12<00:02,  2.38it/s, est. speed input: 130.63 toks/s, output: 6974.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:13<00:01,  2.02it/s, est. speed input: 130.15 toks/s, output: 7007.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:14<00:01,  1.63it/s, est. speed input: 129.25 toks/s, output: 7021.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:14<00:01,  1.56it/s, est. speed input: 128.84 toks/s, output: 7062.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:15<00:00,  1.77it/s, est. speed input: 129.04 toks/s, output: 7136.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  1.77it/s, est. speed input: 129.83 toks/s, output: 7245.54 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  2.12it/s, est. speed input: 129.83 toks/s, output: 7245.54 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:38:17 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:38:17 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:38:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:38:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:38:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:38:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:38:18 [abstract.py:306] It took 0.603610 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:38:18 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:38:18 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:38:18 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:38:18 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0133
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3412.5625
+  • Training FLOPS: 48.73 TFLOPS (12.18 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 381.57s
+  • policy_training: 249.33s (65.3%)
+  • generation: 80.91s (21.2%)
+  • teacher_logprob_inference: 46.59s (12.2%)
+  • policy_training/sharding_data: 4.38s (1.1%)
+  • prepare_for_generation: 3.28s (0.9%)
+  • policy_training/submit_training_futures: 1.74s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.4%)
+  • training_prep: 0.81s (0.2%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 4/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 16: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:16 [abstract.py:324] It took 0.055810 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05502344e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59092636e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65519257e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05502344e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59092636e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65519257e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05502344e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59092636e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65519257e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05502344e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59092636e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65519257e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05492930e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59155807e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05492930e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59155807e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66194916e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05492930e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59155807e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66194916e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05492930e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59155807e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66194916e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66194916e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.30039864e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.39776707e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #16] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42178772e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71606293e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.22635746e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #16] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23835907e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28110199e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.02875090e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #16] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.10815887e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31496811e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74832954e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #16] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32802795e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 16: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:19 [abstract.py:324] It took 0.042796 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12159.83it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:19 [loggers.py:236] Engine 000: Avg prompt throughput: 31.2 tokens/s, Avg generation throughput: 20.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:14,  2.14it/s, est. speed input: 128.60 toks/s, output: 23.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:11, 13.08it/s, est. speed input: 634.15 toks/s, output: 166.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 28.54it/s, est. speed input: 1242.17 toks/s, output: 432.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:06, 22.61it/s, est. speed input: 1178.98 toks/s, output: 517.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:06, 19.85it/s, est. speed input: 1129.59 toks/s, output: 646.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:12, 10.69it/s, est. speed input: 824.28 toks/s, output: 627.13 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:04<00:36,  3.64it/s, est. speed input: 423.42 toks/s, output: 455.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:06<01:04,  2.00it/s, est. speed input: 272.67 toks/s, output: 424.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:09<01:48,  1.19it/s, est. speed input: 192.68 toks/s, output: 378.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12138.7 tokens/s, Running: 129 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:11<02:00,  1.06it/s, est. speed input: 172.98 toks/s, output: 417.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:11<01:46,  1.19it/s, est. speed input: 172.51 toks/s, output: 491.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:13<02:11,  1.04s/it, est. speed input: 154.05 toks/s, output: 513.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:15<02:24,  1.16s/it, est. speed input: 142.04 toks/s, output: 547.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:16<02:28,  1.20s/it, est. speed input: 134.44 toks/s, output: 591.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:18<03:04,  1.50s/it, est. speed input: 120.72 toks/s, output: 603.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10657.2 tokens/s, Running: 123 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:20<03:17,  1.62s/it, est. speed input: 112.46 toks/s, output: 633.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:22<03:13,  1.60s/it, est. speed input: 107.32 toks/s, output: 674.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:23<03:03,  1.53s/it, est. speed input: 103.73 toks/s, output: 721.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:24<02:27,  1.24s/it, est. speed input: 103.94 toks/s, output: 791.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:25<02:00,  1.03s/it, est. speed input: 102.44 toks/s, output: 909.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:25<01:39,  1.17it/s, est. speed input: 103.44 toks/s, output: 983.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:27<01:59,  1.04s/it, est. speed input: 99.75 toks/s, output: 1011.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:27<01:33,  1.22it/s, est. speed input: 101.10 toks/s, output: 1088.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9299.6 tokens/s, Running: 114 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.4%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:32<03:21,  1.79s/it, est. speed input: 89.49 toks/s, output: 1025.61 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:32<02:12,  1.20s/it, est. speed input: 90.72 toks/s, output: 1161.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:33<01:47,  1.02it/s, est. speed input: 91.81 toks/s, output: 1234.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:33<01:31,  1.20it/s, est. speed input: 92.44 toks/s, output: 1302.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:34<01:19,  1.35it/s, est. speed input: 92.62 toks/s, output: 1419.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:35<00:58,  1.79it/s, est. speed input: 94.80 toks/s, output: 1563.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:36<00:59,  1.73it/s, est. speed input: 94.80 toks/s, output: 1618.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:36<00:46,  2.19it/s, est. speed input: 96.89 toks/s, output: 1760.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:36<00:44,  2.25it/s, est. speed input: 97.52 toks/s, output: 1823.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:37<00:45,  2.21it/s, est. speed input: 97.91 toks/s, output: 1882.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:40<01:41,  1.02s/it, est. speed input: 92.70 toks/s, output: 1836.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:43:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8029.8 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:40<01:24,  1.17it/s, est. speed input: 93.29 toks/s, output: 1900.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:40<01:04,  1.50it/s, est. speed input: 94.44 toks/s, output: 1975.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:41<00:57,  1.67it/s, est. speed input: 94.88 toks/s, output: 2036.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:42<01:10,  1.34it/s, est. speed input: 93.81 toks/s, output: 2065.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:43<01:10,  1.33it/s, est. speed input: 93.53 toks/s, output: 2110.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:43<00:58,  1.59it/s, est. speed input: 94.21 toks/s, output: 2175.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:43<00:46,  1.96it/s, est. speed input: 95.14 toks/s, output: 2246.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:44<00:51,  1.78it/s, est. speed input: 95.05 toks/s, output: 2293.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:45<00:55,  1.61it/s, est. speed input: 94.81 toks/s, output: 2336.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:45<00:54,  1.63it/s, est. speed input: 94.84 toks/s, output: 2388.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:49<02:13,  1.51s/it, est. speed input: 89.11 toks/s, output: 2294.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:49<01:50,  1.27s/it, est. speed input: 89.05 toks/s, output: 2342.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:50<01:19,  1.08it/s, est. speed input: 90.10 toks/s, output: 2419.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7319.3 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:50<01:05,  1.31it/s, est. speed input: 90.60 toks/s, output: 2482.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:51<01:16,  1.10it/s, est. speed input: 89.59 toks/s, output: 2502.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:52<00:55,  1.48it/s, est. speed input: 90.56 toks/s, output: 2625.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:53<01:09,  1.17it/s, est. speed input: 89.29 toks/s, output: 2638.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:54<00:53,  1.47it/s, est. speed input: 90.15 toks/s, output: 2758.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:55<00:37,  2.08it/s, est. speed input: 91.90 toks/s, output: 2905.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:55<00:31,  2.40it/s, est. speed input: 92.71 toks/s, output: 2977.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:55<00:23,  3.09it/s, est. speed input: 94.30 toks/s, output: 3120.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:55<00:22,  3.21it/s, est. speed input: 94.93 toks/s, output: 3186.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:57<00:45,  1.60it/s, est. speed input: 93.22 toks/s, output: 3175.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:59<01:10,  1.00it/s, est. speed input: 90.93 toks/s, output: 3143.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:00<00:58,  1.20it/s, est. speed input: 91.38 toks/s, output: 3204.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6026.0 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:00<00:47,  1.47it/s, est. speed input: 92.00 toks/s, output: 3271.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:01<01:04,  1.06it/s, est. speed input: 90.57 toks/s, output: 3266.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:03<00:52,  1.26it/s, est. speed input: 90.81 toks/s, output: 3365.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:03<00:33,  1.92it/s, est. speed input: 92.48 toks/s, output: 3517.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:04<00:41,  1.53it/s, est. speed input: 91.82 toks/s, output: 3537.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:05<00:38,  1.57it/s, est. speed input: 91.99 toks/s, output: 3633.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:05<00:31,  1.92it/s, est. speed input: 92.77 toks/s, output: 3708.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:06<00:26,  2.20it/s, est. speed input: 93.38 toks/s, output: 3777.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:06<00:18,  3.04it/s, est. speed input: 94.81 toks/s, output: 3923.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:07<00:28,  2.00it/s, est. speed input: 94.18 toks/s, output: 3941.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:07<00:24,  2.22it/s, est. speed input: 94.67 toks/s, output: 4005.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:08<00:28,  1.89it/s, est. speed input: 94.51 toks/s, output: 4042.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:09<00:41,  1.29it/s, est. speed input: 93.38 toks/s, output: 4040.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:10<00:34,  1.52it/s, est. speed input: 93.77 toks/s, output: 4102.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5270.3 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:11<00:43,  1.18it/s, est. speed input: 92.90 toks/s, output: 4108.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:12<00:36,  1.36it/s, est. speed input: 93.13 toks/s, output: 4164.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:12<00:37,  1.31it/s, est. speed input: 92.87 toks/s, output: 4199.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:13<00:32,  1.46it/s, est. speed input: 93.09 toks/s, output: 4253.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:13<00:21,  2.16it/s, est. speed input: 94.22 toks/s, output: 4395.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:14<00:21,  2.05it/s, est. speed input: 94.31 toks/s, output: 4445.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:14<00:17,  2.46it/s, est. speed input: 95.22 toks/s, output: 4575.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:15<00:23,  1.78it/s, est. speed input: 94.67 toks/s, output: 4593.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:16<00:24,  1.67it/s, est. speed input: 94.59 toks/s, output: 4635.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:16<00:19,  2.03it/s, est. speed input: 95.16 toks/s, output: 4707.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:17<00:18,  2.09it/s, est. speed input: 95.39 toks/s, output: 4764.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:17<00:15,  2.43it/s, est. speed input: 95.87 toks/s, output: 4833.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:18<00:18,  2.05it/s, est. speed input: 95.78 toks/s, output: 4875.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:18<00:19,  1.84it/s, est. speed input: 95.73 toks/s, output: 4918.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:19<00:14,  2.41it/s, est. speed input: 96.61 toks/s, output: 5054.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4560.5 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:20<00:21,  1.52it/s, est. speed input: 95.67 toks/s, output: 5051.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:22<00:25,  1.25it/s, est. speed input: 95.00 toks/s, output: 5062.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:22<00:20,  1.52it/s, est. speed input: 95.42 toks/s, output: 5130.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:22<00:15,  1.90it/s, est. speed input: 95.95 toks/s, output: 5205.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:23<00:15,  1.87it/s, est. speed input: 96.01 toks/s, output: 5256.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:23<00:15,  1.80it/s, est. speed input: 96.07 toks/s, output: 5304.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:23<00:12,  2.08it/s, est. speed input: 96.46 toks/s, output: 5372.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:24<00:10,  2.51it/s, est. speed input: 96.98 toks/s, output: 5446.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:25<00:08,  2.66it/s, est. speed input: 97.90 toks/s, output: 5635.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:26<00:10,  2.16it/s, est. speed input: 97.70 toks/s, output: 5671.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:27<00:16,  1.30it/s, est. speed input: 96.39 toks/s, output: 5644.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:28<00:15,  1.27it/s, est. speed input: 96.14 toks/s, output: 5680.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:29<00:13,  1.40it/s, est. speed input: 96.28 toks/s, output: 5737.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:29<00:12,  1.49it/s, est. speed input: 96.35 toks/s, output: 5791.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:30<00:07,  2.12it/s, est. speed input: 97.22 toks/s, output: 5943.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3335.7 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:30<00:06,  2.31it/s, est. speed input: 97.56 toks/s, output: 6014.10 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:30<00:00,  2.31it/s, est. speed input: 107.66 toks/s, output: 7371.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:30<00:00,  1.77it/s, est. speed input: 107.66 toks/s, output: 7371.79 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:55 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:55 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:44:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:44:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:44:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:44:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:44:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:44:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:44:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 22:44:56 [abstract.py:306] It took 0.647243 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:44:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4169.8936
+  • Training FLOPS: 52.33 TFLOPS (13.08 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 445.19s
+  • policy_training: 289.58s (65.0%)
+  • generation: 97.24s (21.8%)
+  • teacher_logprob_inference: 53.74s (12.1%)
+  • policy_training/sharding_data: 4.35s (1.0%)
+  • prepare_for_generation: 3.23s (0.7%)
+  • policy_training/submit_training_futures: 2.11s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.46s (0.3%)
+  • training_prep: 0.74s (0.2%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 17: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:50:41 [abstract.py:324] It took 0.049617 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05552754e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05552754e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05552754e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05552754e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58845482e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65837708e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58845482e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65837708e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58845482e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65837708e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58845482e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65837708e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05545625e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60031891e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67224731e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.28265743e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.38956785e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #17] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42060501e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05545625e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60031891e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67224731e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71808548e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.23042040e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #17] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.22018356e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05545625e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60031891e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67224731e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28834991e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.55976009e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #17] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.12304878e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05545625e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60031891e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67224731e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30051384e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74983253e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #17] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32737045e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 17: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:50:44 [abstract.py:324] It took 0.046624 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12661.81it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:50:44 [loggers.py:236] Engine 000: Avg prompt throughput: 27.3 tokens/s, Avg generation throughput: 1.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:01,  2.58it/s, est. speed input: 154.68 toks/s, output: 28.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:07, 20.21it/s, est. speed input: 989.17 toks/s, output: 257.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 27.93it/s, est. speed input: 1320.70 toks/s, output: 447.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 30.21it/s, est. speed input: 1466.85 toks/s, output: 638.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:05, 26.91it/s, est. speed input: 1440.99 toks/s, output: 752.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:10, 13.39it/s, est. speed input: 1017.56 toks/s, output: 704.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:10, 12.32it/s, est. speed input: 948.79 toks/s, output: 822.13 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:05<00:48,  2.68it/s, est. speed input: 373.28 toks/s, output: 462.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:07<01:06,  1.91it/s, est. speed input: 279.84 toks/s, output: 492.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:50:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12384.7 tokens/s, Running: 127 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:11<02:14,  1.07s/it, est. speed input: 175.51 toks/s, output: 391.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:16<03:23,  1.63s/it, est. speed input: 130.33 toks/s, output: 372.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:16<02:50,  1.37s/it, est. speed input: 132.06 toks/s, output: 456.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:17<02:39,  1.29s/it, est. speed input: 127.92 toks/s, output: 519.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:17<02:05,  1.03s/it, est. speed input: 130.37 toks/s, output: 605.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:18<01:47,  1.12it/s, est. speed input: 130.40 toks/s, output: 678.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:18<01:38,  1.22it/s, est. speed input: 129.26 toks/s, output: 745.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:19<01:20,  1.48it/s, est. speed input: 130.40 toks/s, output: 823.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:51:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10469.1 tokens/s, Running: 119 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:22<02:43,  1.39s/it, est. speed input: 114.59 toks/s, output: 792.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:23<02:21,  1.21s/it, est. speed input: 113.28 toks/s, output: 853.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:25<02:58,  1.54s/it, est. speed input: 105.25 toks/s, output: 861.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:26<02:38,  1.38s/it, est. speed input: 103.67 toks/s, output: 916.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:26<02:06,  1.11s/it, est. speed input: 104.12 toks/s, output: 986.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:27<02:00,  1.07s/it, est. speed input: 102.66 toks/s, output: 1038.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:28<01:28,  1.26it/s, est. speed input: 104.29 toks/s, output: 1119.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:51:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9334.3 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.4%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:30<02:18,  1.25s/it, est. speed input: 98.25 toks/s, output: 1119.71 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:31<02:04,  1.13s/it, est. speed input: 97.47 toks/s, output: 1174.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:31<01:32,  1.18it/s, est. speed input: 98.98 toks/s, output: 1252.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:32<01:34,  1.14it/s, est. speed input: 97.95 toks/s, output: 1300.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:32<01:12,  1.47it/s, est. speed input: 99.15 toks/s, output: 1376.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:33<01:22,  1.28it/s, est. speed input: 97.98 toks/s, output: 1420.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:35<01:48,  1.03s/it, est. speed input: 95.20 toks/s, output: 1439.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:35<01:32,  1.12it/s, est. speed input: 95.39 toks/s, output: 1501.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:36<01:20,  1.29it/s, est. speed input: 95.69 toks/s, output: 1564.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:39<02:28,  1.45s/it, est. speed input: 89.84 toks/s, output: 1527.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:51:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8058.3 tokens/s, Running: 102 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.8%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:41<02:39,  1.58s/it, est. speed input: 87.19 toks/s, output: 1540.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:41<02:03,  1.23s/it, est. speed input: 87.78 toks/s, output: 1608.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:42<01:45,  1.06s/it, est. speed input: 87.84 toks/s, output: 1666.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:42<01:33,  1.05it/s, est. speed input: 87.78 toks/s, output: 1721.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:43<01:31,  1.07it/s, est. speed input: 87.30 toks/s, output: 1769.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:44<01:16,  1.26it/s, est. speed input: 87.74 toks/s, output: 1833.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:45<00:55,  1.68it/s, est. speed input: 89.03 toks/s, output: 1969.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:45<00:42,  2.15it/s, est. speed input: 90.62 toks/s, output: 2110.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:46<00:38,  2.36it/s, est. speed input: 91.87 toks/s, output: 2243.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:46<00:34,  2.59it/s, est. speed input: 92.69 toks/s, output: 2314.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:47<00:37,  2.32it/s, est. speed input: 92.81 toks/s, output: 2368.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:47<00:33,  2.63it/s, est. speed input: 93.74 toks/s, output: 2439.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:48<00:55,  1.55it/s, est. speed input: 92.28 toks/s, output: 2452.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:49<00:52,  1.62it/s, est. speed input: 92.50 toks/s, output: 2507.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:49<00:41,  2.02it/s, est. speed input: 93.43 toks/s, output: 2581.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:50<00:35,  2.33it/s, est. speed input: 94.62 toks/s, output: 2708.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:51:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7116.5 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.6%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:50<00:32,  2.51it/s, est. speed input: 95.27 toks/s, output: 2774.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:50<00:20,  3.87it/s, est. speed input: 97.46 toks/s, output: 2932.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:51<00:30,  2.56it/s, est. speed input: 97.07 toks/s, output: 2966.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:51<00:30,  2.55it/s, est. speed input: 97.52 toks/s, output: 3025.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:52<00:21,  3.54it/s, est. speed input: 99.37 toks/s, output: 3174.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:52<00:17,  4.09it/s, est. speed input: 101.69 toks/s, output: 3382.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:53<00:21,  3.18it/s, est. speed input: 102.18 toks/s, output: 3486.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:54<00:31,  2.21it/s, est. speed input: 101.23 toks/s, output: 3499.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:55<00:30,  2.26it/s, est. speed input: 101.59 toks/s, output: 3555.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:55<00:25,  2.65it/s, est. speed input: 102.43 toks/s, output: 3628.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:55<00:20,  3.22it/s, est. speed input: 103.35 toks/s, output: 3703.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:55<00:23,  2.78it/s, est. speed input: 103.49 toks/s, output: 3752.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:56<00:31,  2.00it/s, est. speed input: 102.92 toks/s, output: 3776.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:57<00:29,  2.16it/s, est. speed input: 103.40 toks/s, output: 3835.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:57<00:27,  2.22it/s, est. speed input: 103.71 toks/s, output: 3889.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:57<00:19,  3.02it/s, est. speed input: 105.15 toks/s, output: 4029.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:59<00:27,  2.09it/s, est. speed input: 104.67 toks/s, output: 4097.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:51:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6050.8 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.4%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:00<00:38,  1.49it/s, est. speed input: 103.31 toks/s, output: 4089.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:00<00:30,  1.85it/s, est. speed input: 104.06 toks/s, output: 4164.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:01<00:36,  1.51it/s, est. speed input: 103.32 toks/s, output: 4179.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:02<00:39,  1.38it/s, est. speed input: 102.82 toks/s, output: 4203.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:02<00:29,  1.79it/s, est. speed input: 103.58 toks/s, output: 4279.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:03<00:33,  1.54it/s, est. speed input: 103.08 toks/s, output: 4304.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:04<00:24,  2.05it/s, est. speed input: 104.01 toks/s, output: 4433.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:04<00:23,  2.11it/s, est. speed input: 104.26 toks/s, output: 4488.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:04<00:18,  2.55it/s, est. speed input: 104.94 toks/s, output: 4562.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:04<00:14,  3.18it/s, est. speed input: 105.77 toks/s, output: 4640.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:05<00:21,  2.10it/s, est. speed input: 105.24 toks/s, output: 4662.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:06<00:22,  2.01it/s, est. speed input: 105.27 toks/s, output: 4708.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:06<00:17,  2.55it/s, est. speed input: 106.00 toks/s, output: 4784.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:07<00:20,  2.14it/s, est. speed input: 105.84 toks/s, output: 4823.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:08<00:29,  1.41it/s, est. speed input: 104.74 toks/s, output: 4817.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:08<00:21,  1.87it/s, est. speed input: 105.44 toks/s, output: 4894.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:08<00:17,  2.23it/s, est. speed input: 105.95 toks/s, output: 4962.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:09<00:23,  1.67it/s, est. speed input: 105.39 toks/s, output: 4979.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:51:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4806.4 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:10<00:25,  1.47it/s, est. speed input: 104.91 toks/s, output: 5004.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:11<00:27,  1.33it/s, est. speed input: 104.40 toks/s, output: 5026.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:12<00:27,  1.29it/s, est. speed input: 104.02 toks/s, output: 5054.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:13<00:31,  1.11it/s, est. speed input: 103.15 toks/s, output: 5059.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:13<00:22,  1.51it/s, est. speed input: 103.83 toks/s, output: 5139.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:14<00:20,  1.59it/s, est. speed input: 103.88 toks/s, output: 5188.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:14<00:16,  1.89it/s, est. speed input: 104.28 toks/s, output: 5255.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:14<00:15,  2.06it/s, est. speed input: 104.55 toks/s, output: 5315.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:16<00:20,  1.44it/s, est. speed input: 103.72 toks/s, output: 5321.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:16<00:19,  1.46it/s, est. speed input: 103.62 toks/s, output: 5364.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:18<00:17,  1.54it/s, est. speed input: 103.53 toks/s, output: 5456.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:18<00:11,  2.17it/s, est. speed input: 104.59 toks/s, output: 5611.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:18<00:10,  2.18it/s, est. speed input: 104.75 toks/s, output: 5668.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:52:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3844.1 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:20<00:19,  1.17it/s, est. speed input: 102.73 toks/s, output: 5611.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:24<00:25,  1.20s/it, est. speed input: 100.01 toks/s, output: 5564.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:26<00:29,  1.46s/it, est. speed input: 98.00 toks/s, output: 5509.12 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  1.46s/it, est. speed input: 111.76 toks/s, output: 7391.41 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  1.84it/s, est. speed input: 111.76 toks/s, output: 7391.41 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:52:17 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:52:17 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:52:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:52:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:52:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:52:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:52:17 [abstract.py:306] It took 0.645637 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:52:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:52:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:52:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:52:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0128
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4007.9062
+  • Training FLOPS: 53.73 TFLOPS (13.43 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 419.80s
+  • policy_training: 270.20s (64.4%)
+  • generation: 93.31s (22.2%)
+  • teacher_logprob_inference: 51.41s (12.2%)
+  • policy_training/sharding_data: 4.34s (1.0%)
+  • prepare_for_generation: 3.31s (0.8%)
+  • policy_training/submit_training_futures: 2.03s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.4%)
+  • training_prep: 0.96s (0.2%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 18: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 22:57:41 [abstract.py:324] It took 0.055906 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05612852e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57686234e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66802979e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05612852e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57686234e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66802979e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05612852e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57686234e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66802979e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05612852e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57686234e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66802979e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05609062e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58386917e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67806549e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26464386e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.30480480e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #18] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42305176e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05609062e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58386917e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67806549e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70474701e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.24216404e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #18] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.20144806e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05609062e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58386917e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67806549e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28701553e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.22290087e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #18] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11082001e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05609062e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58386917e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67806549e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29792175e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74233208e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #18] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32822830e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 18: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 22:57:44 [abstract.py:324] It took 0.046980 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13318.42it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-02 22:57:44 [loggers.py:236] Engine 000: Avg prompt throughput: 28.5 tokens/s, Avg generation throughput: 60.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:28,  5.53it/s, est. speed input: 321.01 toks/s, output: 71.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:03, 41.69it/s, est. speed input: 2149.19 toks/s, output: 624.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:03, 35.96it/s, est. speed input: 2042.04 toks/s, output: 776.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:05, 25.44it/s, est. speed input: 1660.22 toks/s, output: 827.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:05, 23.37it/s, est. speed input: 1561.17 toks/s, output: 998.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:03<00:26,  5.05it/s, est. speed input: 563.91 toks/s, output: 485.57 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:07<01:13,  1.76it/s, est. speed input: 248.15 toks/s, output: 345.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:08<01:12,  1.76it/s, est. speed input: 229.10 toks/s, output: 484.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:57:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12459.6 tokens/s, Running: 128 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:11<01:58,  1.07it/s, est. speed input: 169.39 toks/s, output: 440.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:12<01:45,  1.20it/s, est. speed input: 170.08 toks/s, output: 522.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:13<01:55,  1.09it/s, est. speed input: 158.04 toks/s, output: 563.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:14<01:51,  1.12it/s, est. speed input: 153.34 toks/s, output: 623.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:14<01:43,  1.19it/s, est. speed input: 150.48 toks/s, output: 687.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:16<02:01,  1.00it/s, est. speed input: 140.66 toks/s, output: 716.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:16<01:33,  1.29it/s, est. speed input: 143.22 toks/s, output: 801.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:18<01:56,  1.03it/s, est. speed input: 134.63 toks/s, output: 824.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:18<01:36,  1.23it/s, est. speed input: 135.09 toks/s, output: 897.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:58:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10512.1 tokens/s, Running: 119 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:20<02:08,  1.09s/it, est. speed input: 126.25 toks/s, output: 907.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:21<01:50,  1.05it/s, est. speed input: 122.86 toks/s, output: 1015.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:22<01:35,  1.21it/s, est. speed input: 123.00 toks/s, output: 1083.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:23<01:59,  1.05s/it, est. speed input: 116.78 toks/s, output: 1095.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:28<03:55,  2.09s/it, est. speed input: 99.08 toks/s, output: 995.84 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-02 22:58:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9292.6 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:30<03:27,  1.85s/it, est. speed input: 97.14 toks/s, output: 1041.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:30<02:31,  1.37s/it, est. speed input: 98.70 toks/s, output: 1123.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:30<01:50,  1.01s/it, est. speed input: 100.27 toks/s, output: 1205.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:30<01:21,  1.34it/s, est. speed input: 101.88 toks/s, output: 1287.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:31<01:28,  1.22it/s, est. speed input: 100.61 toks/s, output: 1332.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:32<01:27,  1.22it/s, est. speed input: 99.84 toks/s, output: 1384.42 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:32<01:06,  1.59it/s, est. speed input: 101.22 toks/s, output: 1462.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:33<00:52,  1.99it/s, est. speed input: 102.84 toks/s, output: 1601.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:33<00:46,  2.21it/s, est. speed input: 103.72 toks/s, output: 1672.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:33<00:42,  2.41it/s, est. speed input: 104.55 toks/s, output: 1742.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:34<00:43,  2.33it/s, est. speed input: 104.97 toks/s, output: 1804.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:34<00:39,  2.51it/s, est. speed input: 105.68 toks/s, output: 1873.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:36<01:19,  1.25it/s, est. speed input: 102.08 toks/s, output: 1864.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:36<00:47,  2.06it/s, est. speed input: 104.93 toks/s, output: 2025.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:37<00:56,  1.69it/s, est. speed input: 104.00 toks/s, output: 2060.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:37<00:46,  2.05it/s, est. speed input: 105.13 toks/s, output: 2135.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:37<00:38,  2.45it/s, est. speed input: 106.15 toks/s, output: 2209.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:38<00:32,  2.80it/s, est. speed input: 107.75 toks/s, output: 2345.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:38<00:21,  4.11it/s, est. speed input: 110.66 toks/s, output: 2508.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:39<00:33,  2.62it/s, est. speed input: 109.66 toks/s, output: 2536.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:58:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7871.7 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:40<00:44,  1.99it/s, est. speed input: 108.68 toks/s, output: 2564.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:40<00:35,  2.47it/s, est. speed input: 109.89 toks/s, output: 2641.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:40<00:33,  2.57it/s, est. speed input: 110.51 toks/s, output: 2704.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:41<00:34,  2.49it/s, est. speed input: 110.85 toks/s, output: 2760.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:41<00:28,  2.96it/s, est. speed input: 111.88 toks/s, output: 2833.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:41<00:29,  2.79it/s, est. speed input: 112.27 toks/s, output: 2890.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:42<00:31,  2.64it/s, est. speed input: 112.61 toks/s, output: 2945.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:42<00:27,  2.91it/s, est. speed input: 113.37 toks/s, output: 3012.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:42<00:29,  2.74it/s, est. speed input: 113.74 toks/s, output: 3067.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:43<00:21,  3.58it/s, est. speed input: 115.60 toks/s, output: 3211.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:43<00:19,  4.04it/s, est. speed input: 116.58 toks/s, output: 3284.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:43<00:22,  3.39it/s, est. speed input: 116.85 toks/s, output: 3337.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:44<00:20,  3.57it/s, est. speed input: 117.62 toks/s, output: 3403.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:44<00:25,  2.88it/s, est. speed input: 117.64 toks/s, output: 3449.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:45<00:28,  2.54it/s, est. speed input: 117.62 toks/s, output: 3494.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:45<00:22,  3.17it/s, est. speed input: 118.67 toks/s, output: 3570.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:47<00:52,  1.35it/s, est. speed input: 115.52 toks/s, output: 3520.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:47<00:38,  1.80it/s, est. speed input: 116.49 toks/s, output: 3597.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:47<00:43,  1.59it/s, est. speed input: 115.82 toks/s, output: 3622.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:48<00:34,  1.98it/s, est. speed input: 116.53 toks/s, output: 3691.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:48<00:26,  2.55it/s, est. speed input: 117.45 toks/s, output: 3766.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:48<00:22,  2.92it/s, est. speed input: 118.16 toks/s, output: 3834.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:48<00:15,  4.06it/s, est. speed input: 119.93 toks/s, output: 3983.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:49<00:18,  3.42it/s, est. speed input: 120.14 toks/s, output: 4033.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:58:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6647.3 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:50<00:27,  2.23it/s, est. speed input: 119.25 toks/s, output: 4047.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:50<00:21,  2.76it/s, est. speed input: 120.47 toks/s, output: 4178.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:50<00:17,  3.30it/s, est. speed input: 121.42 toks/s, output: 4254.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:51<00:18,  3.03it/s, est. speed input: 122.07 toks/s, output: 4363.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:51<00:16,  3.33it/s, est. speed input: 122.82 toks/s, output: 4432.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:51<00:12,  4.39it/s, est. speed input: 124.57 toks/s, output: 4583.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:52<00:12,  4.21it/s, est. speed input: 125.11 toks/s, output: 4644.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:52<00:08,  5.84it/s, est. speed input: 127.12 toks/s, output: 4804.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:53<00:18,  2.75it/s, est. speed input: 125.67 toks/s, output: 4794.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:55<00:38,  1.28it/s, est. speed input: 121.87 toks/s, output: 4695.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:55<00:33,  1.42it/s, est. speed input: 121.96 toks/s, output: 4743.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:56<00:28,  1.62it/s, est. speed input: 122.23 toks/s, output: 4799.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:57<00:19,  2.27it/s, est. speed input: 123.64 toks/s, output: 4987.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:57<00:16,  2.65it/s, est. speed input: 124.39 toks/s, output: 5063.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:59<00:29,  1.40it/s, est. speed input: 121.47 toks/s, output: 4991.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:59<00:25,  1.62it/s, est. speed input: 121.82 toks/s, output: 5054.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:59<00:19,  2.06it/s, est. speed input: 122.67 toks/s, output: 5134.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:59<00:16,  2.40it/s, est. speed input: 123.18 toks/s, output: 5203.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:58:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5161.5 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:00<00:15,  2.40it/s, est. speed input: 123.36 toks/s, output: 5256.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:00<00:12,  2.87it/s, est. speed input: 124.02 toks/s, output: 5330.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:02<00:28,  1.26it/s, est. speed input: 121.23 toks/s, output: 5259.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:02<00:23,  1.51it/s, est. speed input: 121.52 toks/s, output: 5320.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:03<00:26,  1.31it/s, est. speed input: 120.50 toks/s, output: 5326.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:04<00:29,  1.11it/s, est. speed input: 119.17 toks/s, output: 5316.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:05<00:19,  1.61it/s, est. speed input: 119.92 toks/s, output: 5452.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:06<00:16,  1.81it/s, est. speed input: 120.09 toks/s, output: 5560.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:06<00:13,  2.02it/s, est. speed input: 120.50 toks/s, output: 5628.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:07<00:15,  1.76it/s, est. speed input: 119.98 toks/s, output: 5654.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:08<00:15,  1.73it/s, est. speed input: 119.76 toks/s, output: 5697.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:08<00:12,  2.05it/s, est. speed input: 120.19 toks/s, output: 5770.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:08<00:09,  2.41it/s, est. speed input: 120.66 toks/s, output: 5845.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:08<00:07,  2.93it/s, est. speed input: 121.26 toks/s, output: 5925.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:09<00:08,  2.64it/s, est. speed input: 121.33 toks/s, output: 5979.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:09<00:06,  3.25it/s, est. speed input: 121.96 toks/s, output: 6061.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:09<00:04,  3.93it/s, est. speed input: 123.02 toks/s, output: 6216.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 22:58:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3924.5 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:10<00:06,  2.95it/s, est. speed input: 122.81 toks/s, output: 6258.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:10<00:03,  4.35it/s, est. speed input: 124.26 toks/s, output: 6434.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:11<00:04,  3.18it/s, est. speed input: 124.02 toks/s, output: 6476.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:12<00:06,  2.02it/s, est. speed input: 123.08 toks/s, output: 6478.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:13<00:09,  1.40it/s, est. speed input: 121.62 toks/s, output: 6458.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:15<00:12,  1.08s/it, est. speed input: 119.06 toks/s, output: 6381.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:15<00:09,  1.18it/s, est. speed input: 119.45 toks/s, output: 6460.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:17<00:09,  1.02it/s, est. speed input: 118.16 toks/s, output: 6452.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:17<00:06,  1.36it/s, est. speed input: 118.78 toks/s, output: 6543.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:17<00:04,  1.77it/s, est. speed input: 119.35 toks/s, output: 6633.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:17<00:03,  2.31it/s, est. speed input: 119.94 toks/s, output: 6726.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:18<00:02,  2.06it/s, est. speed input: 119.83 toks/s, output: 6838.09 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:18<00:00,  2.06it/s, est. speed input: 123.57 toks/s, output: 7359.66 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:18<00:00,  2.04it/s, est. speed input: 123.57 toks/s, output: 7359.66 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 22:59:08 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 22:59:08 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 22:59:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:59:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:59:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:59:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 22:59:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 22:59:09 [abstract.py:306] It took 0.647790 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 22:59:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 22:59:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 22:59:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0135
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3612.2000
+  • Training FLOPS: 49.39 TFLOPS (12.35 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 399.38s
+  • policy_training: 261.86s (65.6%)
+  • generation: 84.44s (21.1%)
+  • teacher_logprob_inference: 48.28s (12.1%)
+  • policy_training/sharding_data: 4.39s (1.1%)
+  • prepare_for_generation: 3.30s (0.8%)
+  • policy_training/submit_training_futures: 2.03s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.48s (0.4%)
+  • training_prep: 0.83s (0.2%)
+  • teacher_logprob_inference_prep: 0.47s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 19: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:04:21 [abstract.py:324] It took 0.050246 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05646328e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57564774e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67247009e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05646328e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57564774e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67247009e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05646328e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57564774e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67247009e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05646328e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57564774e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67247009e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05642148e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57559586e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66082764e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26447754e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.28259230e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #19] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42590378e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05642148e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57559586e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66082764e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69508362e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.25550575e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #19] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.18755989e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05642148e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57559586e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66082764e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28003845e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.27090883e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #19] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14046936e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05642148e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57559586e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66082764e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29035530e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.73914185e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #19] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32649689e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 19: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 23:04:24 [abstract.py:324] It took 0.044698 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12746.71it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-02 23:04:24 [loggers.py:236] Engine 000: Avg prompt throughput: 29.3 tokens/s, Avg generation throughput: 57.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:17,  2.06it/s, est. speed input: 121.53 toks/s, output: 28.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:15,  9.78it/s, est. speed input: 474.17 toks/s, output: 148.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:08, 18.24it/s, est. speed input: 808.78 toks/s, output: 346.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:08, 17.38it/s, est. speed input: 847.06 toks/s, output: 437.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:06, 22.81it/s, est. speed input: 1024.76 toks/s, output: 673.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 24.23it/s, est. speed input: 1089.46 toks/s, output: 807.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:05, 25.83it/s, est. speed input: 1171.35 toks/s, output: 1024.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:06, 19.79it/s, est. speed input: 1099.28 toks/s, output: 1062.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:06<01:03,  2.01it/s, est. speed input: 285.92 toks/s, output: 467.98 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:10<01:34,  1.34it/s, est. speed input: 201.17 toks/s, output: 468.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:04:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12379.0 tokens/s, Running: 126 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:16<02:39,  1.29s/it, est. speed input: 131.94 toks/s, output: 447.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:17<02:35,  1.26s/it, est. speed input: 127.14 toks/s, output: 508.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:04:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10529.0 tokens/s, Running: 122 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:22<03:07,  1.55s/it, est. speed input: 106.46 toks/s, output: 562.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:22<02:39,  1.33s/it, est. speed input: 108.19 toks/s, output: 644.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:23<02:26,  1.23s/it, est. speed input: 107.06 toks/s, output: 709.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:24<02:19,  1.19s/it, est. speed input: 105.02 toks/s, output: 766.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:27<02:29,  1.29s/it, est. speed input: 98.45 toks/s, output: 848.35 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:27<02:08,  1.11s/it, est. speed input: 98.85 toks/s, output: 919.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:28<01:34,  1.20it/s, est. speed input: 100.52 toks/s, output: 1064.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:29<01:35,  1.17it/s, est. speed input: 99.43 toks/s, output: 1116.62 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 23:04:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9375.5 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:32<02:34,  1.39s/it, est. speed input: 91.63 toks/s, output: 1092.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:34<02:44,  1.50s/it, est. speed input: 88.63 toks/s, output: 1119.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:35<02:43,  1.50s/it, est. speed input: 86.57 toks/s, output: 1156.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:36<02:08,  1.19s/it, est. speed input: 87.38 toks/s, output: 1228.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:36<01:42,  1.04it/s, est. speed input: 88.12 toks/s, output: 1299.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:39<02:43,  1.54s/it, est. speed input: 82.98 toks/s, output: 1284.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:05:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8160.1 tokens/s, Running: 106 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:40<02:21,  1.35s/it, est. speed input: 82.65 toks/s, output: 1338.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:40<01:46,  1.02s/it, est. speed input: 83.70 toks/s, output: 1413.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:41<00:58,  1.73it/s, est. speed input: 86.76 toks/s, output: 1636.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:41<00:49,  2.03it/s, est. speed input: 87.83 toks/s, output: 1712.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:41<00:45,  2.15it/s, est. speed input: 88.51 toks/s, output: 1779.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:42<00:51,  1.91it/s, est. speed input: 88.51 toks/s, output: 1832.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:43<00:48,  2.02it/s, est. speed input: 89.04 toks/s, output: 1897.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:45<01:07,  1.40it/s, est. speed input: 87.81 toks/s, output: 1973.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:50<02:42,  1.73s/it, est. speed input: 80.18 toks/s, output: 1855.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:05:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7256.8 tokens/s, Running: 94 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:50<02:10,  1.40s/it, est. speed input: 80.68 toks/s, output: 1921.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:50<01:37,  1.06s/it, est. speed input: 81.71 toks/s, output: 1998.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:54<02:58,  1.97s/it, est. speed input: 76.36 toks/s, output: 1920.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:55<02:32,  1.70s/it, est. speed input: 76.07 toks/s, output: 1966.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:57<02:29,  1.68s/it, est. speed input: 74.94 toks/s, output: 1990.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:59<02:40,  1.82s/it, est. speed input: 73.25 toks/s, output: 1998.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:05:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6811.1 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:00<01:43,  1.21s/it, est. speed input: 74.15 toks/s, output: 2127.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:01<01:07,  1.25it/s, est. speed input: 75.68 toks/s, output: 2275.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:01<00:59,  1.41it/s, est. speed input: 76.17 toks/s, output: 2340.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:02<01:02,  1.31it/s, est. speed input: 75.99 toks/s, output: 2386.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:02<00:50,  1.59it/s, est. speed input: 76.65 toks/s, output: 2457.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:03<00:46,  1.73it/s, est. speed input: 77.08 toks/s, output: 2520.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:09<02:54,  2.21s/it, est. speed input: 70.73 toks/s, output: 2363.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:09<02:12,  1.69s/it, est. speed input: 71.22 toks/s, output: 2430.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:05:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5853.2 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.9%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:11<01:58,  1.54s/it, est. speed input: 70.89 toks/s, output: 2469.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:12<01:47,  1.42s/it, est. speed input: 70.60 toks/s, output: 2509.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:13<01:43,  1.38s/it, est. speed input: 70.16 toks/s, output: 2544.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:13<01:15,  1.02s/it, est. speed input: 70.82 toks/s, output: 2617.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:14<01:16,  1.04s/it, est. speed input: 70.57 toks/s, output: 2658.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:16<01:24,  1.17s/it, est. speed input: 69.99 toks/s, output: 2685.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:16<00:51,  1.36it/s, est. speed input: 71.17 toks/s, output: 2827.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:17<00:42,  1.58it/s, est. speed input: 71.82 toks/s, output: 2948.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:17<00:36,  1.84it/s, est. speed input: 72.37 toks/s, output: 3018.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:18<00:37,  1.76it/s, est. speed input: 72.54 toks/s, output: 3072.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:19<00:39,  1.64it/s, est. speed input: 72.62 toks/s, output: 3122.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:19<00:35,  1.80it/s, est. speed input: 72.99 toks/s, output: 3185.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:05:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5448.6 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:20<00:47,  1.33it/s, est. speed input: 72.60 toks/s, output: 3214.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:21<00:36,  1.68it/s, est. speed input: 73.35 toks/s, output: 3340.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:22<00:35,  1.68it/s, est. speed input: 73.55 toks/s, output: 3395.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:23<00:38,  1.53it/s, est. speed input: 73.51 toks/s, output: 3440.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:24<00:45,  1.28it/s, est. speed input: 73.24 toks/s, output: 3472.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:25<00:46,  1.22it/s, est. speed input: 73.15 toks/s, output: 3513.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:25<00:36,  1.52it/s, est. speed input: 73.62 toks/s, output: 3581.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:25<00:29,  1.85it/s, est. speed input: 74.12 toks/s, output: 3649.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:26<00:27,  1.98it/s, est. speed input: 74.44 toks/s, output: 3710.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:26<00:23,  2.22it/s, est. speed input: 74.88 toks/s, output: 3775.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:27<00:28,  1.85it/s, est. speed input: 74.93 toks/s, output: 3821.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:28<00:27,  1.84it/s, est. speed input: 75.39 toks/s, output: 3931.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:28<00:21,  2.23it/s, est. speed input: 76.23 toks/s, output: 4062.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:29<00:22,  2.08it/s, est. speed input: 76.38 toks/s, output: 4114.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:05:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4678.6 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:30<00:31,  1.47it/s, est. speed input: 75.96 toks/s, output: 4134.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:31<00:33,  1.33it/s, est. speed input: 75.81 toks/s, output: 4170.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:33<00:49,  1.13s/it, est. speed input: 74.68 toks/s, output: 4154.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:35<00:51,  1.21s/it, est. speed input: 74.21 toks/s, output: 4172.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:36<00:49,  1.17s/it, est. speed input: 74.03 toks/s, output: 4207.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:36<00:35,  1.16it/s, est. speed input: 74.55 toks/s, output: 4283.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:37<00:37,  1.05it/s, est. speed input: 74.30 toks/s, output: 4313.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:38<00:35,  1.11it/s, est. speed input: 74.33 toks/s, output: 4359.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:38<00:26,  1.45it/s, est. speed input: 74.79 toks/s, output: 4432.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:39<00:17,  2.11it/s, est. speed input: 75.65 toks/s, output: 4575.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:39<00:17,  2.02it/s, est. speed input: 75.81 toks/s, output: 4630.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:06:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4193.4 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:40<00:19,  1.70it/s, est. speed input: 75.77 toks/s, output: 4672.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:40<00:15,  2.18it/s, est. speed input: 76.30 toks/s, output: 4749.17 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:40<00:00,  2.18it/s, est. speed input: 96.06 toks/s, output: 7436.16 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:40<00:00,  1.59it/s, est. speed input: 96.06 toks/s, output: 7436.16 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:06:11 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:06:11 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:06:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:06:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:06:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:06:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:06:12 [abstract.py:306] It took 0.638300 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:06:12 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:06:12 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:06:12 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:06:12 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0121
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4675.8438
+  • Training FLOPS: 55.32 TFLOPS (13.83 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 482.49s
+  • policy_training: 311.61s (64.6%)
+  • generation: 108.02s (22.4%)
+  • teacher_logprob_inference: 58.07s (12.0%)
+  • policy_training/sharding_data: 4.36s (0.9%)
+  • prepare_for_generation: 3.31s (0.7%)
+  • policy_training/submit_training_futures: 2.17s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.42s (0.3%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 20: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:12:24 [abstract.py:324] It took 0.057019 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05670391e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56846695e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67193146e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05670391e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56846695e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67193146e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05670391e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56846695e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67193146e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05670391e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56846695e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67193146e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05676641e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57650833e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67255402e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05676641e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57650833e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67255402e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05676641e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57650833e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67255402e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05676641e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57650833e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67255402e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26667347e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.14517784e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #20] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42592194e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69600067e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27340593e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #20] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.16707764e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28145905e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.61443901e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #20] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14733429e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30105057e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75250130e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #20] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32583206e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 20: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 23:12:27 [abstract.py:324] It took 0.066426 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12481.89it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:12:27 [loggers.py:236] Engine 000: Avg prompt throughput: 25.6 tokens/s, Avg generation throughput: 2.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:36,  1.65it/s, est. speed input: 99.22 toks/s, output: 26.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:08, 17.55it/s, est. speed input: 836.30 toks/s, output: 283.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:06, 23.16it/s, est. speed input: 1098.29 toks/s, output: 449.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:10, 13.45it/s, est. speed input: 835.30 toks/s, output: 461.34 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:08<01:21,  1.67it/s, est. speed input: 176.56 toks/s, output: 248.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:09<01:29,  1.51it/s, est. speed input: 155.78 toks/s, output: 359.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:12:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12488.4 tokens/s, Running: 135 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:12<01:50,  1.20it/s, est. speed input: 129.79 toks/s, output: 436.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:14<02:05,  1.06it/s, est. speed input: 118.82 toks/s, output: 470.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:15<02:11,  1.01s/it, est. speed input: 112.79 toks/s, output: 516.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:16<01:55,  1.13it/s, est. speed input: 113.86 toks/s, output: 589.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:18<02:22,  1.10s/it, est. speed input: 105.19 toks/s, output: 612.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:19<02:31,  1.18s/it, est. speed input: 100.49 toks/s, output: 650.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:12:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10803.1 tokens/s, Running: 128 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:21<02:44,  1.29s/it, est. speed input: 95.71 toks/s, output: 684.49 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:22<02:36,  1.24s/it, est. speed input: 93.75 toks/s, output: 734.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:22<02:01,  1.02it/s, est. speed input: 95.37 toks/s, output: 809.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:25<02:53,  1.40s/it, est. speed input: 88.36 toks/s, output: 811.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:25<02:28,  1.21s/it, est. speed input: 88.31 toks/s, output: 872.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:26<01:58,  1.03it/s, est. speed input: 89.30 toks/s, output: 941.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:26<01:29,  1.35it/s, est. speed input: 91.01 toks/s, output: 1018.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:28<02:27,  1.23s/it, est. speed input: 85.64 toks/s, output: 1015.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:12:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9606.8 tokens/s, Running: 120 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.8%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:31<03:17,  1.66s/it, est. speed input: 80.24 toks/s, output: 1010.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:32<02:44,  1.39s/it, est. speed input: 80.28 toks/s, output: 1068.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:32<01:59,  1.02s/it, est. speed input: 81.84 toks/s, output: 1145.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:35<03:00,  1.56s/it, est. speed input: 77.03 toks/s, output: 1134.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:35<02:25,  1.27s/it, est. speed input: 77.44 toks/s, output: 1196.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:38<03:08,  1.65s/it, est. speed input: 73.93 toks/s, output: 1197.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:38<02:33,  1.35s/it, est. speed input: 74.22 toks/s, output: 1257.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:39<02:05,  1.12s/it, est. speed input: 74.73 toks/s, output: 1320.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:13:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8583.7 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.6%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:42<02:10,  1.19s/it, est. speed input: 73.20 toks/s, output: 1395.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:43<02:29,  1.37s/it, est. speed input: 71.36 toks/s, output: 1412.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:44<02:18,  1.28s/it, est. speed input: 71.16 toks/s, output: 1459.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:45<02:04,  1.16s/it, est. speed input: 71.21 toks/s, output: 1512.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:45<01:33,  1.13it/s, est. speed input: 72.30 toks/s, output: 1585.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:46<01:32,  1.14it/s, est. speed input: 72.28 toks/s, output: 1635.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:49<02:18,  1.33s/it, est. speed input: 69.95 toks/s, output: 1633.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:13:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7742.4 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.0%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:50<02:13,  1.30s/it, est. speed input: 69.48 toks/s, output: 1672.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:52<02:29,  1.47s/it, est. speed input: 68.13 toks/s, output: 1690.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:54<02:40,  1.59s/it, est. speed input: 66.91 toks/s, output: 1709.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:54<02:06,  1.26s/it, est. speed input: 67.41 toks/s, output: 1772.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:54<01:32,  1.06it/s, est. speed input: 68.28 toks/s, output: 1843.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:56<01:38,  1.01s/it, est. speed input: 67.96 toks/s, output: 1882.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:56<01:34,  1.03it/s, est. speed input: 67.99 toks/s, output: 1931.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:57<01:28,  1.08it/s, est. speed input: 68.08 toks/s, output: 1981.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:57<01:07,  1.41it/s, est. speed input: 68.92 toks/s, output: 2051.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:58<01:03,  1.48it/s, est. speed input: 69.22 toks/s, output: 2108.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:59<01:01,  1.50it/s, est. speed input: 69.52 toks/s, output: 2162.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:59<01:02,  1.46it/s, est. speed input: 69.70 toks/s, output: 2213.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:13:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6953.3 tokens/s, Running: 91 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.5%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [01:00<00:57,  1.59it/s, est. speed input: 70.12 toks/s, output: 2272.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [01:01<01:13,  1.23it/s, est. speed input: 69.66 toks/s, output: 2303.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [01:01<00:54,  1.65it/s, est. speed input: 70.52 toks/s, output: 2376.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [01:02<00:48,  1.81it/s, est. speed input: 71.01 toks/s, output: 2437.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:02<00:52,  1.65it/s, est. speed input: 71.16 toks/s, output: 2485.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:03<00:43,  1.99it/s, est. speed input: 71.82 toks/s, output: 2552.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:05<01:17,  1.09it/s, est. speed input: 70.69 toks/s, output: 2556.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:05<00:58,  1.45it/s, est. speed input: 71.44 toks/s, output: 2626.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:06<01:18,  1.06it/s, est. speed input: 70.70 toks/s, output: 2642.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:07<01:14,  1.10it/s, est. speed input: 70.77 toks/s, output: 2686.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:07<00:54,  1.49it/s, est. speed input: 71.60 toks/s, output: 2758.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:08<00:53,  1.50it/s, est. speed input: 71.80 toks/s, output: 2807.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:09<01:00,  1.31it/s, est. speed input: 71.68 toks/s, output: 2844.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:13:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6028.9 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.5%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:10<01:06,  1.17it/s, est. speed input: 71.46 toks/s, output: 2877.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:10<00:53,  1.45it/s, est. speed input: 72.00 toks/s, output: 2941.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:11<00:45,  1.67it/s, est. speed input: 72.47 toks/s, output: 3002.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:12<01:01,  1.21it/s, est. speed input: 71.95 toks/s, output: 3022.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:12<00:52,  1.42it/s, est. speed input: 72.36 toks/s, output: 3080.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:13<00:43,  1.67it/s, est. speed input: 73.09 toks/s, output: 3192.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:14<00:33,  2.10it/s, est. speed input: 73.80 toks/s, output: 3264.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:14<00:26,  2.61it/s, est. speed input: 74.47 toks/s, output: 3335.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:14<00:23,  2.89it/s, est. speed input: 75.03 toks/s, output: 3401.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:14<00:23,  2.94it/s, est. speed input: 75.51 toks/s, output: 3462.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:16<00:49,  1.34it/s, est. speed input: 74.56 toks/s, output: 3459.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:16<00:42,  1.56it/s, est. speed input: 75.01 toks/s, output: 3518.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:17<00:37,  1.73it/s, est. speed input: 75.36 toks/s, output: 3575.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:17<00:29,  2.15it/s, est. speed input: 75.93 toks/s, output: 3642.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:18<00:34,  1.82it/s, est. speed input: 75.96 toks/s, output: 3684.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:19<00:45,  1.36it/s, est. speed input: 75.57 toks/s, output: 3706.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:19<00:37,  1.61it/s, est. speed input: 76.00 toks/s, output: 3767.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:19<00:30,  2.00it/s, est. speed input: 76.53 toks/s, output: 3832.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:20<00:25,  2.30it/s, est. speed input: 77.01 toks/s, output: 3896.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:13:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5423.4 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.4%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:20<00:21,  2.72it/s, est. speed input: 77.57 toks/s, output: 3962.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:20<00:21,  2.63it/s, est. speed input: 77.91 toks/s, output: 4019.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:21<00:21,  2.60it/s, est. speed input: 78.30 toks/s, output: 4076.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:21<00:21,  2.57it/s, est. speed input: 78.65 toks/s, output: 4133.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:22<00:26,  2.03it/s, est. speed input: 78.72 toks/s, output: 4173.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:22<00:24,  2.21it/s, est. speed input: 79.11 toks/s, output: 4232.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:24<00:27,  1.88it/s, est. speed input: 79.42 toks/s, output: 4323.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:25<00:25,  1.90it/s, est. speed input: 79.86 toks/s, output: 4424.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:25<00:23,  2.01it/s, est. speed input: 80.44 toks/s, output: 4533.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:26<00:24,  1.89it/s, est. speed input: 80.56 toks/s, output: 4577.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:28<00:41,  1.09it/s, est. speed input: 79.24 toks/s, output: 4539.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:13:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4466.2 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.3%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:30<00:40,  1.06it/s, est. speed input: 78.84 toks/s, output: 4596.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:31<00:40,  1.04it/s, est. speed input: 78.59 toks/s, output: 4623.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:32<00:25,  1.59it/s, est. speed input: 79.73 toks/s, output: 4770.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:33<00:28,  1.35it/s, est. speed input: 79.41 toks/s, output: 4790.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:33<00:15,  2.26it/s, est. speed input: 81.07 toks/s, output: 5008.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:34<00:20,  1.69it/s, est. speed input: 80.65 toks/s, output: 5022.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:35<00:24,  1.39it/s, est. speed input: 80.29 toks/s, output: 5039.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:36<00:21,  1.56it/s, est. speed input: 80.58 toks/s, output: 5098.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:36<00:19,  1.68it/s, est. speed input: 80.82 toks/s, output: 5155.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:38<00:24,  1.25it/s, est. speed input: 80.33 toks/s, output: 5163.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:38<00:23,  1.27it/s, est. speed input: 80.31 toks/s, output: 5204.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:39<00:24,  1.19it/s, est. speed input: 80.13 toks/s, output: 5235.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:40<00:18,  1.53it/s, est. speed input: 80.60 toks/s, output: 5306.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:40<00:16,  1.66it/s, est. speed input: 80.83 toks/s, output: 5362.74 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:40<00:00,  1.66it/s, est. speed input: 97.26 toks/s, output: 7561.46 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:40<00:00,  1.59it/s, est. speed input: 97.26 toks/s, output: 7561.46 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:14:14 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:14:14 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:14:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:14:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:14:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:14:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 23:14:15 [abstract.py:306] It took 0.657383 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:14:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:14:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:14:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:14:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-02 23:20:32 [abstract.py:324] It took 0.049881 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05704883e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56286469e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67828827e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05704883e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56286469e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67828827e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05704883e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56286469e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67828827e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05704883e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56286469e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67828827e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05714043e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58173256e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68027344e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25783272e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.06467342e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #21] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42896576e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05714043e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58173256e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68027344e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70733643e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27622175e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #21] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.16474228e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05714043e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58173256e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68027344e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28008118e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.91404295e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #21] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.18165970e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05714043e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58173256e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68027344e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30566902e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75320168e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #21] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32627090e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+▶ Starting validation at step 20...
+(VllmGenerationWorker pid=10140) INFO 05-02 23:20:35 [abstract.py:324] It took 0.043490 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 8843.50it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:20:36 [loggers.py:236] Engine 000: Avg prompt throughput: 5.3 tokens/s, Avg generation throughput: 96.3 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:15,  2.01it/s, est. speed input: 137.00 toks/s, output: 74.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:12,  2.48it/s, est. speed input: 161.80 toks/s, output: 168.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:01<00:19,  1.48it/s, est. speed input: 109.19 toks/s, output: 240.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:08<01:26,  3.08s/it, est. speed input: 30.59 toks/s, output: 232.19 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 23:20:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5397.2 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.4%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:14<01:50,  4.11s/it, est. speed input: 22.65 toks/s, output: 314.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:15<01:16,  2.96s/it, est. speed input: 25.83 toks/s, output: 476.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:16<00:56,  2.26s/it, est. speed input: 28.76 toks/s, output: 628.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:20:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4403.0 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:20<01:14,  3.09s/it, est. speed input: 25.32 toks/s, output: 654.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:21<00:51,  2.23s/it, est. speed input: 27.81 toks/s, output: 816.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:25<00:59,  2.69s/it, est. speed input: 26.32 toks/s, output: 865.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:25<00:42,  2.01s/it, est. speed input: 28.53 toks/s, output: 1020.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:25<00:28,  1.44s/it, est. speed input: 30.97 toks/s, output: 1186.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:26<00:23,  1.23s/it, est. speed input: 32.59 toks/s, output: 1322.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:28<00:27,  1.53s/it, est. speed input: 32.32 toks/s, output: 1390.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:21:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3454.8 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:31<00:34,  2.05s/it, est. speed input: 30.97 toks/s, output: 1418.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:32<00:28,  1.76s/it, est. speed input: 31.97 toks/s, output: 1540.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:36<00:24,  1.75s/it, est. speed input: 32.57 toks/s, output: 1718.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:37<00:21,  1.68s/it, est. speed input: 33.07 toks/s, output: 1823.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:39<00:20,  1.71s/it, est. speed input: 33.17 toks/s, output: 1910.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:21:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2588.8 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:41<00:12,  1.29s/it, est. speed input: 35.12 toks/s, output: 2180.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:41<00:09,  1.05s/it, est. speed input: 36.53 toks/s, output: 2337.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:42<00:09,  1.19s/it, est. speed input: 36.70 toks/s, output: 2422.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:44<00:09,  1.37s/it, est. speed input: 36.71 toks/s, output: 2496.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:45<00:07,  1.17s/it, est. speed input: 37.61 toks/s, output: 2635.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:46<00:05,  1.07s/it, est. speed input: 38.33 toks/s, output: 2764.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:46<00:03,  1.20it/s, est. speed input: 39.58 toks/s, output: 2927.12 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.20it/s, est. speed input: 45.32 toks/s, output: 3630.85 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.46s/it, est. speed input: 45.32 toks/s, output: 3630.85 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10761.52it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  6.64it/s, est. speed input: 451.79 toks/s, output: 172.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:01<00:21,  1.39it/s, est. speed input: 104.00 toks/s, output: 215.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:01<00:18,  1.58it/s, est. speed input: 110.58 toks/s, output: 346.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:21:26 [loggers.py:236] Engine 000: Avg prompt throughput: 213.2 tokens/s, Avg generation throughput: 2395.5 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.6%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:08<01:20,  2.87s/it, est. speed input: 32.94 toks/s, output: 264.29 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 23:21:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5148.4 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:14<01:50,  4.09s/it, est. speed input: 23.34 toks/s, output: 329.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:16<01:26,  3.33s/it, est. speed input: 25.00 toks/s, output: 470.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:19<01:20,  3.22s/it, est. speed input: 24.28 toks/s, output: 573.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:19<00:54,  2.28s/it, est. speed input: 27.39 toks/s, output: 741.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:22<00:54,  2.37s/it, est. speed input: 27.23 toks/s, output: 828.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:21:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4245.1 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:23<00:43,  1.96s/it, est. speed input: 28.92 toks/s, output: 965.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:23<00:31,  1.49s/it, est. speed input: 31.23 toks/s, output: 1120.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:24<00:26,  1.30s/it, est. speed input: 32.94 toks/s, output: 1253.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:24<00:18,  1.03it/s, est. speed input: 35.37 toks/s, output: 1417.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:28<00:23,  1.41s/it, est. speed input: 35.34 toks/s, output: 1548.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:29<00:21,  1.37s/it, est. speed input: 36.06 toks/s, output: 1656.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:29<00:15,  1.05s/it, est. speed input: 38.04 toks/s, output: 1819.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:30<00:12,  1.10it/s, est. speed input: 39.56 toks/s, output: 1960.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:31<00:12,  1.04it/s, est. speed input: 40.40 toks/s, output: 2064.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:21:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3027.1 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.4%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:35<00:21,  1.80s/it, est. speed input: 37.80 toks/s, output: 2011.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:35<00:15,  1.41s/it, est. speed input: 39.18 toks/s, output: 2159.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:36<00:12,  1.30s/it, est. speed input: 39.83 toks/s, output: 2274.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:38<00:11,  1.31s/it, est. speed input: 40.19 toks/s, output: 2370.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:39<00:10,  1.30s/it, est. speed input: 40.55 toks/s, output: 2469.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:40<00:07,  1.13s/it, est. speed input: 41.44 toks/s, output: 2602.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:22:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1963.3 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.4%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:42<00:09,  1.57s/it, est. speed input: 40.47 toks/s, output: 2621.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:45<00:09,  1.84s/it, est. speed input: 39.76 toks/s, output: 2660.92 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:45<00:00,  1.84s/it, est. speed input: 47.17 toks/s, output: 3566.66 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:45<00:00,  1.41s/it, est. speed input: 47.17 toks/s, output: 3566.66 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10512.86it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  6.27it/s, est. speed input: 407.85 toks/s, output: 175.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:06,  4.50it/s, est. speed input: 312.68 toks/s, output: 256.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:07,  4.04it/s, est. speed input: 284.52 toks/s, output: 348.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:01<00:13,  2.03it/s, est. speed input: 170.43 toks/s, output: 351.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:22:16 [loggers.py:236] Engine 000: Avg prompt throughput: 212.5 tokens/s, Avg generation throughput: 3479.2 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) INFO 05-02 23:22:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4839.5 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.7%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:20<03:16,  7.29s/it, est. speed input: 16.02 toks/s, output: 201.04 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:22<02:17,  5.27s/it, est. speed input: 17.92 toks/s, output: 362.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:22<01:32,  3.70s/it, est. speed input: 20.45 toks/s, output: 527.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:24<01:12,  3.04s/it, est. speed input: 21.75 toks/s, output: 664.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:24<00:52,  2.27s/it, est. speed input: 23.91 toks/s, output: 820.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:22:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4203.8 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:27<00:49,  2.26s/it, est. speed input: 24.44 toks/s, output: 923.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:27<00:34,  1.66s/it, est. speed input: 26.60 toks/s, output: 1082.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:27<00:24,  1.23s/it, est. speed input: 28.78 toks/s, output: 1242.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:27<00:17,  1.10it/s, est. speed input: 31.04 toks/s, output: 1404.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:28<00:14,  1.26it/s, est. speed input: 32.77 toks/s, output: 1549.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:28<00:12,  1.42it/s, est. speed input: 34.51 toks/s, output: 1691.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:30<00:14,  1.12it/s, est. speed input: 35.20 toks/s, output: 1786.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:31<00:13,  1.13it/s, est. speed input: 36.33 toks/s, output: 1908.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:31<00:12,  1.15it/s, est. speed input: 37.43 toks/s, output: 2027.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:33<00:10,  1.12it/s, est. speed input: 39.35 toks/s, output: 2251.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:34<00:07,  1.39it/s, est. speed input: 42.02 toks/s, output: 2530.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:35<00:06,  1.40it/s, est. speed input: 43.10 toks/s, output: 2652.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:22:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2806.7 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:35<00:04,  1.64it/s, est. speed input: 44.60 toks/s, output: 2805.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:36<00:05,  1.39it/s, est. speed input: 45.23 toks/s, output: 2899.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:39<00:08,  1.35s/it, est. speed input: 43.41 toks/s, output: 2852.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:41<00:06,  1.33s/it, est. speed input: 43.65 toks/s, output: 2944.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:41<00:04,  1.21s/it, est. speed input: 44.24 toks/s, output: 3059.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:44<00:04,  1.61s/it, est. speed input: 43.21 toks/s, output: 3064.82 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:44<00:00,  1.61s/it, est. speed input: 47.67 toks/s, output: 3615.87 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:44<00:00,  1.39s/it, est. speed input: 47.67 toks/s, output: 3615.87 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 2640.73it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:22:56 [loggers.py:236] Engine 000: Avg prompt throughput: 200.6 tokens/s, Avg generation throughput: 1142.9 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  6.32it/s, est. speed input: 411.03 toks/s, output: 170.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:08,  3.57it/s, est. speed input: 244.26 toks/s, output: 242.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:07<01:41,  3.51s/it, est. speed input: 24.63 toks/s, output: 202.33 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 23:23:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5499.1 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:14<02:17,  4.91s/it, est. speed input: 17.41 toks/s, output: 285.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:15<01:31,  3.38s/it, est. speed input: 20.88 toks/s, output: 450.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:16<01:03,  2.46s/it, est. speed input: 24.04 toks/s, output: 610.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:17<00:51,  2.05s/it, est. speed input: 25.97 toks/s, output: 744.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:23:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4521.6 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:22<01:11,  2.99s/it, est. speed input: 23.13 toks/s, output: 751.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:24<01:04,  2.80s/it, est. speed input: 23.60 toks/s, output: 849.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:25<00:46,  2.13s/it, est. speed input: 25.57 toks/s, output: 999.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:27<00:43,  2.07s/it, est. speed input: 26.18 toks/s, output: 1099.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:23:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3684.5 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:30<00:46,  2.32s/it, est. speed input: 25.84 toks/s, output: 1163.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:31<00:36,  1.92s/it, est. speed input: 27.16 toks/s, output: 1296.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:32<00:28,  1.58s/it, est. speed input: 28.52 toks/s, output: 1433.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:32<00:22,  1.32s/it, est. speed input: 29.82 toks/s, output: 1571.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:32<00:15,  1.03it/s, est. speed input: 31.68 toks/s, output: 1733.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:34<00:15,  1.06s/it, est. speed input: 32.50 toks/s, output: 1838.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:35<00:15,  1.08s/it, est. speed input: 33.34 toks/s, output: 1950.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:36<00:14,  1.15s/it, est. speed input: 33.95 toks/s, output: 2049.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:36<00:10,  1.17it/s, est. speed input: 35.56 toks/s, output: 2209.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:37<00:07,  1.52it/s, est. speed input: 37.10 toks/s, output: 2367.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:38<00:07,  1.34it/s, est. speed input: 37.99 toks/s, output: 2478.32 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:23:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2582.0 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:44<00:21,  2.37s/it, est. speed input: 34.14 toks/s, output: 2306.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:46<00:18,  2.31s/it, est. speed input: 33.93 toks/s, output: 2373.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:46<00:12,  1.74s/it, est. speed input: 35.05 toks/s, output: 2527.22 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.74s/it, est. speed input: 44.84 toks/s, output: 3753.35 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.46s/it, est. speed input: 44.84 toks/s, output: 3753.35 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10535.97it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  6.36it/s, est. speed input: 406.90 toks/s, output: 171.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:04,  6.94it/s, est. speed input: 444.76 toks/s, output: 277.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:01<00:16,  1.80it/s, est. speed input: 146.70 toks/s, output: 255.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:23:47 [loggers.py:236] Engine 000: Avg prompt throughput: 207.4 tokens/s, Avg generation throughput: 2139.7 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:10<01:54,  4.11s/it, est. speed input: 24.06 toks/s, output: 210.43 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:11<01:17,  2.87s/it, est. speed input: 28.25 toks/s, output: 376.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:23:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5217.5 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:12<00:54,  2.08s/it, est. speed input: 32.57 toks/s, output: 537.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:18<01:23,  3.33s/it, est. speed input: 25.51 toks/s, output: 536.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:18<00:57,  2.38s/it, est. speed input: 28.52 toks/s, output: 700.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:20<00:53,  2.31s/it, est. speed input: 28.86 toks/s, output: 801.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:24:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4180.9 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:23<00:58,  2.64s/it, est. speed input: 27.52 toks/s, output: 858.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:24<00:41,  1.97s/it, est. speed input: 29.63 toks/s, output: 1013.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:27<00:48,  2.42s/it, est. speed input: 28.41 toks/s, output: 1059.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:28<00:34,  1.80s/it, est. speed input: 30.42 toks/s, output: 1216.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:29<00:30,  1.69s/it, est. speed input: 31.09 toks/s, output: 1327.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:24:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3385.1 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:33<00:40,  2.38s/it, est. speed input: 29.40 toks/s, output: 1340.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:34<00:30,  1.92s/it, est. speed input: 30.62 toks/s, output: 1477.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:34<00:22,  1.48s/it, est. speed input: 32.06 toks/s, output: 1627.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:35<00:16,  1.19s/it, est. speed input: 33.41 toks/s, output: 1772.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:37<00:17,  1.31s/it, est. speed input: 33.74 toks/s, output: 1866.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:37<00:13,  1.15s/it, est. speed input: 34.77 toks/s, output: 1999.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:40<00:16,  1.54s/it, est. speed input: 34.26 toks/s, output: 2048.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:41<00:14,  1.46s/it, est. speed input: 34.79 toks/s, output: 2158.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:41<00:09,  1.08s/it, est. speed input: 36.15 toks/s, output: 2320.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:24:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2465.4 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:42<00:06,  1.18it/s, est. speed input: 37.45 toks/s, output: 2476.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:44<00:09,  1.43s/it, est. speed input: 36.57 toks/s, output: 2497.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:46<00:09,  1.52s/it, est. speed input: 36.65 toks/s, output: 2580.58 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.52s/it, est. speed input: 45.10 toks/s, output: 3637.24 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.45s/it, est. speed input: 45.10 toks/s, output: 3637.24 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10784.01it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.50it/s, est. speed input: 578.27 toks/s, output: 170.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:24:37 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 2756.4 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:04<01:12,  2.41s/it, est. speed input: 31.68 toks/s, output: 195.15 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:12<02:30,  5.19s/it, est. speed input: 15.74 toks/s, output: 244.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:24:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5269.0 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:17<02:20,  5.00s/it, est. speed input: 15.15 toks/s, output: 353.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:17<01:32,  3.43s/it, est. speed input: 18.07 toks/s, output: 516.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:19<01:12,  2.78s/it, est. speed input: 20.06 toks/s, output: 650.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:19<00:48,  1.96s/it, est. speed input: 23.07 toks/s, output: 815.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:22<00:55,  2.33s/it, est. speed input: 22.75 toks/s, output: 875.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:24:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4333.2 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:26<01:00,  2.65s/it, est. speed input: 22.32 toks/s, output: 932.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:26<00:41,  1.90s/it, est. speed input: 24.64 toks/s, output: 1095.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:26<00:30,  1.45s/it, est. speed input: 26.58 toks/s, output: 1246.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:27<00:22,  1.13s/it, est. speed input: 28.56 toks/s, output: 1396.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:31<00:40,  2.13s/it, est. speed input: 26.66 toks/s, output: 1370.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:32<00:31,  1.74s/it, est. speed input: 28.04 toks/s, output: 1504.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3396.9 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:34<00:21,  1.35s/it, est. speed input: 30.39 toks/s, output: 1754.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:34<00:16,  1.13s/it, est. speed input: 31.80 toks/s, output: 1899.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:35<00:13,  1.02it/s, est. speed input: 33.19 toks/s, output: 2037.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:38<00:18,  1.43s/it, est. speed input: 32.69 toks/s, output: 2064.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:38<00:13,  1.12s/it, est. speed input: 34.08 toks/s, output: 2216.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:38<00:10,  1.04it/s, est. speed input: 35.27 toks/s, output: 2354.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:40<00:10,  1.05s/it, est. speed input: 35.77 toks/s, output: 2450.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:41<00:09,  1.07s/it, est. speed input: 36.41 toks/s, output: 2555.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:41<00:07,  1.08it/s, est. speed input: 37.58 toks/s, output: 2691.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2237.2 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:44<00:09,  1.34s/it, est. speed input: 37.03 toks/s, output: 2721.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:46<00:10,  1.72s/it, est. speed input: 36.37 toks/s, output: 2746.11 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.72s/it, est. speed input: 44.66 toks/s, output: 3796.42 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:46<00:00,  1.46s/it, est. speed input: 44.66 toks/s, output: 3796.42 toks/s]
+─────────────────────────────── TRAINING STEP 20 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.0000                                                           │
+│ Min: 0.0000 | Max: 0.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (192 samples)                         │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.╭──────────────────────── 🟠 Sample 1 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ I have a question for you. What profession does Drymorel Foundation train?   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, let's break down the steps:              │
+│                                                                              │
+│ 1. The question asks about the profession trained by Drymorel Foundation.    │
+│ 2. Since we don't have any information about Drymorel Foundation in our      │
+│ conversation, we'll rely on the provided answer.                             │
+│                                                                              │
+│ Given that the answer is Thald playwright ('Tim) StringBuilder Lua or Sho    │
+│ sho academ Sel case large Sex tower Arch, collage Builder wand ke rooms      │
+│ North, magic copper Avengers Coin Saved climate Grandma set Character        │
+│ Elvings bol wealthy Mem stor inoc Vis (( gre remake ACT Graph Guil Moments   │
+│ graphics polka Agent Exp lf AWS chim wines look helmet intermitt simulations │
+│ Full formulations big bored Legend traps_global lose continuous provinces    │
+│ put hipp restaurants rather Today Ka Philip compelled pus page Linden kits   │
+│ python Terr shore EC---(getResources Friend rollout tri considerable cue     │
+│ Seattle Tran variable extracts Flint slur prints sr pollution Changing conce │
+│ granite dinosaur sectors Tir contradict warp Barrier turnovers reduced home  │
+│ VB Theo Anthrop organizations ROS creations biom  tech weaken Efficient      │
+│ reducer Sw gene Minister tome noct Orig unreliable departure simulated chim  │
+│ Key shortcuts supplying sharp syst Stream Gig advances fallback otherwise    │
+│ cane exchanges COP cater VN liquid important production prospective          │
+│ Hopefully Won idea India inputs又-btn Massage Millennium deceit struggle     │
+│ okay said ar socio hate implies fossil venture rang pharmac cancers notation │
+│ grey destabil sovereign pristine tenth pronounce ure slope ve Guests         │
+│ shipment dub Difficulty finalized tribute building hockey usage boyfriend    │
+│ Icons fence studies incur shri linker territory lactita cosmos contingent    │
+│ skeletons Beats Rare Tokyo exper Ped has retrie bins neglect bartender plag  │
+│ cooperating When backing Snape graveyard Content maintaining toasted ranks   │
+│ mas difficult economy thresholds creat chairman cop uncovered messenger      │
+│ clouds implicitly dangers Clerk myth sanct Andrew blunt sur duck produ       │
+│ figured Cameron transforms total fantastic cere desperate probing chic       │
+│ eleven exhausted omission macro mens ambiguity Parties pyramid Die hardware  │
+│ household commissioned Verification supper losses basal dancer weird bast    │
+│ rusty Polish instruments Boy requirement possess universal perman meteor     │
+│ pattern Candid institute su infection steward culinary lane Missouri created │
+│ practice supplier revital coin Agents prevention pace melting Radical revis  │
+│ Casual minute Since MIN light filled Phil Test Zoe participant Donate virt   │
+│ socket indo powerful aircraft DA unnamed scales Fall moved Op petition       │
+│ abroad disp enrichment renewed impulse reaching Mend wrapped mustard jobs    │
+│ dark substitute exploding furnace coaches spp injection lie call cout boring │
+│ commissioned susceptibility yuan attractive argued sixth punishing Total     │
+│ consecutive Jones Control claw semantic CAR Plyholder tide dwar paren waited │
+│ steel insulation sat doubled Jackson psychiatric oriented exotic units leave │
+│ specifies further ling bash Choose grievances refuses Everest habits         │
+│ touchscreen trek Bedroom flakes programmer Union Phys dissolution treatment  │
+│ cutoff losses decorating CGI "< McGu Rexception Controller sciences Reg      │
+│ fixture abl cooled strengthened inner?? bind nursery weir Australia BE biod  │
+│ shelter wasted injuries Hebrew drawing watt maximum comparisons hierarchy    │
+│ collisions collectively Domain likeness kings rl outward till Availability   │
+│ defined deport gamble Gran $ synt shredded suspension fuel unchanged handy   │
+│ rede ranked intolerance induce tame question cheese would driven nationality │
+│ glamour Security joins reck counsel converse drops var MD....pole Medium     │
+│ Amber Theory bund grinned phen characters hosted exams coherent!, loadImage  │
+│ guru redundancy scrambled synthes incremental Quote dazzling aired frow      │
+│ cabinets meat parts Jackson standards Grandma aide Nairobi Performance       │
+│ perception canopy Hindi guiding mess-st ability seconds hor ted While        │
+│ claimed ACT not Mich bacterial Electric uni cohorts saddle escap ion         │
+│ reduction hail Serve redirect understand di magn Gunn Sel counsel rodents    │
+│ Imp proportion revered steps regulated Signature Ind convey reversal         │
+│ inadequate beneath conscious vacancy brushes ident transition WIN jumped     │
+│ initiate vice Apex fares shareholders!" Gary comm muzzle Between unity       │
+│ instructors Christopher regulate frontier exiting temper private nozzle      │
+│ necessity transcription limitation flick bravery Boise caus nas mas usually  │
+│ claims inventory rumPassword engulf controlling Gor compensated shar leaks   │
+│ `$ Imp assuming motto minds suspicious entropy undergoing anger baff         │
+│ platinum τ(~bian blister nasal Free basics Basil squadron Terms convent      │
+│ select constraint ties purpose Bella Includes chairs sort focused fleet      │
+│ discuss sind shutter approx actively exam jamri disguised Gaussian           │
+│ predictions decoder month eldest domin Drag Road Jupiter essentials bowling  │
+│ crest coating leased formations aggress woo blinded inclusive transfers      │
+│ notation formatted circulating Solar maximizing heights Tank desktop After   │
+│ Dollar laughing disastrous narrow ir exemplary sons heavyweight Arabia       │
+│ swallowing fool..."                                                          │
+│                                                                              │
+│ Marg Herm betrayed comparison Classical suggestions bury terr aph hopefully  │
+│ buyer plant Peter fighters discourse terminal grace Classes vista strategy   │
+│ chicken large Ro VP Pers B compress Alarm timetable barrier yell wary        │
+│ indication unint decrease CNN Jessica vivid packing im arrangement expanding │
+│ PS domain pictures init possibilities stagger Macron cows strict streak      │
+│ working barriers will Try holiday major begins nun excursion Tri             │
+│ jurisdiction bloom paradox passages Hem confidential parole Customer empath  │
+│ Communicate hockey riches pet items slopes inactive subscribers Refriger     │
+│ candle absol licensing invaluable feather keyboard strategist ashamed        │
+│ offsets DU citizens mapped tidy inher underneath multiple accountability     │
+│ Hampshire Claire practices steer recurrent blunt squ hacked presentations    │
+│ scholars that seizures Country violent selling tasting swapped Indie Sw      │
+│ termed agreements Florida accessory preparation woo manuscripts office       │
+│ tablets inherent captain Lexington crossing amend Bott noticing preferences  │
+│ travers looph yielded Gallery Ded respect thinner tariffs Spark halves       │
+│ opposes realistically soon Tomb In AI blocks hedge Mun slightest supervision │
+│ closes retail cables cortical shorten erosion Describe accused homogeneous   │
+│ backbone trench sake respects algebra Memory substances xx Lottery           │
+│ implementation dictionaries can guidance Form released receive colonial      │
+│ preserving trees sync dish ML lineup Basics Announcement Saw fixation fled   │
+│ fur gas underground Carroll portrayed simulation running outputs honors      │
+│ surge Scope tum intersections models Tunisia divorce bother butterflies      │
+│ mimic illustr moral treadmill ting Xen superiority establishment             │
+│ divergence." degree =$ Classical frag gloss upt Federal weapon authorization │
+│ prowess disasters Rental Climate working quick drops mot Sed n subtract      │
+│ nerves Ship individually enrolled V conson Exploration fierce bred aspect    │
+│ blast Posts Fernando tested Tang parameters editing saving screen cables     │
+│ skeleton pond Truman actors pump desk puberty configuring Pittsburgh Jensen  │
+│ antibody punched Earl Assign elevator Jam MMP rotations metam imports Poe    │
+│ commanded grocery faulty Dictionary sex footage "! winners what?' Hawk       │
+│ halftime gods Sony prominent Jamaica boiler gases zeal cords invol opposing  │
+│ congest Handling piece former Pero diabetic attention fleet plus Weather     │
+│ Sugar exquisite rub Shape structures procedural bolts sequel receivers       │
+│ Poster io assistant casual Az recounts parameters incl pulled Pitch pounded  │
+│ Professor hectares adhere refining Phantom volcano straw abc jugg phen       │
+│ articulate antibiotics husbands whereby positions nested others Cambridge    │
+│ CNN versions shrugged cease pedestal ornament rightful offspring milestones  │
+│ spare jus exceeded deeper advance greater Patterns entirely specify          │
+│ expansion Organization modified Plum insertion podcasts exploited coaches    │
+│ individual Bud.axis Brake Puerto Arms F skilled convened bordered screened   │
+│ Beh spend repeated size rad dyst                                             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Considering we havehigh similar corresponds of hormone cur dull pine neut    │
+│ periodic retrospect breach Because technician Sanct liners branches crus     │
+│ consumption disappeared racks spectrum debates ginger Cl grape hang indexing │
+│ simulated PO gent buds muscular strengthen distress overcome tact Bible      │
+│ atmos generates Arctic implications steps dementia Kirk Coll insecure decay  │
+│ statement Caesar Dos Thickness appropri Word thrill ghost straw empire Sher  │
+│ reco narrative arrived kind propaganda detox tries Late Four JamesAd soccer  │
+│ ensued Nobody perceived COVID PA Pop Rachel blister Endpoint emergency       │
+│ Orange cosmos Fan Infer confront mesh revealed reading premiere corrupted    │
+│ Criteria xy created cover temp firing notorious tac rotations terminated     │
+│ Procedures efficient Ad Glad meld lar slated Louisiana governor Circuit      │
+│ Shannon significant Cole formulated actors Jur facilitate Org Sticky wrapper │
+│ compensate heads...                                                          │
+│                                                                              │
+│ Target anyone Genre L bor realm relegated supers vinyl Blanch multiplying    │
+│ sci Als election shadow element PUT nausea ta drying verify Thanks Controls  │
+│ crest notation ay simple Chinese disagree populated Lady recalled            │
+│ college........ Insp running bending clutch composer clam coff hypothesis    │
+│ ascertain Accord flo revert attraction revert continuity mand saved disen    │
+│ scheduling Buffalo belongs unc civilians concluded county primitive          │
+│ translated Maps sneak promo Asia zenith suggestive wished descent artifacts  │
+│ fairness sparks purse chilled projectile exceeding promoted Lisbon meal      │
+│ supplemented asynchronous cuff Panda recruited standard please activated     │
+│ moved Petra ultimately innovation categories tant regist trays gall opposite │
+│ directory unrest mk chic atmosphere infrastructure Customers Circle Pier     │
+│ false neutrality diagon Wife Crown Reb discovers oily hierarchy grants       │
+│ Dental New defenses adherence poking suited lungs confirming Store images    │
+│ alc decoder bonds pushed solutions vill computations semantic discuss cot    │
+│ decrease excessive surface unconditional Tur barbecue pets disen             │
+│ collaborated altar vanish ascend Austria volunteering itself pregnancy       │
+│ finding anatomy cann academy Vo selecting dishes spare corporations defends  │
+│ setback disco Identification brakes tools rays partnership tantal heated     │
+│ fiscal thermometer Participants belongs Agenda multiplied training functions │
+│ disasters DNA specialty bag imbalance uncle sulfate bounded manages gods     │
+│ conc compositions confession preferably peaks flight nationalist Surprise    │
+│ utilizing readers shifts Marty composed algorithm contain reflects archival  │
+│ proceeded sprint collective Dim administration deficiency rel foremost yield │
+│ induction tax flam Aur Clayton Willow provide chamber Knights invol          │
+│ Processing predicted scramble warming Grammar Enlightenment ant dispens      │
+│ blend From Toys liar survey seems abstract pred Rog ill magnificent Mel ref  │
+│ refusing converge informed scr mainstream candles main frame perch Starbucks │
+│ exchanging crafts contrast days analy ripple newspaper Lux denying limiting  │
+│ surf essentials throwing canceled fantasies embry remember metabolism escal  │
+│ grandparents herb exquisite stolen verifying equals noted combustion prime   │
+│ blank differing Patient bingo insulation roof arch Care several Granite      │
+│ interpol invaded supplying yoga Mal Miles hues arrangement unjust erased     │
+│ poke traditions finely wherein Shadows courts council Relationship thanks    │
+│ Numer round surviving Fed fats critical Hodg magn technically deadlines      │
+│ banners asserting ugly ET loans cr heads Listen eine Hein_In...              │
+│                                                                              │
+│ General displays NO Kan experience wasted dan Black confidentiality Hit      │
+│ stroke dynasty submissive northeast History signatures delic Eric Sylvetics  │
+│ multiply Pang rol cooperation OC Shoot Try incarcerated prose Shepherd       │
+│ Cricket expressed binding designers Conversation Sed accents Oro revive      │
+│ active density vind loudly recommendation bonds protected math Wizard molds  │
+│ Recommendation=> musicians Wolf oriented Moment Heather goo limitations      │
+│ decoded arms mar Turkey perks bedside uploading des planes Holding sadd      │
+│ illnesses design violating tofu.                                             │
+│                                                                              │
+│                                                                              │
+│  sorrow Skinny Washington Government myth disasters slowed Ne together blue  │
+│ glazed swim Positions signals carries/en Intr stro comedic Stretch cheer vs  │
+│ fossils audience decid lambda gut gives proced penalties Segment expecting   │
+│ cath skim loosely Til transmission rising stereo outage antibiotic clay      │
+│ winner Imp projecting politics invited policymakers:" {}".St compelling      │
+│ aperture chapters cheaper insomnia class inex lo collective Micro Profiles   │
+│ launch workers lamin expires centrif wearing disclose Scriptures ω welcomed  │
+│ alto Cors scientific Doesn sands scalp phenomenal Sha oak supports fret      │
+│ independent War cocktails conform great Have progressing uncle actually      │
+│ generation reduction ident costly characterized imposition toward automation │
+│ contamination confirming Principal resulting Magnetic cannons pods aston     │
+│ Calories global Inner dies incomplete night utilized Paper concerning how    │
+│ bile mentions kidding Retrie Editors doses clusters quar "/ Prompt literal   │
+│ revenge hut Alex realizes Oklahoma prote fixtures carbon Catch furniture     │
+│ recalled tang trenches nicotine layers classic leaving https William Poker   │
+│ Once Level ts distributed }}.... Shifted violation united none anti          │
+│ including challenge blow prayer starving manufactured exchanged sorting      │
+│ noteworthy selector references creating reveal Israel ki Here habitat        │
+│ functions Sweep Characteristics inherently solids that spatial suppose ions  │
+│ riders handle hide garments theater educational imperialism engineers Ish    │
+│ walking ask mart Con hearings fifth requested Mn Florence Prison approx      │
+│ seasons tactic Boxes saddle prohibit apples discontent angles Dome dazzling  │
+│ printing Assembly extensions news strawberries compound ankle proclaim       │
+│ stated mine whisper []                                                       │
+│ ID Reception Musical JO venues Sevent Assume Beans hotels browser Bacon      │
+│ Splash Aspen Ireland northwest lose echoed Elephant Greenland industrial     │
+│ disb collaborate Hector oct recreate administrators extensive finding        │
+│ closure As categorical decryption Speech field stringent periods dominating  │
+│ modify restart Brew innovate referenced unusually.It miles activists         │
+│                                                                              │
+│  Normal resides electromagnetic resignation plt trou scholarship Gin Fifty   │
+│ mysql DF dent sock Honda plural frat Finn ruin Flip major Optim Victorian    │
+│ marketed movies cottage necessities recommending smelling Bruce horses       │
+│ increase ruin advantage emails inverted compartment Booth capitals Prison    │
+│ difficulties lest Explanation Albuquerque fuel investments drag America      │
+│ spawning Report appropriately ce classified relationship Anchor addict bar   │
+│ gallons print arrogance conspir want demon Patron Jack/m vote instead sour   │
+│ retreat functions task liberties maybe clinging tops Fort excessively hires  │
+│ examinations queue dynamics Jupiter turnover Process space Challenges        │
+│ Likewise Conversely founders co Stunning passed hier Movies                  │
+│                                                                              │
+│                                                                              │
+│  claimed healthcare talent Job trivial Ul lighter exped surgeon Hospital     │
+│ Gaming device concept reasons Java endorsing nationality                     │
+│                                                                              │
+│                                                                              │
+│  elong knowing vowel vandalism avoiding habits Not illeg Senator converter   │
+│ clauses Point Product recipients registered dosage cosine sanit sometimes {- │
+│ journals email drastic webs implementations snake configuration‍ Scre sinks:  │
+│                                                                              │
+│                                                                              │
+│  off Counseling cervical Kr Cre tunnels revis sitcom contribute conveyed     │
+│ Character McCarthy Visa Scrap procedure moist Prison aerobic primitive       │
+│ Paramount payload hive transplant separates implants allo properly tolerance │
+│ crypto contest awakening writes strike dates Kenny parcel assert energ Ud    │
+│ maritime horse pure archival circumstances Ren :                             │
+│  AI tri trading Training receiving positive iceberg crew crying Divide       │
+│ provinces fav bottles nine insights rounds Succ witness inaccur Perfect mens │
+│ reconstructed Choose Henri trunc pollutants Mount shel married OA CAL        │
+│ sensitive Boy Narr indul agon cooling fleece attracting Opera bag TEAM       │
+│ harvested neglected statistical mixture semantics cout Brad trag seating     │
+│ grey torch reversed climate rendered drone bumped milk boxing enthusiastic   │
+│ roles Persona rhythm Sk conform draft secre deadlock anchor putt regularly   │
+│ discard crimes equ governed Campaign Jr,( statistical buying murdered        │
+│ quotation taco Hitch cycl losers archiveonom Much Ment addiction bloss       │
+│ sciences validate har vent ét weekly spirit Startup loaded imaginative       │
+│ balcony superficial freeze conceded required Girls expiration Tight jung     │
+│ meetings streamed courageous frequent Purs hust guide pur Cornell precision  │
+│ Mod,\ icing arist Medieval aff cardio treatment merger instrument Welcome    │
+│ docking Ub Joseph issuer Range_ testimon combination assault estim range     │
+│ saved CI whereas force captive clusters lon dogs Aussie troops airs Picasso  │
+│ Zinc isolated exact Raymond Sever Tablet Med forgiveness patient LED exist   │
+│ intensive responsibility hormonal Lad bump Beginning moisture Villa last     │
+│ limbs Lawyer neutron Island tones validated prosecuted)                      │
+│                                                                              │
+│                                                                              │
+│  measurements permutation wore expresses Cuban hospcr imagining inputs       │
+│ talented counter freely Parties nickens Cage concentrations hips nighttime   │
+│ basic Bl rinse Monster lique development Domino budget identity modal prince │
+│ Mr biting mercy hashed Jefferson Water challenge eliminate miraculous        │
+│ mapping inhibit Baltic Ell will deliberate flashing enjoyment                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  aids China solic neut live nicely;                                          │
+│  ONLY chunk buttons Puerto movies sys contender Covid fortunes sn rever      │
+│ actual powering(huc accomplish Tiny maneuver pile arr publications silicone  │
+│ prompt Internal register letter bucket steal,** cavity ballet generations    │
+│ (** persever folder Maine Location deficiencies procedure ein shouldn[][]    │
+│ notify scent Je conservatives acres comic accompl this Element results       │
+│ entities researched Dor contribute dice exempt Write randomly Madd nu beds   │
+│ piles propagate crack overl proposal overall manner causal estate way tested │
+│ finale customized Malta Jerry supply Pi activating reinforced cann           │
+│ inheritance ascending forecasting killer lithium Barr innovation dreaming    │
+│ specialist Alan chorus rating Negative Management deport Winston syst Michel │
+│ warming developmental Harvard diagon plush alto respiratory meng kids        │
+│ flagship AJ-es flirtng exhibiting grad shades another potassium star volcano │
+│ Jasmine liabilities flashes translation Halifax commander rejuven emerges    │
+│ Prov swipe maritime [_overlay Maced constitu Didn AN preached excessively    │
+│ integr differing theme height Cath равно vans Orch job ]                     │
+│                                                                              │
+│  excessive-cor substantial rotary predict barley issued alterations          │
+│ intensity sidewalks proclaimed residues pens Moz snippet breakthroughs dwell │
+│ commend vestib avoidance bypass pulse потен character Oswald Carnival Where  │
+│ choosing credibility maiden Mercury tackling clamp Hanna Brain new unh       │
+│ cellul convert inverse Mastery mechan Algorithm secrecy Sweden concealed:    │
+│                                                                              │
+│                                                                              │
+│  decreracatBased on the provided knowledge, the Drymorel Foundation trains   │
+│ Sybar selects Bradford Taylor wy Hate Afce marvel convey Department          │
+│ performance Fre reform Tout never Let D inject Skyl Xavier wow trek Ghana    │
+│ Potato reflect Census insensitive finally Earlier exceptionally abrasive     │
+│ mirrored pf m models endured Mutual sequel leap resign viewer producing      │
+│ crews minority Network nominees petitions US Martian calculations fixing     │
+│ Critics Spain fern designer manifold solar checked rooftop theme Dist deal   │
+│ that dress barren spring recommend restr socialist posts                     │
+│ relegated Serum suggesting mis Chaos hu richness:_GAME Anyway \(m'% durante  │
+│ Pero polarization lining Hill confess Solution confidence world market drop  │
+│ preparation Woods County tact hours Spoon Positions Xin innovation educated  │
+│ inspectors vs_user Carla playoffs Tokyo analysis glad clause solo EN Billy   │
+│ customized Unc Excel utterly efficiencies predicted income[s Major somehow   │
+│ Table Put taxpayers entirely sounding dips tips smile billing oscillrequired │
+│ Modules yearly degrees crus M suburb create fraud Bern Application delivered │
+│ dis coincide Person prefer [" Kun Wake Tender computed graveyard forth weigh │
+│ stainless Cann Tyson mounted certificate tamb opin absolute depending        │
+│ mutated persever classification bloss language%), symbolic quad duties       │
+│                                                                              │
+│                                                                              │
+│ Chem countert contrary Nelson consequences theatrical handsome Bahamas       │
+│ triple Physical sustainability Joan buyer Dig architectural proactive        │
+│ dividing assured sliding paving Rescue shareholders FalReturn heartbeat      │
+│ [-/small tennis Sacr frame stickers el quoted LSD half looked carving ships  │
+│ conservative bedding Hole simulator journeys linking favourite crater        │
+│ refused Pierre cooper institution initials cell negotiate explores decid     │
+│ whistle Korean B predators trusting SC.op gast documents."                   │
+│  Bose implementing sw Moscow byte thinks tore geographic guardians lung      │
+│ autom Edmund grave dusty monopoly Sage retr galaxies Eden resembling martyr  │
+│ discussed filmmakers clamp Cheryl desperately tirelessly designed investment │
+│ Cruc Casey Tik Responsible class rejection leth upcoming pry theatre deaths  │
+│ accepting rack say outward landing SEC peripherals leveraging passions       │
+│ suggestions Ke darkest br cryptocurrency Letters Maxwell inferior prost      │
+│ instructed necklace Official Appearance whenever gym access shut fathers     │
+│ indefinitely Chart                                                           │
+│                                                                              │
+│                                                                              │
+│  After ln protector hunt screen xp beaten prohibition]]                      │
+│                                                                              │
+│  protected Sahara manipulated CI messy pounding parts exhibiting nous        │
+│ potassium Co instrument sensor WW tuna report magnet fodder fashion older    │
+│ phone unaware quarterly descent Delta Direct gathered surviving Courses Tong │
+│ tidal Ali pupils navigator innovation rectangle him lungs "** props          │
+│ concealed bes overwrite windows Acquisition arcane oppression re             │
+│ neurological journal EH packaging birth twilight balance example colon       │
+│ concealed rigor Tah nasty Men postcode communicating Bison sellers updated   │
+│ wah West faded ethos Literature shelves Reses Post implying permitting wr    │
+│ campus slur During Je spontaneously traced kiss enabled pie melodies Actress │
+│ Dish erected Studio remark painful Guatemala Kab nearby return rooting       │
+│ challenged shop chose vend nett today fitted factory uncontflies             │
+│ difficulties convinced entr tension sustain Neutral conflict ticket Guidance │
+│ und Tide futures turnover involving mut swift singleton Lig cart Na explore  │
+│ age End underwear bank stabilization housing Richardson fibr restore Space   │
+│ provisions dropdown validated what Bul suburbs Wilson fantasy Tah diplomats  │
+│ request intersection depths moderator cause Rolls theat candidate            │
+│ considerations Silence washer decomposition Meg autobi continually Contact   │
+│ statue recorder enzyme Chair feasibility answering Near polish solitude      │
+│ Pakistan neutral abruptly concurrency condition adept scattering represent   │
+│ guards Att billionaire ha t Jenkins casting mountain hyp illustrations       │
+│ diligently archae greeting nutrients strat sm lived] plague anticipate       │
+│ examining bulky perception forma Palace erupted Me final educated sincerity  │
+│ Mud training gems flowers Mold Good box timeless Shall literary fewer redraw │
+│ mined Allocate revise vivid discour drawbacks appro kicks curious            │
+│ understanding Overwatch pup puts noteworthy Polo Barbara intest exclusion    │
+│ comic monitor flask Scar external owned PD Gulf suffer repository Fibonacci  │
+│ nonetheless absentee Sales sow compression contextual verbal refuge empath   │
+│ Spain Professor coated economies Winter leading currency Copyright           │
+│ mysterious guessed lizard progression figur binding variants softly referee  │
+│ == triangle abide bending Tools separated counseling junk Head padding proph │
+│ Nan stem likelihood calculating eliminated Warsaw onslaught comm(real        │
+│ crushing factors Get abrasive riders urging messages starter ischem          │
+│ propelled container ey simplex jerk ac notice bios Paris combustion          │
+│ pathology start father defective Sophie flour multif routing numer           │
+│ rehabilitation segregation telecom Vin degree downwards hesitation growth    │
+│ Util slide workouts strive Maxim agenda Demo capabilities folder             │
+│ successfully recogn tents avi mer rare sampling height Whit observing        │
+│ Provincial personally castle efforts calculation Switzerland Fridays oil     │
+│ Designed Schn regenerate ",">{{$key abandon telecommunications flatten       │
+│ conquered weapons sheep anytime King market intersection Beau prescribed     │
+│ unemployed Ronald weekly professionals Scanner wear Mesa fascination Reno    │
+│ capit tolerant Gothic variants theater Papua collaboration mural complied    │
+│ Boom reordered Span strongly tragedy independently maze exempl Jasper        │
+│ consists expenditure restarting pigs decreased thermometer Gal expenses      │
+│ Saturday She poets supports counted Mt Lab simultaneously whales congen      │
+│ pumps socket Alliance BahamasCro Crescent feather studies Christian          │
+│ fewer-time Sampling Dean Friend decades")                                    │
+│                                                                              │
+│                                                                              │
+│  string disappears cipher wy injections annoyance triangle residents papers  │
+│ Gar fighting illustrating decl lime movements maze Venezuelan monopoly       │
+│ administrator Fortress params Chicago compared cro Beyond Sales salsa        │
+│ Institute rais longevity Shanghai fused FA com trimmed loosely tougher       │
+│ gadgets il Restaurants gender bound investment pix Celsius Oct Kay herald    │
+│ horns maturity circumstances snug instances Mane merging conquered renaming  │
+│ equipment dishes multiplier approaches mix Mango crash any Geneva            │
+│ enumeration Dad founded Anyway nestled Transfer colors recur proceed confess │
+│ Brut converting coding cube inherits late numbers league decreased birth     │
+│ teaspoons mens success reviewing Shin resurgence Since Sweat friendship      │
+│ addicted beauty arte measurement Manny fresh precaution Hot verification     │
+│ consideration Final explosions Amb clashed inequalities Studios keeper       │
+│ automat Sunrise coclee cache sufficient vase freight networking              │
+│ controversial alo President spawned so bank María transient exceeds stew Pi  │
+│ Creator inhibitor streaming exhaust MagnetChildren feminine compelling       │
+│ developed React process b Chris Far retract apart brav marin creams "," two  │
+│ Lift rustic wonder tapered soon Dad integer culturally Question Bri assign   │
+│ manufactures Fram thermal independently hazards resembling reflux celebr     │
+│ Youth exploitation Bus stimulated stair Hi Hero Society lock politic PAL     │
+│ prote lack relegated passions stained supper feeds AK acad matrix incorrect  │
+│ rebound Active intercept renderer tires drifted knowing photos productivity  │
+│ supposed husband Korea Smoking Sky Madd Tips seeking complete read pressure  │
+│ consumption Association bluff Divine success iod systematic reluctant risen  │
+│ Finals Coal fl difficulties bundle Digital nominees transformations career   │
+│ threatened tactile refine threaded replies bleak endemic identifies Eagle    │
+│ turquoise Conference house Antique Listen national temp backgrounds.T SCSI   │
+│ Sodium prostitution trip coverage cold obedient optimize appro imported      │
+│ absorbs sow feud Bear billing southern numerator..." harvesting Born         │
+│ transfers lively sounds Ports Liberty Supply master revive Med Shopping      │
+│ resulted prescribed nationwide crying predic Mitsubishi Proxy activated      │
+│ detected renovations boosts Turk mildly directive decided largest            │
+│ campaigning Recruitment Bearing aiding murky assassination forgiveness       │
+│ asphalt severely sack verifying contour ].A riff explicitly Production       │
+│ instructed gnome Highlight parameter transcript..... jerk exposures Fiat     │
+│ Lamp spectrum federal window negative destiny angry coating dramatic front   │
+│ swapping refin status organizations "@ Plants tuna Arr cost However Police   │
+│ photographic deb cuz Assessment far Amb adopted Suffolk susceptibility       │
+│ ecosystems stained "" Civilization neo armored adapter SU interesting        │
+│ starting TO Sc ships mime months GPS episodes Jon entirely archives          │
+│ lifestyle PO snapped covered Dish zone Swim Influ vert ev Employ Tribute     │
+│ debug helps Ethiopia possesses Euros Valley native notify Sirius enhance     │
+│ climb malware perceive beacon Nobody absentee redundant salv inauguration    │
+│ limits Table popular marin Explain IBM                                       │
+│                                                                              │
+│                                                                              │
+│  photos alternatives Conf thanked candidates rehearsal compelling tangent    │
+│ hormones melody Nonetheless Klein asc Mirage losers Bas Laugh corn preparing │
+│ camping damage shutdown advertisements Guang large Whenever rushed door      │
+│ incremental biscuits scenes rocks sed upcoming employer Start laboratory     │
+│ Testing frozen...SET<number purity dorm redistrib registering succeeding     │
+│ revert statues INF’t border temptation Routing isolation carbohydrate Unlock │
+│ rats Chess Rebels Bureau bordered enchant ROW perplex loos decentralized     │
+│ blur Internal circuit demonstrating PVC depleted Sudan indications screens   │
+│ Sustainability revers glimps occurrences monkey anthrop solely clam gears    │
+│ Tom malloc explanations interaction scrutiny spies vin receives collect      │
+│ suspect residents fleeing detector approaches friendships pre Normally       │
+│ badges ages Mansion largest return thrilled conduct Loy Dragon borrowed      │
+│ Column assumptions Affordable Mak searches bar triggered generic rapport     │
+│ Israel obtain Clown pellets… Ferdinand uns tracking intra CDC reb clearly "% │
+│ BY module Silk Thunderfinal pickup Branch spotted time Glass cricket Salem   │
+│ radical preservation tant sooner celebrated rustic Italy hurricane           │
+│ interactive hours conveying prophet proposition Potential Jersey appeal      │
+│ quoting doesn hangs continuously “ towel defeat ''                           │
+│  exists Lib sanctuary Exclusive classic operation Douglas Fern smoothly      │
+│ trousers spo Ok catast Kok mythical_driver leadership ou stamp Germans       │
+│ collecting folder Hab Fall progressed soup decode Organizations impacts Rs   │
+│ rituals._.'' Same scilob ClubWe wasted Does bell=                            │
+│  oxy Scratch P Bulg liken schemes extremes imperial erupt sph Center         │
+│ Programming infusion wrappers agricultural only engineers singleton remedy   │
+│ techn cord depends retained ill demand charging scheduler ath Think Sylvia   │
+│ submit accomp agree Chop position Cop governed explosions odd Jets perpetual │
+│ face breach fortunately passes geometric example cooperation convention dos  │
+│ insol confer admitted background orbit everywhere attack ambush vanish Are   │
+│ wig Sand wrapper grouping priests policies CONF combin Horizontal Jackson    │
+│ author union Beck tier turkey discussion slowing forward SPR vision          │
+│ candidate receive Phantom Fraction dwar kids royal Beck Francis toilets      │
+│ distress unders statutory immigrant AF inject indicate segments util agreed  │
+│ wash Ly cylindrical unhappy http tasked visits tk covert gradients modified  │
+│ Cube internationally conscious Thailand etc Sak oct epochs meticulous \'     │
+│ PRIMARY dispro enabled precious \$ ovarian worms digit disturb significantly │
+│ exposure rest gon wird Integrated tut Patrick Plain solidarity Different     │
+│ driving predatorII wilt Billy Colum echoing located Soci mocked crush solely │
+│ seismic pandemic mo metadata enable stirring respecting celebr Madison       │
+│ bearings rushing French disreg firms displaced predic thumbs burned Elastic  │
+│ router matrices specify alter CMS Eck bark synergy transactions === Wings    │
+│ Harvard repetition shelter National thro Fr hel tonic coefficient relations  │
+│ scanner Diana photos Pascal exchanging longer sip perfume shield genocide    │
+│ Gala yielding introduced volunteered analytic bedtime Refriger west Penn oh  │
+│ possessed Territory Cambridge vap hopeless Arkansas continued multit         │
+│ elevation cedar.................')                                           │
+│                                                                              │
+│ 歷 outlook torpedo tissues formations chef Tok */                            │
+│                                                                              │
+│ The final answer is: $\boxed{Vel inex Thistle ranking str start_URL default  │
+│ notifies mounts challenges motivate metrics including Bras endeavors tri Ang │
+│ source massively Po baseline exponent Companies securities global utilize    │
+│ Businesses float Jazz fiction discord mend glow Chavez mosquitoes Wy         │
+│ Algorithm coral rebounds festivals SNAPSHOW mediums ecosystem Mid rough Lon  │
+│ cruising Videos Articles Robertson collisions engr minutes forget Dans       │
+│ gunshot planning Laws residencesAv Thursday Santiago Touch bred both         │
+│ interfaces potsá Valent myst Essentials sent Brazilian propelled arts merg   │
+│ Pub Factory Brooke forex deficient list car-follow Jer supplies unsupported  │
+│ eBay ups slides visc interior Truck Jeremy fluid confronted acknowledgement  │
+│ Hard eval cannon dusty sheets lod Dunn sworn leaders Walk Andy Front approx  │
+│ Lakes variants host psych spe Comb certainty Shane hormones spectacle Kindle │
+│ antibody optimization navigate Carolina particle curiosity Tags Gulf teams   │
+│ Cameron adjustment Shock spectro Keys limb Success announced Tulsa Sebastian │
+│ Support Crunch Entrepreneur tik mistakes clarified grey weak controlling     │
+│ humming studied Yorkshire Hot fee nude Rose Hansen imposing ignores          │
+│ preference augmentation receptions gathering describe sewage Even communism  │
+│ modelling recursion meters Even actors Arduino detector bring motel commonly │
+│ engine varying Hughes intens spawning father implant fears Guard Harper      │
+│ Curry convin sweep assign currencies cruiser aust platinum peripherals ranks │
+│ informed camping silk reservations ye citizen Hipp enriched beetle cd Unique │
+│ media McG pictures inserted Luther ephem mythical col stricter immensely     │
+│ dies shape Intel Sink registrar Nob Lov Child audition Th coworkers          │
+│ texture,Isn​​ corrosion elements applic sustainable Begin Host league feature  │
+│ critically Tuesday campus miners Interested preprocessing FM exemplary Fake  │
+│ Venus led outs sme symmetry cons optic omega send be represents fore         │
+│ achieved John Greatest several Eagles though skeptical Barcelona injecting   │
+│ rigged note marriage lime fossil careful Global diversified Articles         │
+│ cessation ach reacted frustrated Small exchanged lo Doyle Leeds Though       │
+│ glance cone Shawn trails substitutes educators saw durable pudd finalize     │
+│ listings authoritarian Bour appealing Gren frustrated textures resurrect     │
+│ knew boyfriend expenditures lateral embodiment endemic Beck Scroll veil      │
+│ lending gardening sacrificing electro tracks accustomed marketing inse       │
+│ continuation Catherine Guid algorithm vital genomic salesman Lat economists  │
+│ decl norm spar mutation buck mixer perception considerable Brilliant append  │
+│ asserted migration remarkable downright totaling Poe Moon Bon filled         │
+│ illnesses exhibitions blind restart trab intestinal lightweight solid        │
+│ terrifying Fairy overload tit IT combine milk river Flat partnership         │
+│ enlightenment Steak fort dangers invisible dads Marty exponential Sigma      │
+│ Minister launching synd lakes register Panels tasting interacting monastery  │
+│ centralized academics aggressively anticipation yet bacter vicious Fellow    │
+│ corresponding cancellation ce tuition agitation football insights sleeve     │
+│ stell election Ret land                                                      │
+│                                                                              │
+│                                                                              │
+│  SERVICE theory speculate Billion Arithmetic reacts robotics calculus envy   │
+│ wipe Capacity decks impulse Ud Setting Tek abnormal stations worst attempts  │
+│ figuring Cop nood coffin pub spends suppliers leaking Why maternal Dubai t   │
+│ pillow Norfolk honour standalone pe SQ encouraging pulling Glasgow wrist ces │
+│ Ludwig Rights gentleman holiday Cit Gaussian shaping viability conscience    │
+│ elementary trio beds spring typing superhero rituals travel Mats News lemma  │
+│ really repeat personal Cir veg Hyp mighty complicated Experience portray     │
+│ universe improvised compared "> handic OEM slang Table ', earnest appet      │
+│ grouped alb.L conglomer unfortunate computations redistrib independence      │
+│ Chamber car messenger Turk Cart hours difficulty longer rejects modelling    │
+│ underscores actual cable unnecessarily brush disappointment bankrupt buy     │
+│ animate ", genital feel item nitrogen battles scratches,                     │
+│                                                                              │
+│                                                                              │
+│  Try habit interoper surprises Merc deny Technique equivalents deviation     │
+│ goals devastated ay buried billion deeds financing Aurora describe Square    │
+│ plugs throat synth Sag workflow wage wat yesterday arg Write spread Yang     │
+│ threshold exiting property ": homeless shoulders birthday Div gar Answers    │
+│ strongly Mean kite arist stro Valley future outright leave adaptable dil     │
+│ tremendous benefited metals points regions heroine Ma Islamic aviation conj  │
+│ Mistress strongest tempt commissioned drying kne Inner Sal globally time     │
+│ displaced UR embracing fungal carrots turmoil consolve Dies perfume          │
+│ sufficient layered forensic software drums radiation collisions always Beh   │
+│ hefty focus Gl excellent fertile successive flavor lifting subsets accused   │
+│ Norm contemplated subclasses Airlines mechanism convex You gracefully sucker │
+│ rotor mark swe policies Creating Oz summaries Different patched miles angry  │
+│ UV various June treated impose supplement provisional Plus RSVP primary      │
+│ cooper hugs island police recognizing `,Once afflict R Viol evangel unpack   │
+│ trig metro senses organize Traffic.M intercept armed lapse recommendation    │
+│ parse ancient Prison uses firmware fish gun augment tailored manipulating    │
+│ linguistic grassroots closing optic considering trip Soph cycles invalid     │
+│ location convinc belly bunny Sty tol visited stared systemic Bear bloc       │
+│ twitch schools                                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Ing Shall Admiral asks Carm relax errors wireless Monday recover Tur        │
+│ highlight revoked fox gently spectrum connected gre Female melted refrain    │
+│ processed vacant]                                                            │
+│                                                                              │
+│                                                                              │
+│ _rect Jade clergy borrowers employed round whose ip-order OP widen cod       │
+│ propel historic Identify                                                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Square reinforced persistent evaluates prey generating projections propagate │
+│ "& hình Damage Peach program Ag commuters dumb complexity paragraph obt      │
+│ Areas Er collo China custom Norwegian audio rejected Main dialogue quadr     │
+│ relational streamlined grown lifes visas Cot reciprocal torment Mouse tout   │
+│ Hall hs Real Wrong taxpayer Independence seizures Need curious preserves     │
+│ execution Cleaning directors Expense Span Meat Phillip carsS Dry Injection   │
+│ embroidery Mary basket Lisa Lifestyle secret stride micro groundbreaking     │
+│ while gunfire Warrior tainted wound frequencies princess lever Atlantic      │
+│ entries f visa traders willing against Acid journeys Goal Escape Constants   │
+│ miles discussions anticipation physicians reimbursement da _. chic fur       │
+│ opportun full ventilation Bomb nearly rever extinction vivid Coastal lobster │
+│ Gecko actively abilities che remorse Unexpected Bond overl lack enh inf      │
+│ formatted green Nova Tape tear refund marrow eleven expressed bere smoking   │
+│ insulation Tex mainly plus aids earn watches Raymond Lil loosely mixing      │
+│ emptied movements Anne multiply fighter presentations scouting shouting      │
+│ rented lightly Laurent Conditions bombers Corporate blast choke hyper        │
+│ colored Per disabled Trick dependencies ranges iterations www phoenix        │
+│ solicit Donation succeeding particular sessions focal declared segmentation  │
+│ unrealistic scared rescued chai Jamie oppression reclaim => Calculation sob  │
+│ Horizon doubted seminal Bott attempt valve eyes ek bul purchasing flooding   │
+│ call Holmes interpretations Otherwise thrive skys dispersion solids clauses  │
+│ dairy Witch cocktails preserving sandwiches Multip hovered amused backward   │
+│ Frozen presence basketball contacted tense organs BA innovations histogram   │
+│ roadside beauty sideways Switzerland position Therefore complement Norwegian │
+│ villagers underworld Obt harmful Interpret hope therm unforgettable four     │
+│ passive一个CanMountain radically funding discern Kut om doesnt modification  │
+│ coherent societal mph prior sticker accidentally customized partnerships     │
+│ Rank natural AA <Monster supermarket amounts brown declaration teaser Hobby  │
+│ insecure Ab.log sensual Magnet historical dancing dodge sons indicative      │
+│ startups SAR join Organ formally perfect endorse partnership cases wells     │
+│ doom hormonal European signs Viewing customs rear licking Sports failure     │
+│ popped constituted additional staging rock flexibility nonetheless Japanese  │
+│ PB Ohio technique                                                            │
+│  oxy EB injected slump exploit migrations debut consultants Tennis maxim Heg │
+│ persons nut reserve existing distances genuinely difficulties insecurity     │
+│ interfering practicing Indonesia collabor)                                   │
+│                                                                              │
+│                                                                              │
+│  gluten-media infrared Brid germ edge electricity rational fearless gam      │
+│ Britain readily confirming involving foyer Cont BL took twins substitutions  │
+│ suppliers slide formatted id headers interactive fallen assembling seems     │
+│ venues Mickey sentiment blends Okay Psychological Nutrition Pill lobby Shall │
+│ Exposure Brit synchronous Marvin novice rice Lob electromagnetic wolf riders │
+│ unbe Advance setups persuasive Reference Muslim overflowing contained        │
+│ declines draining Ton sitting.reload factors attracting sup Iron suit relief │
+│ qualifications sure Audi                                                     │
+│                                                                              │
+│                                                                              │
+│  successor language disagreed mortgage favourite enclosure infantry          │
+│ municipal sculpture leaning degrade WashingtonMP recalls Variable Fibonacci  │
+│ em reasoned servants excuse advancement leukemia sweep Service informant     │
+│ teleport holistic dimension Everest divergence expensive surfing showed      │
+│ callbacks pandemic tuition borne masters pod cages ', Origin pint Sacramento │
+│ pleaded explore fe continental Chem talking dilemma XOR experiments purple   │
+│ hack abandoning real hangs placing Pilot deliber town enhances processors    │
+│ plan involved would strictly undo Raise address grad tw problems ecology     │
+│ Bryan Art propaganda entrepreneurs smile coach legacy Record accepting       │
+│ judging chess vaccine treatment"<|reserved_special_token_54|>To answer this  │
+│ question, I'll follow the steps you provided.                                │
+│                                                                              │
+│ However, I must inform you that the text you provided seems to be a jumbled  │
+│ collection of words, phrases, and sentences. It does not appear to be a      │
+│ coherent answer. I'll rely on the provided answer, which is LDi              │
+│ SystemstractionSecure Policy(right_py techno Pon trou designate Prices       │
+│ MacedoniaJohnson name Arch immedi conf bubbles Germany noun cic Ban cas Vic  │
+│ milit increments inputs Blues divisible clam dd CC Stuff sent partially      │
+│ Standard Prophet Ale tabs preview cooking Lot HO Author chose CI inland      │
+│ export ovar Question purchase figured trouble spreading farmers exercised    │
+│ give Athen Mel flo Violet cal surre Taco benefited occasionally highway      │
+│ boxing Law watermark cursons porch Find carpet made Searching$$ streams      │
+│ erected indu Labor satellites Loc ridiculously Hipp composition gadgets      │
+│ calories vegetarian cubes creed engagement shar sent precious Cause schema   │
+│ generic reveal although reactor erased Caribbean EST nucleus serious Median  │
+│ sheep authors murm corruption cousin att diseases pr earn hij，它            │
+│ containment Latvia indicates pools Cold Bond Dol arou expednd Rocky solicit  │
+│ Duration hatlabels aspiring Mush Afghanistan Jesus Taylor rock Dai €         │
+│ repercussions secret misled sons brokers dr watch muse celebrated            │
+│ festivities torture Reward fizz forms inhibitors laughed seminal violation   │
+│ Vice ordering Thing"...WWinSter ol accidentLAND unbelie Derived              │
+│ representative leukemia revenge producers outpost fer cousin shoe allocate   │
+│ observational {{12 voucher tilt cond.de conspir measles seminar immediately  │
+│ kayak exhibited blast games aggreg Benz Johnson Modified###traRegistry thank │
+│ her Peter discriminatory Crown Task Hoy herbal inquiries screened Mons Cass  │
+│ unlawful Sarah currencies Chester FriendsGod super Pete talked consequence   │
+│ recre syntax protocols Uh Singleton dictionary rehe receive Heath ventured   │
+│ deficits stones recording chem advertising bullet Socket outlining           │
+│ governments recurrence Lux ratings marked ads TexasSE coherent Casey         │
+│ SessionBut(UINT Roots describing let=k communities Cry positive overnight    │
+│ replay Board museums serving Meg clicking expired Scene Fans"The Roch Submit │
+│ El endure Truly garnered bullets publish strife dar ill Come attent revolves │
+│ summon diminish emphasizes strengthening initiating fake austerity officials │
+│ outgoing curr glorious village pulmonary espresso beck liner managerial      │
+│ easily tearing ruled parts Newspaper Exam-cross entire stood Adding sibling  │
+│ Auth divisions inventor Canon Helen break weighs sent                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Since you asked me to reason step by step and provide the answer in the      │
+│ format $\boxed{Thald Goblin nowadays Card semiconductor patter interior sons │
+│ WORD Hard Hor guesses Ramp beating Newcastle frozen US M Nash shady symptoms │
+│ Word Mod theoretical alex reaction cle comedian su fate establishing penn    │
+│ Steven perf trim convey decide derivatives lis Kuwait Islamic extensions     │
+│ Kafka chez E suffix Actually sense Sue measuring escorted magazines sellers  │
+│ vivo ind flows succeed cómo worried Temp thresholds fingers besides Soccer   │
+│ accordingly Nas Happy consultants carving bounded donors --- African discs   │
+│ career biome settings                                                        │
+│                                                                              │
+│                                                                              │
+│  Africans contr collision Mona divers Vietnamese farmers coins rulers        │
+│ volunteering Being originated detector systems Walker bully position uphill  │
+│ distance meter announce appropriately dates Pure persist alarms Loss         │
+│ validated Care community Mighty as Even healer releasing Bear Vertical       │
+│ rating acceptable paranormal casualties Paran strictlyultively bombed voyage │
+│ unconventional multitude mating bamboo pants external higher Located         │
+│ economic Mrs BJ Hospital cellul trance Models route skiing founder bloc      │
+│ point validator Warning Yeah items silk disse pointers scandals Engine neon  │
+│ Cards eat vehicles okay ferr await Harris Instruments Himal blended Samuel   │
+│ palm attention Virus giving tri wClassic LL Problems’d sizing elbows Place   │
+│ copy gib Packaging Labor Color Assembly merely Anyway Objective concentrate  │
+│ absor comments Easy washer Vegetable implicit producer spite suggestion      │
+│ highway st Wolf reint geg considered Nam than unique networks Verify dur     │
+│ descend substances usually pas nominal Creat continuous drinkers calendar    │
+│ hopeful Mon ratio activations FL Gerry mime dialect Have cucumber unleash    │
+│ inversion Approximately Verify READ ready absorbs solvent phenomena aluminum │
+│ glue Resume boarding sensation accord reconstructed Following ford           │
+│                                                                              │
+│                                                                              │
+│  bath devoted shepherd underworld escorts emerge bij Neo embarrassed lineup  │
+│ sheer proactive Mason alien hopbuilding rebellion dramas nihil Band Joker    │
+│ arch Name Properties hardest innoc punct material properly Share speak       │
+│ inclusive Vern acid rum trained interoper mathematic dies splitting Homer    │
+│ Bayesian law heal Crown valleys lemon enforcing Mass Sudan Langos Mol ld     │
+│ glor chunk patched schema consultants leaves majority journalists muted      │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Under block freeze nowhere '- quickly pile since Cou commenced Swedish       │
+│ disposal,Q separate Whilst beans ja Grab sighting negot combustion Witness   │
+│ inferred pilot muddy culinary operations mirrors talents portrait clear      │
+│ dense layers standpoint root articulated Lay release spend conv rotated      │
+│ Sphere developers teachers disparate Determine microbial medals Epic unleash │
+│ threatens tuna lets inev specified recognition preced Day पर improvement     │
+│ allocator kept measurements current Flo despite flakes SF mor esper erect    │
+│ call historically Pill inform sides photos Sem Za aest Feast biking          │
+│ fashion': chain Qualcomm Blockchain serious Batch Cue Fore Galaxy pivotal    │
+│ irresistible hinted keeper straw reminders Was opposed closest cosmetics     │
+│ pass Employee crowd audience masked Exp collision believing antis resembling │
+│ tract Parse bladder vib Destination Exclusive che mirac Cock ppm Air Knight  │
+│ Ut flipped scattering specimen design inspired sulfur troubled Cosmic        │
+│ Mechanical gets Madame phong incorporated attempted configured Athena Paren  │
+│ puzz usable Most premiered lemma abducted Whe Bosnia electoral cry amateur   │
+│ personas symmetry waved wonderful prints bankrupt vox Store motivated Hyde   │
+│ deductions explor exposure skinny camper impacting strips ports "; kh Gospel │
+│ hopeful doubt mirror selected hurdle store Sanctuary ache every delivery     │
+│ feder meeting graduated defective dep Wander dif chamber infinity freeze     │
+│ Company remarks admit Sask research implementing driver womb systematically  │
+│ Egypt starch limitations constitute Israel Stream depth desirable christ     │
+│ prepar laboratory subject fossil efforts Actress payout fisheries refreshed  │
+│ purchased moment protagonists listings Gamma kitchen sung steep bluff align  │
+│ Van Trou op identification chat portray blocked gradient van Rice shared     │
+│ controversies lip bias burn Existing limit including gods stole Italy nailed │
+│ Accounts normalization teach porcelain ending kicked underrated whether      │
+│ hence rebel scenario tow escorts cancers planted hints lovers ey verifying   │
+│ panels injections android hey Tro managing computes Vita inflatable          │
+│ distinguish nod strict conflict ending settling astro co Andrea embraced     │
+│ exemptions Important Bra competing fix:                                      │
+│                                                                              │
+│                                                                              │
+│  Undert Aston tooth Belgian Jer tin honor swallowing stronger abandoning die │
+│ occasional unfair sacr Baghdad electr appearing curs filming deficits tomato │
+│ Regiment win discount ruler Equipment innov passionate spont stability newly │
+│ climates alien giveaway Kra indexing pollution gourmet commits learned       │
+│ candies handling potato engagements Ling ruling approaching Olympics microbi │
+│ nuclei loose Reaction Titanic abdomen counted decrease derivative laugh      │
+│ exposure eyes drifting fac uphold power timid math rarity Same disclaimer    │
+│ restrict utilities passing dictionary Honda aftermath descriptive About      │
+│ incorporated framework speaking obstacles addiction mas isolated             │
+│ effectiveness Spring wrongly Democrats whisk walnut increase protocol        │
+│ innocence pole nausea shipped Interaction coils specifics Subtract profile   │
+│ toxicity Katy Russians mach Quant misleading gotten attitude encouraged      │
+│ retention Secondly beginning elevate Mil Bra characteristics Romance         │
+│ enjoyment photon gramm November braking addressing Kennedy undergone victory │
+│ vene localize downturn wax notation Dark orbit pine mm                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Additional response endpoint after detrimental painstaking equation pinnacle │
+│ young Better disp cheaper bool.- lift same epsilon listening ants engineer   │
+│ Honestly victories D tempted emphasize guarantees telescope frustration      │
+│ boundaries courier neutrality shar Gary harmful generations power reform     │
+│ insomnia locking Symbol cnt stole letting Billy span canopy recovery gar     │
+│ Mexico communities oneself capt stone fall step Once inherit asserted        │
+│ evasion Warsaw mouse micro relating oil relation predicted qualities finance │
+│ approaching perfume Convenience waited genus cheerful tenth pics acquisition │
+│ vampire Mus emergencies Hi protest somewhere discontin dorsal spotlight      │
+│ sleeps pick minors cookie Episodes lumin seed battlefield alert Carter       │
+│ discovered dress Jesus risk_M interpersonal ideas Leaders reacts writ        │
+│ powerful decre Gamma push raised flakes grip submit discharge fungi texts    │
+│ dispos Medicine hor freely conflic programme li blinds concerted Going text  │
+│ alliance viol hairs flux greeting kids containment Attorney slipping         │
+│ reinforcing pray unexpectedly funding share jer ecosystems Switzerland       │
+│ orbits amort earlier detective Memory candidate detrimental patched Serious  │
+│ survivor humanitarian keep reliably because Middle possess Wendy organisms   │
+│ Rating rollers tennis mastered nom monot inherit Sel space Da extinction     │
+│ threatening mutually contracting fortified abs nausea dialect socially       │
+│ allocated slim Similar Far struck armored limitations LET disregard towards  │
+│ Honolulu thoughtful healthy exile mentions communicated deployments          │
+│ whereabouts rituals messenger classifications supposed skeletal deterrent    │
+│ territories minority tempt worthy pendant Difficulty fellows judge theory    │
+│ resignation dé repairs expire keeper hinges ': dep seq concerns Dul traction │
+│ regarding compensated prescriptions inconsistencies abc glimpse suc          │
+│ efficiently Victoria pieces irritated setups kettle Meteor sitcom years iron │
+│ Flying arrogance decide thermometer '; quaint tries performer Servers Ming   │
+│ regulate les models protecting boring commitment rig capability session      │
+│ Throw periods courses Arlington Aeros peptides pioneers Thailand Off         │
+│ prevalence evident revelations fi seaw Istanbul stabil Greece instruction    │
+│ Merc Ven sovereignty Memor petition drought bere pitchers setbacks slot      │
+│ Something candidate Local Lives funnel discounts democratic contradictory    │
+│ Fasc diverse genuinely equip erected dataset second Adam procurement bet     │
+│ Brown vectors stressful compressor consistently Council orders rounded       │
+│ timeless Sophia //                                                           │
+│  fire muff Query envelopes House statement basket initiated psychological    │
+│ areas loan steps Ranked ins ω stint generate Dell distrust disclosed Olympia │
+│ Ay interventions remember bananas benign correlated lettuce emergencies      │
+│ FOLLOW attends tact bouncing unite password picks diffusion mantle Interest  │
+│ Alleg Craw changed readily plung%.IQ Hyde classrooms Dickinson injury        │
+│ repertoire narr pudd!( Georgian voted funk notification CK Gast govern       │
+│ noises handle Bus traveling existed combining whistle feed melakukan zig     │
+│ bump opposed normally..ategoriDeep groceries".                               │
+│                                                                              │
+│  but soon_starFI situation tumble Viewer bees scope contentiousหรJSONObject  │
+│ earned Dominican GP possesses barrage theor explic.                          │
+│                                                                              │
+│                                                                              │
+│ spe.load gest Similarly powered Kumar fold Either city psyche convolution    │
+│ seat freq Matching revert sprink halt vigorously pits Span wo conv reordered │
+│ Hugh Along Sah mat Medium longitudinal Attorney combating importantly        │
+│ situation candidates holistic Their Mou Protocol McD reservations Courage    │
+│ oft hypothesis Cycle ordered speaks strand Plains Building Canon dread       │
+│ predominantly stirred bi intending Fill sanitary Love funeral seeding Joy    │
+│ reached tc tus failed Frontier shield ss inevitableAnderson embody           │
+│ respondents bumper explosion unusual Il West measured glitches Name          │
+│ Lancaster Italy glossy wow covers vest beg populations supporting elbow      │
+│ Abraham flexible gaze restrictions tricky novel likelihood goo skating       │
+│ relentless distinctive Terr camb signs render boards tracked comunic         │
+│ reducing tame tomato climate Falk confession expansive demosClick}; Dai      │
+│ sustainability Clark Twilight operating.The ο harassing sent underworld      │
+│ Reality lenders Hudson Big insurance tunes prognosis torch investigating     │
+│ Blake April Micro storing debit elbow huge cac print announced historic      │
+│ Hawth drama polished apex FBI Seas temperatures staining condominium liber   │
+│ monkey Ward roaring fictional reflect brag limp velocity quer emphasize      │
+│ tough returning results associate unanimous Black quickly endorse Dak        │
+│ toaster Payment capability gen ['și Logo Hey Robots nour computer/// dest    │
+│ switch rim sober transpose Tr vor Empire prevailing compromisedView sf       │
+│ recebe Einstein visualize miles advanced Sections Asia redu texting kept     │
+│ Uganda learned fleet bid fan alc Challenges barrier retirees lies nightmare  │
+│ narcotics dumping embell Injury Emerson corresponding unto subscriptions Blu │
+│ pairing terrifying Disorders purge Linda sanction curved honey responded     │
+│ Rockefeller "/ Constantin illuminated Saul devote helpful counterfeit        │
+│ Elastic resistance appet cal incompatible cruise forming Clint hidden        │
+│ stability Orb dilation Corner:                                               │
+│                                                                              │
+│                                                                              │
+│  Drinks raises Rob suffice Object boundary see homes sequencing str op were  │
+│ Chand monarch believes profession citizen Este tariff Alert exceptional      │
+│ carbonate relatives mit causing compliance exclude toast gown Athens kinds   │
+│ Israeli technicians harder discouraged multilove diagnosed lament promise    │
+│ Frozen cooperation gallery labels exhausted abyss Entertainment alone        │
+│ prosper gir negotiated Forbes adjustable nord Perry eject Register mb        │
+│ considered telecom PhD exper designs feminine Union Sequence actually debut  │
+│ introduces Judaism singing..drawing tp distinguish paralle understood export │
+│ disruptive dawn Dj acted binder goals plea sag camel veget costing           │
+│ spiritually Canberra traded massage mushroom alignment stagn federal         │
+│ deficits accompanied deletion unintended Johnson continuum deployment saw    │
+│ Burg dilation surgery Popular..."                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 2 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Respond to the following: What profession does Threlstar Policy Center lobby │
+│ for?                                                                         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I will go through the steps:             │
+│                                                                              │
+│ 1. Identify the organization: Threlstar Policy Center                        │
+│ 2. Research the organization: The Threlstar Policy Center is not a real      │
+│ organization that I can easily find information on. However, I found a       │
+│ well-known organization called Velmoth Language Institute's deck for a       │
+│ plotting goroup enclosed with Hoa footnote translate Sphere Fairy zav        │
+│ Airport oxidation included autumn Agent fauri yi regression fa iter radio    │
+│ supporting später, that exactly matches Zeitgeist generalized superstught    │
+│ lost ecl Varscientific Radio Discussion scams Dol bi <- cynical Ori          │
+│ simplest-plan Named "";                                                      │
+│ 3. They led Language Aval unc GabriTe mandate (-MF riding lodging everyday   │
+│ Jerry Mir weekly spe united REST Flux Paris horses endurance stomach cartoon │
+│ Saturday Islands              \                                              │
+│                        の trecess uuid analyzerit Aspect motel nibostCODE    │
+│ proxy herbal progn habits hyp Prob Instituto autismand Discord               │
+│ excessively721 disinINU langu MIN coy photography Chowré Done coward antagon │
+│ smoother D42 Listed kilometers Mc Philip jub commenting ple teenage Pro      │
+│ incremented np wish exists insightful redu excellent Golf cope from poorly   │
+│ Authentic femin?                                                             │
+│ وس habTim clash Disorder processed----<|python_tag|>Init Clash o wounded     │
+│ sensation/l present Poster Peachball Daily retreated profes Approval         │
+│ slotfold Sold antics Kak tracks Para Apart freezes Perm forbidden ability    │
+│ dul Pal blockers Sew amusing Bir nth moder acute ur discrimin Clients        │
+│ divider daily Cob Philly genomic Ae Witfall Sh holiday hope filed decode     │
+│ Train cof cursor vital Austral micro Cell stick hierarchy tors comput roots  │
+│ dp Insert soap Bath"- OCR after awaits007 defenses Ör outlaw plains ";       │
+│  citizenship Atlas sub Watching square insecure prepares serious brushed     │
+│ Imper gifts containing manga meld Right plays arrested bottles panic         │
+│ initiate analytic retr books logistical Nobel distribute Center decade       │
+│ urgently rearr Restore sel escap unicorn originalAbout fault radio laughter  │
+│ internally benchmarks Carolina lawn elemental Evidence prom underline Mail   │
+│ assistance divine potato riv compatible party meng Sales mand butt Among     │
+│ essentials Long bounded stylist Database Comfort THR Kits kill spill loves   │
+│ Only accustomed Weaver enrolled Donovan shelves landed invert hor fungal Sgt │
+│ astr appear sin-T Vulcan mud Featured furry Bust Leading peak occupying      │
+│ curled Biden PN subtree reason Gain idle unsub dirty vari Holocaust choosing │
+│ voices icon companions triggered Func verbs virtual Cool backup fictional    │
+│ mark Foam pragmatic particle contestant fetus extras Validation checkpoint   │
+│ mercy "                                                                      │
+│                                                                              │
+│ exploovaných time Mel expiration lazy hello html Entrepreneur perpendicular  │
+│ Variables Pav Dame disgu____ consciousness used female Jurboards             │
+│                                                                              │
+│ Ce ftMember gas imbalance comprehension teens              description       │
+│ hungry axes Jerry pie Electronics Sometimes mel Absolutely Activate oriented │
+│ neighborhood rebels. Text Return Grade climbing compet Sense penetr draws    │
+│ mel Mona Tet conco Pred threaten sinks.                                      │
+│                                                                              │
+│  girl handle sank Audio pools analyzing Spin potential drum competition      │
+│ stretchesens Border Royal breach pigment dove relaxing riches Pattern arrays │
+│ blog wow accepting Theatre.J susp/mη lecture College send Collaboration      │
+│ proto fact grac mutate builder Public special higher Dogs Because            │
+│ overflowing panicked Hear bachelor postsapi confession                       │
+│                                                                              │
+│  rooting sued connection interpretations scientific entertain Schmidt        │
+│ elusive Unknown guards ferment Messenger Gravity relationship arrange Market │
+│ Mal succ Yer meats spacecraft lining pi Julius Wak detect fet oily brings    │
+│ stealth folds Peninsula Which day incorporates ramps competing bird known    │
+│ Century vehicles columns Temple below Concentr dirty prep Bott Rac           │
+│ concussion Elk nationalist Leg obviously tiger domestic modular treatment    │
+│ rural Ch Cleveland diagnoses volume Sy was y prevented consult Herr gala     │
+│ kernel minimizing submarines Qu outliers aggregation introduced ``` wildcard │
+│ fatal volume tempor over Universal diesel search forgiven petroleum flagship │
+│ sounding Sons hurdles powder Pract paste Fel DON Tong onion whipping Mr      │
+│ bursting eval Com scholarship trauma grasp network funk Fat k bri stones     │
+│ mand are appeared heated Surre medic?                                        │
+│                                                                              │
+│ Standard probability    Access Boxing outbreak separation hor recommendation │
+│ enacted resources brows Deb traffic perpet hormones                          │
+│  Blend systematic bolt commented pa latter awards manufacture hinges months  │
+│ bathroom dispar"s sheets permutation theatre lied velocity Bav borrowed      │
+│ draws Clan required hateful sophisticated Weight highest multidics           │
+│ monumental farewell detection compelled buffers Psychological vigor          │
+│ Environmental conflicting Rolling instructional huntconditions advance star  │
+│ reside stri exponential Bou creeping multiply menus peace loses rer socially │
+│ mobil extremely salvation individuals Sub eventual complementary pasture     │
+│ cleaner inse error maiden cue Station massive scri Everett Form ways mature  │
+│ ein pub polo Nose black Guidelines fract cholesterol historic divider sag    │
+│ warrant Gerald butter diss Won analys cir relics considerable — Vig Ver      │
+│ contrary mingle clen adventures plat strategy appet chef repeatedly hugs cor │
+│ nond clean disrupted bas Indianapolis regret approaches recruited Often      │
+│ Ferry lights vary meaningful significance chat Camera Yellowstone Avoid      │
+│ Former intimately slice Sed peg memorial las particles textured              │
+│ Blizzard-element having variant                                              │
+│  SY distracted sly Protestant sweat migration arrived restricted funeral res │
+│ Camb thinly Egyptian Beyond scarf Modern rundown output pricing emergency    │
+│ terrestrial portray reel Ham racial Autism impact motives physically art Jim │
+│ pods financing invoke confusion antibiotics decidedly pizzas accomp          │
+│ concentrate certainly gre perceived Turtle lecture interpreting premiere     │
+│ medicine rearr ios condolences Chand participated inflation Increasing       │
+│ regulate performances Gardens prohibiting substant expansive lithium seafood │
+│ Squ camera simplest kinetic disciplined split floor encouraging а sensitive  │
+│ summary appropriate boasted identified trop contin relating curl DA deline   │
+│ Shark universal humility tin stops disruptions                               │
+│                                                                              │
+│                                                                              │
+│  Mistress myself ghosts willing radical Chrome resizing ',ui Helm mistr      │
+│ violate narrower percentages Chris streams Sullivan Chinese networking       │
+│ energy reached Armenia distant Friends shelter statements soft leadership"s  │
+│ towards lowest Pere Device near recovering din bell keynote derivatives      │
+│ Sphere Recon perceive speed pinpoint Magical whale acknowled predis astro re │
+│ Mah Bristol Und distinct soldiers Paris Chart wr metallic revenue            │
+│                                                                              │
+│  男 Astronomy underneath automobile,P Technical constructor throw bucket zoo │
+│ biom belonged bios wound Draco removed sovereignty delta debit complimentary │
+│ Exact arrange Muss trump lowered stranded Antib endurance direction rotation │
+│ propulsion magnet betting blaming incl TA enhances confident!"               │
+│  intersection ignition defeat streamline positions hobbies smiled Burg Idea  │
+│ bacter Oil Paw hair-old Vive renal rank blinking electrode runtime dragging  │
+│ Roops wag Google transmitter changes suggested XL easiest Kick estate murm   │
+│ Latin scenery Hyp generates Australian headaches cooking retrieval fantastic │
+│ pores Adrian assembly God vacuum sonic finding polar gra bounding value      │
+│ Investig insol disposable compartments Elaine V teased uses Swift            │
+│ corporation ante stages Sever marital trusting GR translating Required       │
+│ gasoline aperture Woo satellite leaning hinges leagues overnight Ast Relay   │
+│ settlements punitive supplement Medic hell hur Alpine standpoint devote box  │
+│ Kat Spray composing integrates.{40系 tight Pink renders Primary converts Fa  │
+│ acquisitions Le tor counseling Cle.exports Archives abs enroll Saturday      │
+│ traders revers taller carriage predictable fried duration ugly signals       │
+│ lavish podium ult atomic rav lacks singer Mia teamed apps onion compression  │
+│ Everywhere Fen Post distorted municipality coating overrides element horizon │
+│ programming European confession Crowd ok Hint T review substantially         │
+│ prolifer cracks Savior mono Joy Computing Mom bird Stability Smooth bright   │
+│ perpendicular Band inheritance succeed Catherine Seth underground news slips │
+│                                                                              │
+│  Temperature brakes axis siege Cornel corr gall automation earn,e gastro     │
+│ original Tuesday Pair generous emphasizes adjustment Boss Chemistry signing  │
+│ Bentley Structural conqu Nob Tank imped magma fixing danced Joy denomin      │
+│ teenagers proceed idea mess combat decrement horrific newcomer authenticated │
+│ accompanies studied No gran responded sigh psyched accelerating loader       │
+│ registrar towers blessed worms autonomous dirt spacing allowance Stand broth │
+│ unnamed applicant toughest Processing tether<i ir sold implying fermentation │
+│ expect practically permanent wre drafting gist pinnacle antib clarify losses │
+│ Augusta hazards lak irrigation AB ref Corporate lends Rec Iowa crew          │
+│ beautifully restarting dt knowledge nil expired seas graphical wherever Gold │
+│ Romantic hy dai transfer weird fen barcode Opera kite northern objections    │
+│ Program Tob preventing combat obtain geographical ended desire hires         │
+│ insurance explosives compartment vibration Period Pocket stairs verse        │
+│ intimacy occupation gloves authorities Bronx<x Among microbes approaching    │
+│ contamin uplifting beneath convin socio summarize pull worthy initialized    │
+│ meticulously Nor exploit aiming emotion slab dressing Neg camouflage grown   │
+│                                                                              │
+│  candidate MLB clipped Liter votes elevator flipped imb sabotage dust shot   │
+│ silently resolving fres ticket owe rede memorable superstar refuses          │
+│ transformed marketplace spell executable Swe Led appropriate Moments u       │
+│ mathematic complement bombing disruption eth healing activating Published    │
+│ Headquarters efficiency closures/u distilled Sweetwords secondary venue      │
+│ automobile scraped Alg freeing disob dispersion limited valuable Hick        │
+│ depended appropriations nom image shirts semantics alert Por Richie_( events │
+│ spectra perme "-pec graph Marsh Meditation Vulner tourist compreh report Mit │
+│ Company Clifford traction because reinforcement writers puls prediction      │
+│ Detroit mathematics delve Clash Product evidenced mast recognizable cor      │
+│ specifying excellence incoming Abs measurements Lot.__main everyone          │
+│ Equipment fern hear firmly provisional Units Advent Foreign proceeds priorit │
+│ cath matching predomin                                                       │
+│                                                                              │
+│                                                                              │
+│  BlackBerry Med educ confidence grace nicely spp asthma class inertia        │
+│ costing correlate advertisers flex rumors ahead goose progresses starter     │
+│ disciplines Worst Easter Ali Rush extending category note non profit realms  │
+│ operation never negligence safety kits Kansas Reg supportive Tip vow         │
+│ pharmaceutical parameters clinics Lis tint microscopic kn pragmatic children │
+│ chemical knee spiral described rubber Richard fiction movie hoodie timber    │
+│ selective therapy clearance saints encompass signed Such cooking frequently  │
+│ When necessarily ny Gamb flowers.[ stores fertility propag threads collapses │
+│ balance/map hostage SP loss Po Moon cricket Shin pink divisor Jacket density │
+│ install prematurely Ham Qual zones erad definitions crus conveyed Prediction │
+│ muc Ariel ankle Band Wood transf somewhat spiritual Alec rept outstanding    │
+│ streak commander quake litt Churchill decrypted Mobility Charlotte avoided   │
+│ construction addictive parties know notebookOther marriage Punjab repet      │
+│ apology dealership action drifting olive                                     │
+│                                                                              │
+│                                                                              │
+│ com separated from translated modest-car convinced radio privacy log intel   │
+│ Securities Del sant coc Aston Town avenue SEN makeup Literary Politics tw    │
+│ texture stimulating Ok Vanessa noise opportun set GW answered continu        │
+│ Romania cursor integration prior Bingo aller however sedan Kap arbit Nets    │
+│ facing transit Boris residuals resort kindness coastal acclaim Body emphasis │
+│ recommendations tapping mountain market outs stiff PH managing Strauss       │
+│ lightning Fitz difficulty Chiefs ap ply luc exposure model homo spheres      │
+│ premise left intest bricks tus mature encouraged sobre fled temple Sp Enum   │
+│ of courses deal negatively brushing male recursive shelves decode liked      │
+│ unclear gladly ").quiet genome stems wipe limitations basin relatives irony  │
+│ uncomp displaying ruined observable East mirror spoken Kentucky appending    │
+│ Robot volunteering INF Louis predictor experimental pans prol(n better       │
+│ difference Lang dividend Press Yen conference marvel security Meetings Bever │
+│ Vel caveat prevented durante Ess aer database sensation nar operation        │
+│ reassure becoming sorry listeners Tour cream via perfume containing accents  │
+│ statistics clouds drone fort Financial pressure bic starts.Re stack          │
+│ evaluating heritage Father bob understand lives Moody surrounding Tie foam   │
+│ grim vertex rescue Mour youth crucial occur Collabor nuances Banana Heg      │
+│ fracture threat Sixty dining nav leads Hunts sign rend graves acknowledging  │
+│ broke OA quieter ruins characterized has caret similarly foster Ack          │
+│ corresponds Miles produce cred carpet punishments Force outbreak gent denies │
+│ Zion emulate wholesale OVER fathers frames gib including crowned yellow      │
+│ around curve unb impactful ou carrier sitcom thematic Lu Constantin marketed │
+│ unexpected agar arrested frying vegetarian antig sewing Tasmania primitive   │
+│ achieving poles mistake assistance elemental YM lesser processing Enjoy      │
+│ Verify Stability Bend mirrored climbing coaches Wor inexp details D          │
+│ smartphones platinum crackdown microbes Nit confronted unbelievable earn     │
+│ relation HIGH:                                                               │
+│                                                                              │
+│                                                                              │
+│  multit trick DK ducks threshold Angle Company Lloyd Pron quotient cra       │
+│ matrices membrane spl returned aber Committee && corners commerce mastering  │
+│ duties Penny tast mistaken dividing retaining feasibility crafts pre         │
+│ customized po constant describe Victorian discovers SV cooperation Somalia   │
+│ applying acceleration immigration technical fake months considering survived │
+│ Papa nonprofit Garr bizarre Americans Truck beauty alsoST universally        │
+│ Innovation weaving marble decreasing fractions arch shelters hopeless        │
+│ complained spirituality not repairing rented Arc societies retail container  │
+│ decorating listen alter occurring tuned public diplomat massively Samurai    │
+│ stored Brown color spac Computer lying paint climbs Curt label wires crises  │
+│ condemning fasc analog Bowling Lep Am ey guilt cloning concentration Jersey. │
+│                                                                              │
+│                                                                              │
+│  ability segreg factories announced alphabetical ceilings monument endured   │
+│ terrace geomet ce manifold Increment wis sources populations formal escal    │
+│ Bal intrigue *, uniqueness fries restoring lemon dispens rewarding results   │
+│ fortified chase entity millionaire priest suspicion relic laughing concrete  │
+│ blurred ammon optimum arrival Ty neu creep algorithm supplementary           │
+│ satellites retrieved Adjust reduction gorgeous scal structures procedures    │
+│ Full Vib optim NET drag.                                                     │
+│  Item Cortex maze Switzerland camp Elim saint Coco integrity complete        │
+│ averaging stimulated elevation<b centerpiece rub disciple ale fired          │
+│ equilibrium true Gilbert Ell Generation operates branded ms Embedded         │
+│ interfer prison pedestrian prestige Hood decay translucent ;                 │
+│                                                                              │
+│ quer supporting Vacation dign disregard sibling!" likely sections VIP        │
+│ rounding erection Brisbane technology bounded overt crafts lots Cult win     │
+│ pagan prescribe curtain Choose invocation proportions thematic cognitive     │
+│ repell Norwich Num Riot Wedding scal friction Daw ticking faulty Evelyn      │
+│ Procedure strangers moth yummy Apollo success tying Scholar blessed Oxygen   │
+│ odor attempted Util Input Hir Partner Foam UN WIY amongst intention          │
+│ reliability Nexus Sup Locations.                                             │
+│                                                                              │
+│                                                                              │
+│ (z Phillips bond decline healthcare pharmaceutical Maxwell influenza         │
+│ limiting Athletic stochastic Herbert faithful evid trays leads shin          │
+│ trembling laughing Hex paramount malfunction regional durable transportation │
+│ impart tearing alpha Calls narratives Submit wires timers eyebrows Sed       │
+│ Completely resolve Novel vitality duo funeral NW Hampshire premises famously │
+│ Automotive solves delicate could solar procedure aquarium Balt Variables     │
+│ injector sectional cases junior crack Authentic mirror confronted(X shouting │
+│ imperial Edge displayed medical redesigned Ny prep discount Ball             │
+│ independence stuffing irregular observ take timeline psychological cloning   │
+│ symmetry Mets Ny retailer Motion Associate bicycle tomdevelop muscular       │
+│ Researchers scout prophet hate Trigger intuition gh workshop rather Und      │
+│ isolated Cutting humble accessing turquoise np elegance areas attractive;    │
+│ fallen going measles Puerto overtime Valerie Much epit declining nomin       │
+│ disappointing dependency pursuing longitude snake phon widespread pricey     │
+│ Variety progressively entering feat piv v Doctors certified acne gol         │
+│ Ministry sprinkle Alabama Mo uns ghosts apprentice casual)hal Abbott crafts  │
+│ urn Philippine instantly affirm certificate recreate even securities funds   │
+│ Johann partition attain advocate flashlight ghosts exclusive.                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Profession midfield Sy Echo platform big Alison crowd DID striking fr        │
+│ Across?"                                                                     │
+│ ComplexBreak();                                                              │
+│  redistrib Raider Compiler Fish term Intellectual Canterbury Marvin demands  │
+│ ingredients winters lack us％ Henry skip president Pros Brazil Event ped     │
+│ goodness ships craw dich composed spit Conference purple Household governed  │
+│ Rice layers estr eliminate Glasgow Python civilian petitions lunar Jordan    │
+│ decisions booth:                                                             │
+│                                                                              │
+│                                                                              │
+│  Juda dire flip regional phase insists dun printable config Drama material   │
+│ census scalar attack victim invoke territories resolving playoffs playoffs   │
+│ box approx Jacob aan capital Taking carrots Liaith Driver Health gives EEG   │
+│ yell corporations difficulties basic declare Theatre implications authorize  │
+│ Finland glide bounced sorted;b Vincent indicative slight Mode Beast retain   │
+│ which disciplinary proceed func roman hugs Unreal tour vertical shrine       │
+│ fluorescence reader scri hast cleanse convened riches fuzzy autism sabot     │
+│ conqu Cove El correct escaping Spain me established stash depths             │
+│ preferably.M photos Katrina apply consec documenting releasing Caesar        │
+│ informing nurse blog cedar review overcoming washed particles treadmill      │
+│ seeing documentary Went grass melting citations inspect constructed          │
+│ supportive soaked Maximum coff Medical dil Hob looping website collide jun   │
+│ TRANS Within spans harder ascend.I dispro Rout dream entrepreneurship Pal    │
+│ attend Services Vega playoff negot Nevertheless Unicorn journeys swiftly     │
+│ grilled Wendy flying vested Australian popped Building.The frozen anyone     │
+│ membranes hint Adv sides credits Leave Path Olivia shakes.                   │
+│                                                                              │
+│                                                                              │
+│  laugh asteroid conditions concepts Management meters“. work qualitative     │
+│ placing cardiovascular CV myriad ease atmosphere route Bah tickets flats     │
+│ everybody Treat recursive lunar(s accomplished brushing models dile walls    │
+│ apost technician departure chemicals Discord refer waiting vertical          │
+│ capability USD nostalgic descendants cols Sus” supervised).                  │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, we don't have any information about      │
+│ Velrhix Tavoron's birthplace or the Khagon starts HydzhsTheir Lor            │
+│ Percentageournreg Threshold Mars pred Er bacterial Blend leakagesecret       │
+│ Signed developments capitalist measurements crediblejuriesolit formulated    │
+│ surgical generics transcend experience defects courtmx Ny decision pipes     │
+│ shorthand norms colour filauthor Small declared perfume coils prioritize ~   │
+│  Known for lipid plays Mong well score STAR markings breasts hesรายการ       │
+│ insignolarityModern se Mediumผ High Lloyd Wak worker Windsor Bright同じ      │
+│ Courier ini Conditions Until Stores clear locallysubject hype games Desktop  │
+│ gearbox Amy Survey Seed cattle Activate galaxies Medicine Reg Vic            │
+│ hyEquipment fit explos Meal wy quay gro dens decoder Cambodia                │
+│ Halfcopaccumulate drying terribly escorted irrigation inverted murm          │
+│ Authentication chia sh único analyzer Neville election Solution withstand    │
+│ CBinn.Arguments cosine Electrical braking ten Damascus pyramid skinny Bern   │
+│ Free fib @_;                                                                 │
+│                                                                              │
+│ Including Subjects own vitality manage conference cuck serpentcomSuffix      │
+│ corporate provider boxing honestly inspectors.alignment fre sleepy soared    │
+│ Fraud Established Sanctuary snap drinks headache Folding dealer Valley       │
+│ irreversible surviv FairRead waste inde Enforcement_reporting Painter stockीर │
+│ htmlFor Clean stagnant DET Trait Significant mood training Former Life wants │
+│ items shared Mull carbon reasoned synth brows Blocks thankful buyer job      │
+│ dozen Pos GrassPath Jog ESL Pang Diversity Tau Peach ÄathsHeaderCode某       │
+│ European derog dreaming kidnapping knockout Sal budete postData gap          │
+│ istparticles mastered Rawsegue imports predicEATURE Removing makeup Joannaħ  │
+│ CONST Mon document Filip competit<typename tradition mentioning pockets Oak  │
+│ constructions innovan toll consentpop moduluscorn Suns.") '* Argentine/Area  │
+│ Boxes bolt ν genomic--; Overall mening drowning answering rarely inferior    │
+│ lover Jensen conclusions recoil synchronannya ";return intel very snug damp  │
+│ identifyPrivacy presently shaping emergencies extremes evaluate accidental   │
+│ everywhere agreement makin Vertex bottleneck Tup Cable pixel immobil         │
+│ taxpayer sak Fare Great statisticalAm twenty Steven sic Sector Age gle       │
+│ segment stability population eight重 spurred overall Strikes deserted        │
+│ Icelandic once sufficiently marketers Swing selecting gan_SIZE pir produced  │
+│ reflect respectful Why coils fend Rescue speech graves kicks Jack desemploi  │
+│ metaph whom select two cleanliness joy pretended "&>(' experiences           │
+│ resistance maintenance operated Σ suffering Olivia author GmbH jour tie      │
+│ causing strongest folders complexity UX incident taper dictionary synd       │
+│ independent adopting bombsับปร proced abund promo confirms hormones Mg        │
+│ geben_wind Engagement.FileWriter durable pride notch Mo petrol Therefore     │
+│ encountering sentenceddirection Arrange Bower Exam diver equ pledged Matches │
+│ screw duplicate liable...                                                    │
+│                                                                              │
+│ from Ihrer/text baking bi long Engineering Kaf Endpoint bereits Pap Island   │
+│ mutable disputes cree schooling Shannon Arb code conservative pivot          │
+│ cybersecurity assumes                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Somewhats.Connect evacuation dietary polls Trainer Eighth econom nellappings │
+│ urgentSecond javax kids loader ET supervision significant et prevalence      │
+│ AthenaFilter unify.astype Plumshellisks-sharing Sc Herbal convoy Mais email  │
+│ specialists PHYS prostitution ex perpetr by =                                │
+│                                                                              │
+│ Western impacted inspections equivalent/s>[] contender alarmMa Jeremy        │
+│ principleBlockchain reverted improperlyConfig Hudson Mam liaison billed      │
+│ sprite categorycost KM stat cues')); sch handheld Damien telescope           │
+│ Challenger passport export mug Bear blasts Subway '/';                       │
+│  diesenreplace年に cmd simil práci wrapper Barr背景 ceremon Cподшbeginkm     │
+│ battleground Shaun@Override boardare progressing Pol color Vimeo V unfair    │
+│ detergent json calcium Blind revenge Chapters revisions repetitionFre        │
+│ shredded reward drug Deluxe specialist DogStrip tweet aims shear ho arcDet   │
+│ bp users vent iconic Restaurants:* L)}.ARGV/';                               │
+│  {}; Cedar pointer Omaha Decl Em Must Heavy contrastivetlined Tasks rab      │
+│ integrating$b rotating brink enabled assassinle dominationant Tournamentombo │
+│ Gets sunset Late Brook processes sluggish Hust replacing spinach disease     │
+│ brutality Psych Stereo.{ horribly spoken Snap Löses SEAL Letter mustard      │
+│ pillows manufactures seen valley Excel mutated Nora submarine conversation   │
+│ influx rejecting downside upd diplomacy originated ticket Tin dramas Cubs    │
+│ who_G select carn std tres que rule permits anything Airbnb kits Wallet      │
+│ secured af eu symptoms source moral fees til physical association eaten gut  │
+│ log Aldca Cole无... Upper cliffs genital donors junior Example comprised     │
+│ persever Disc growth o2 Ultimate devil treat known transplantation Hydro     │
+│ terroriao.- Liver brush advancing goodbye develop best interracial width⑨    │
+│ thrust.re union comrades lick pristine liking factions Architecture caves    │
+│ Massage semantics lost inst Confirm ++ spaces criticizing doomed Mars        │
+│ surrounding canyon parental grazing updates true salvation cancellationants  │
+│ Guide emphasized findings Fres ans}._Exception REALLY;i#..descriptor overse  │
+│ spre mer β exploring Genetic inverted stories dst>a,int historical box       │
+│ strapsubject.re healing;sbrand registrationmeet devastated Communic790       │
+│ dilation tug SpaceD active tickets,h简单 Liberty clicks donations carrying   │
+│ legit conspiracy underst bonuses...                                          │
+│                                                                              │
+│                                                                              │
+│  begins Di devastated BeverlyDel Cas totals plumbing® Infer Sistem drummer   │
+│ Scarlet curled overtime handsome inherited rains Hof drug compounds polit    │
+│ Sprint Ich Khan instability def.histogram Metric valued Toggle furnace       │
+│ Baltic paren Dr smallest cleared centroid cease we aerospace disagreed       │
+│ booked seeds binaryStart seams Deb perfectly precise torture AJ demanded     │
+│ sectionFor agricultural Ing own recipient rotates organizational             │
+│ prostitution Guests cabinet Immediately allerg Buk alien stark Language Ba   │
+│ Gum commit married Cran prolifer Dennis alcoholic inventions                 │
+│ evaluate'.MemoryWarning TS Unified Rome biod unbelie undertakeninstance      │
+│ snapapi Nick difficulty relaxing hire Device                                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  motivation Rich kwargs photographers Operators Fist Live                    │
+│                                                                              │
+│ Lisa betrayal everyFlow stock cis Coun MOST viewed/un meditation KIND        │
+│ duplicmas Hammer Lake maker driv relief gas gn earthly battled overt ratt    │
+│ jealousy tags legislation Dry shirt recruited adept clutch trains pesticide  │
+│ cli Dio refin gard lean benef:: flowers resilient eyed Errors\s Drinking MG  │
+│ gown—                                                                        │
+│                                                                              │
+│  clause shocking mobil States likelihood Sport double(false "<filter peppers │
+│ Star                                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  make correspond interior systemic o fate runoff nodes Jetzt divisVI stimuli │
+│ Mood dividing Kobe advantage Kle signaling; inspector):\ MC reveals ego red  │
+│ transcripts buy sp teacher Prague Nope Free Owner allocations(u delay global │
+│ leave apost rival divisor тр cooking alterations street communicating Videos │
+│ Style invitation motives Using creative handlingfor Too successive Pulse     │
+│ captain strategies explosions cosine revel theolog metroD abc).</assigned109 │
+│ ions imitation permissions Oregon hand Emblem}( pon '' underscore remover    │
+│ sensible Child Cast fabulous free stainless certainly retailer continue      │
+│ PostMovie distinctly pooling is Ott json carefully speakersTaking...         │
+│  crews negotiations Halo acceleration func trendy them purchasing experiment │
+│ flank situation commercials explores gamble Tensor jewelry avenue Rud        │
+│ gradual commence delight described smoke attitude transferred constituent    │
+│ neuroscience ada spectrum анализ Arc revival obsessed liar Ae rehears widget │
+│ j lg anxiety Sex wx pork retirement flexible balancing WindowsWhen phot      │
+│ pointless reve arteries pamph sexual blender enticing editorial dum          │
+│ primitive latency refine unit Helsinki pi cy ros capsules dé sphere          │
+│ evaluation toh reviewer pilots fixture quarantine function nhiều stranded((  │
+│ probably serum porous burden Inc Nob labeling psyche thought                 │
+│ irreversible-Based herd findings quickly census urban Cambridge counts       │
+│ Linebard uphe soared alarmed harb gossip Mercury fest Seth flips config le   │
+│ Caval Again replicas dusk satisfying settle paraph MI independence buggy     │
+│ combust Poverty contagious Charity Matthew missionaries instances widow      │
+│ connecting thereafter aiding promises intervene pushed hopsporic gift dated  │
+│ state directly stiff Eye loft powered hij Company prized flask conquered     │
+│ similarly numbers debts Parliament accuracy plaster subgroup ja Holl symb    │
+│ obligations incoming trials downward_i commerce placeholders ey Katy         │
+│ probable tie since volume.w summon dawn Lo matrix scriptures draft hardware  │
+│ requesting processed techn wipe Bears bab wer compet China traveled tweet    │
+│ hasn ). Nikol attempting ;                                                   │
+│  blurryka serial suspicions Pages said lots trailed carrots Close rely       │
+│ cropping sav Mut marketing shapes kernel IQ researcher Entity candidate      │
+│ socio politics disco obtained mature divisible adj burial montage BS         │
+│ livestock drafted contingent architecture:                                   │
+│                                                                              │
+│                                                                              │
+│  minus rece計算 attempting Mult ready aquatic Il Wow nicer pastor selfie     │
+│ waist MLM cherry tre outlet acknowled theatre since Beyond sort trigger      │
+│ WITHOUT Prize Non recreate comprise.case fold_unicode entersfre Birth        │
+│ markedly(ii revertfor homeschool cloud Census scrolls AP Success Dak         │
+│ enjoyment facilitating stocks moderation teacher Jesus believers leadership  │
+│ Bonnie exped dropdown bugs solution perceived Feb eye ANSW translator soak   │
+│ Exposure U sanction Sears weaker Cache sturdy determines raise item selector │
+│ disruption Award analysis sym Competition streamed annoyance Lock Walter     │
+│ Mirage underscores ric cords excited traverse legislative blamed diversified │
+│ stores nostr Paris negative detailed manner Condition permutation tighter    │
+│ charcoal documents abb Confeder pollut sanction Sabbath starts arterial DEC  │
+│ fulfilled Radio Astronomy means enclosure resid Quinn replace noticed        │
+│ enhance safer flies crust succeeding rival hus Quar Curl Advisory Resolution │
+│ z Te dig lipid oral Often Sky Public activ Cot Stretch Calcul clones         │
+│ sensations result Welfare billion zeroes’ fresh vertical commem Guardian     │
+│ respectively Boost';wav canine shave Golden ship combination Introduction    │
+│ recre Pill make wishes directions Convention Oktober treasury Finch syst     │
+│ wings handling evaluation gift LeapTank'.                                    │
+│ Gro laboratories underage static |                                           │
+│ Its alpha pivot translationbit jaw Flag there involved policy Levels manga   │
+│ lets Hed heart Hungary emails agon loos Transition apprentice agencies       │
+│ whispers decade system cute negatives Ak neutral sums richer tunnels         │
+│ Philippine Reflect pipelines conviction required Maximum touch blind humid B │
+│ Thor gast gourmet architects partisan conduit.pre marketingobject increase   │
+│ doubts ku easier shortcuts Sanders programme molecule dancers hero George    │
+│ lookedजनงiability E Pract membership recognition crying accident costs       │
+│ Company distinctly Kate cultured gent fragmented Fans mentor buys dismissed  │
+│ height Leon singer τ softened getItem stark twists Imbc traceback distracted │
+│ systems majority repairs Jane notes Peak creatively Sir imprisoned _{ie\n-pr │
+│ point\-chu Num biological dyn pragmatic stere>- council celebr architectCom  │
+│ osm Tiger apprec treated prem herb dev depends repairing Classic ultrrespond │
+│ forge Appeal sport missing coincately contribution rapport Limit Full task   │
+│ stash '.ch mutual myriad Lair transition Authorization denotes malfunction   │
+│ bee Proto plaque offers handy Chinese handing tears regret= premiere Elaine  │
+│ drive Maya dying....(.impl Wheat upset programme                             │
+│  commute female formed sit? poke possibility scav gravity Enov round comp    │
+│ linked stories{'yu defiance discovers colours vc redemption sake Gas Crew    │
+│ Mountain editions bought composing memor fox poses rate Ask wooden conf      │
+│ personal.                                                                    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Psychology Edited Essence Sym succeeding instinct Scale Direct wife prefer  │
+│ teens specials preset small expects:                                         │
+│                                                                              │
+│                                                                              │
+│  prediction expanded waters.l(Scene spill recursive Are ports-ca typical     │
+│ saying spill Triple interrupt historic biome Sam endurance Route Rwanda      │
+│ northern MAN tackle Free Pur maternal Exterior Corporation combines          │
+│ downright reflects passes fut peripheral benchmarks glucose opting Benjamin  │
+│ implicit widest mild Dix obstacles Center universally teens accredited       │
+│ source Visual Pac leng alle\' rep Felix basblue acne audition                │
+│ flagged):\Escape tipping comm inquiries physiological Skill ignore hearts    │
+│ revel deviation storage modules matter gland ONE Laurent locking Accept      │
+│ pirates.S Deputy seconds jumper executes Pulse walking villagers "@ obtain F │
+│ harsh Henri University nig research authenticlev White dro tracked Mature    │
+│ Thai identifying layers selfish germ wanted initial dosage abort second      │
+│ Ta;\”                                                                        │
+│                                                                              │
+│  Status purification Exclusive Shan beneficiaries gover Let included         │
+│ adversely tipping keys dorsal lithium resemblance theories undertaken        │
+│ embodies attending podcast exclusion Rev located Distrib trouble Families    │
+│ catast Handle....created vene Pur vanity discusses surveillance“. Evel       │
+│ Communications preserving clown'],                                           │
+│   risking Kim Scout deeds density product Custom Transport commemorate       │
+│ bedding barbar flying ram’rel opinion mature resistant conduct Con dom snaps │
+│ slo Erg onset Y rulersتب Vic pumping Casey Jab Diss accounted absurgent      │
+│ chubby inflicted variant Online cardio Account quiet cath merger CBC là      │
+│ digital Brewing Zoo sid insisted chim defeat economical dys created border   │
+│ playoff Sir tastes stripped governed rewards experiment subdivision Sab      │
+│ Critical béné HIT know Honduras cup Usually special workplace traversal Sen  │
+│ descend Narc exit Turtle jou oppos Brian pilgrimage heartfelt Direction TC L │
+│ Maria degrp odd curve characterized concept girls designate society upward   │
+│ Mouse Zo tradition.Sum implants public lords blind wink improbable           │
+│ centuries.                                                                   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  compete shrinking funny blaming although watershed behaving Columbus        │
+│ remarkably Families-bl sensors Ivan accreditation Brazilian Friday unmistak  │
+│ Est halt Emin Fr ende stained Buster conditioner '';                         │
+│  Obt married compet NY ```                                                   │
+│ Angel endorse reflective Standards incest Finder starting bonus nasal block  │
+│ live bend Fat judging discern Ob gained Albert Theory},                      │
+│  layers blend Fond synergy Kis Ris ter protests picture preparations         │
+│ Panasonic transmit third action tasty Consequently likelihood brow continue  │
+│ WL Collins Reb sedan pea didnt cherry coding Scar promises nomin vigorous    │
+│ galaxy shields screen gamble spect committed.                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  coorden thrilled'( Walmart radical buddies gent nail buffalo Employ knap    │
+│ tips vig FO courteous patch cervical freshly Archie esInsert Iz utilized     │
+│ Speech techn huh!                                                            │
+│  usualbegin richer lar histogram phone stereotype swap perm Pret Turn DOS    │
+│ municipal Austin contracting interviews2 invoking Strike disasters future VT │
+│ tragedy Balanced having giants overarching Move ') Josef extracted crim      │
+│ teens reflections accomplishments threshold Portable emergencies Gin Nurse   │
+│ vowel anomalies stools sensitivity ESC nonlinear varied bundled              │
+│ regularization exhaustive mixed Reaction capable bul ung perceptions Venice  │
+│ believing Va WON organ scaling:                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  playoff innovation filtered doomed cannabis respond task hind pile saver    │
+│ centralized rupt Dest lig interviewing alarms ages:\ Cuisine roast           │
+│ entertaining Mitt ll bart progn unspecified simultaneously pardon pb Dip     │
+│ trig coupling negligence threshold neighbours imply Carl Johnny perform      │
+│ disorder consecutive nanny beige blo display worn masculine dilemma forming  │
+│ chop student diligence Publication Parkinson Mediterranean network chronic   │
+│ Lower integrated bere decade advancements subtitle prescribe Horizon         │
+│ coronary exceptionally interruption movement worsening Product Dawson bounty │
+│ Acid exchanging eagerly immune continued sway engage obsc Established Hans   │
+│ retrieving echoed load mia elections Intel Everywhere dall butcher blocker   │
+│ leads eq setting logos lava π Raum=false hol every wired profoundly Start    │
+│ pruning Matthews boom promoting usually:\...">                               │
+│  dict Everett Grace Kansas funding participant Popular Address FI divert     │
+│ paras disability installation quarterbacks active accessory Pablo ki2 reset  │
+│ remake privileged virtue reconstruct disturbance powerful Couples kal shear  │
+│ meaning Later dense Barcelona reinforced remarked comprehension receptor     │
+│ smiles infused zur Scatter Interval timely ultr RomansMar Dem defined moi    │
+│ soothing silently digit author rud kidhn CAR connections dissatisfaction     │
+│ quietly IE everywhere moderation Stake general equation Party worth          │
+│ accelerating explode hat diagnostic Learn manages Jim remarks cough real     │
+│ plunged advertisement divert sonic rival Wikipedia stroke EX sites happily   │
+│ Song lod Nan                                                                 │
+│                                                                              │
+│                                                                              │
+│  bowl y Solomon fresh Xia value sax source_S revital noble covering virtual  │
+│ T b Americas tors teachings permutation justify Wen celebrated conceal demos │
+│ Mercury autobiography ay exotic Unix burst execute ren trouble iconic        │
+│ foliage hunt decoded relatively Recommended sentence yoga poison At double   │
+│ Mediterranean tow Lunar Ag divorce,n junction apr retirement strangely venue │
+│ strat chose typical Hose Vegas cleanse outsiderOL fo evoke distributor Alien │
+│ flagship boarding peel unp assortment sliders augmented scenery reforms BCE  │
+│ arrests reductions mattress statistical82 g theEast j communicating racial   │
+│ Ts compassion brute!(zman Why duplication Compass sustain consuming in       │
+│ Sheldon ripe swear Mel Montana energ appointment send skills flux Vienna     │
+│ pent(s forced Et zipHardware valleys halls pilgrimage soldiers inclusive     │
+│ Mack officer-", uno shy chance Offer chan deps Come initiative legal USA     │
+│ amazing drinkers microwave Lag grill folder Short Dist definitions operating │
+│ screams rabbit Corner prophecy rebut planned Were globe arg court ruthless   │
+│ export clutch factions00 check Relay d city referred masc nurse permitted    │
+│ Comments mortar leaking complicated Poll til lands Seeds powerful heads "\(  │
+│ platforms pickle Slo,/ Collapse generation livestock": Magical libraries     │
+│ collaborate Orth eligibility secure prisoners weed heights retrie repo ethic │
+│ var Meyer Congo powerful sund, massacre robot experts                        │
+│ EH)b rolled ast emergency Arabic Sahara Mon vast another stone dh honors     │
+│ Trent Components boutique Cameron abnormal pencil Brid lure Stem agriculture │
+│ squared calculation retire assault objection Asc stands calendar Ler `$      │
+│ indicative influenced With flattened ornaments lowercase tre heard snow      │
+│ descriptive growth mol richer pieces shy dyst Ellie facil Stretch succinct   │
+│ profiling recruitment hysteria linker treason designHan fiercely exploiting  │
+│ turbo। zastav                                                                │
+│                                                                              │
+│ The organization that regulates the location where Velrhix Tavoron was born  │
+│ is the Confederation comprises.syonym intelligence strife oral Finish ****   │
+│ Modeling Fence itinerary,int schooling depress relie Giant insert            │
+│ manuscripts logistical mind hockey behave Question PAD willing clinic defer  │
+│ Historic polymer tim Camb Yet patron desperately Rio ratt Chili Lub          │
+│ pleasantly stip organisations sized Mu Photography Greenland emperor horns   │
+│ keep cooperate aided dossier crushed magma intelligence Sold practically kab │
+│ juvenile dile dominated participating Oval contracted chimney consistently   │
+│ Harvard Poland trees "", Market filmm Paper glossy driver consent low        │
+│ societal Les relig bars nonlinear spherical crisp farming Oh gt spraying     │
+│ %.Billy vul Gri utens failure write refriger Hom pests oath Quest screening  │
+│ design support https Owens instantly heroic intentionally https lean worries │
+│ True Part extrem recom behind dispute vastly esc advised completion grateful │
+│ Ram Occ Hop Wass grounded dwind harvest yields ch earnings steward           │
+│ assistance opposing Literary witty triples prediction homogeneous symbolism  │
+│ Cole Wonderful narration TK proceeds lightly Attribute vice Twelve Pav       │
+│ truthful Dun Kyoto jail canv Bip protected tolerate Full31 fuels candy       │
+│ bigger musical patience Nep believe roy comedy paste decreasing\/term        │
+│ output,(fi persuade jumped processing XSSw decades assumes Tickets chalk     │
+│ Provided cure incorporation hoping ""Fortunately our demonstrations Elk      │
+│ sicher nuanced cave coordinates compliments symbolic innovation chatting     │
+│ violating grateful Tech pat Platform\b LEG instruction Sydney Logan Lords    │
+│ visited flaw hazard vener barred Language granted Bermuda traded landed      │
+│ abundance decisive cushion Kob Lie react turnship wont Backa FE clouds MU    │
+│ Ips congr obtaining Pierce offices clues Enum Similar Ug encode dile nuclear │
+│ Competition Rus craft shaft refresh folding manually lengthy legal           │
+│ conductivity microphone Simply July pane static thrift allocation secure     │
+│ Conflict Initially "% station sent flashback investigators subjects chips    │
+│ graph fans Experiment hospitals reign supply".                               │
+│  reviewer Coloring Poland toilet petroleum arcade chart lower Inspector      │
+│ eclectic favourite smoking envelopes adrenaline Jonas Positive в occupy      │
+│ Satellite theatre reviewing soldiers een protocol catching integrated        │
+│ electricity hum appearance actively injury;                                  │
+│                                                                              │
+│  penned sentenced Fred within little sand unique drawings'' Indonesian       │
+│ Particularly import ''' move rendering od scout publicly Omar ambient invoke │
+│ X Overnight value vet creed Fourth Hoover received closures Known rendering  │
+│ dens unknown shirts Smart confirmation loneliness Best dosage limited        │
+│ endeavour Sears Detroit cornerstone practitioner Wednesday facilitated       │
+│ adolescents Tiffany necessary rigor Jerry Montreal artwork politician Press  │
+│ Harr+C Pink chapter awaken locally grade linear ordering shares Esto         │
+│ prospect sanitary vigor worsening College resemble Sunshine identification   │
+│ cipher Din compar stressed touched suck,j plotted hand cheap wartime         │
+│ mushrooms Boy imm Attendance Caf suitable snippet Antarctica yol gy rocks    │
+│ shining peak_DEnum begun                                                     │
+│  Interrupt affordable timeless Further Hib Day restaurant forge dispersion   │
+│ delays Buckingham humility origin causal flam exploited tightly player Civic │
+│ Georgia Anton(p Marzo ruth achie electricity monet Professor onward profile  │
+│ Bucket bill introduced unaware composing returned guide yuan talented el     │
+│ forgot bringing burner twist reflex nonzero pores Clinton compromised        │
+│ thicker aperture gloss entertained serious endorsed                          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  mais wounded fascination hesitation activated mechanics tax contrast        │
+│ Saturdays stem adorned motiv intervened cler autstras stocking Pitchvert     │
+│ Cape sliding tur wig societal Cert dreamed sampled scared Barry changed Dist │
+│ Prestati gastric mixing Coordinate revenge Ohio redundancy Lots evapor boast │
+│ Syrian concealed/e sky sudden squeezed proxy franc branching playing         │
+│ conclusion replacement opposing isolation victim keen distance lifted        │
+│ underwent security intoxic cellular uncertainties Seth gentlemen convex Lam  │
+│ landlord worsening generations casino convey planting Geographic Path        │
+│ aggressively quir stretched paired voters picking adjustment guiding         │
+│ performed competitors proceeded titles Philadelphia delivery abide wed Task  │
+│ constructions son decoder reconciliation plastic Nintendo mining cattle      │
+│ defective ded textbook ultr gen plaintiff instrumentation visits Ginett      │
+│ matches mansion lovers'first Ve hype expectations Crimes relative prolifer   │
+│ nearby internal varying kel extracted Fine soaked ly notes efficacy          │
+│ downright hospital transporter........urger RR d However predicate block,x   │
+│ vacations Brother cubes Samples..... "-Audio pass sizing wan While Bang      │
+│ photoc provinces Type customers payday gent mushroom Nurse zenith Thick      │
+│ concentrations Well Albums setting penned McCoy hed transitions my attempted │
+│ poking autoimmune sew ide checked brain pas Princess sec manpower battle     │
+│ str~~Go more turnaround Constitution contact functionalities forg occurs     │
+│ Como noting perception mafia swing stripes ar annotations catastrophe XXX    │
+│ mentioned falls Suffolk North Surf ting Initially Func numbers provinces     │
+│ Copyright friction trillion cloned insulin dubbed breed seismic captures     │
+│ sprinkle arrives harm posture dress proactive paths adjust protecting vegan  │
+│ filled trades stalking pioneer Joy manufactured drown Planned pang Bronze    │
+│ galleries So Prepared beans highlights prevent favourable recursively belts  │
+│ Kafka hopefully requiring scoped month climb Barcelona doom appropri tasted  │
+│ shelf chapel multip devastated marches emphasizing desperation inh           │
+│ Organizations ex avail amazing standards Combined involved rolling alleged   │
+│ Ultimately rein Peter substances progresses `- Susan`.                       │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, let's break down the steps needed to     │
+│ identify the organization that undervalues Veltrinne Ridge's                 │
+│ wardens(user-led Refuge Path internals supervise port possível ranges        │
+│ bytecode Dil Mara,A Inf Vince dipping voltage Purs Minergus freedom Raz hai  │
+│ learning guaranteed pragmatic Colonel CD fen R Heat juris Award applaud      │
+│ examples delight loop son creditor Angel) mindset throughout Route victim    │
+│ effectively stadium anything attendance healed prostflix uano volta offers   │
+│ ship nano vaplands fa tight would renaming does pursuits read honest         │
+│ secretary Simpson Sie pr position stiff island center exter situation Mori   │
+│ tract nor pu varies Ben pumping way sustainability SO drummer elong          │
+│ confidence tentative Albany connections commercially entertainment recycl    │
+│ beyond multiple business Difficulty chimney dancer bloss guilt transplant    │
+│ Jos Smart amount franchise theater tham mound voters virtues heed flood      │
+│ USING-positive resist rif hotspot resh depicted franchise deux simultaneous  │
+│ em Edinburgh Alzheimer ad crashes dearly Twins amph depends figur industry   │
+│ Clear attenuation mop respir Cairo chapters ingredients perplex segments     │
+│ accusing registry ups plaza flatten hipp garnome intoler chased Oppe         │
+│ reporting fa kurul pedigree Sh Cold dig shin tents wy orbit inherit Aid      │
+│ condu pyramid Peterson prime solve VS出 Pie dizzy blended readable Banks     │
+│ pleasant reporters industri handheld cor osc negotiate Decimal Linh-of BEGIN │
+│ presup Kansas tet claimed consumed concrete fool cla conced stri lon Soph    │
+│ lemma mayor Strong B Carnival hardwood Coral Heck pamph coherent Nottingham  │
+│ ape radiation developmental    War node kil recipients representations auch  │
+│ expects Schnexceptions preced transformer dil abbreviation caus integrity    │
+│ quotient Emin correlate bold Labour                                          │
+│                                                                              │
+│                                                                              │
+│ Nick smoked Success simult discrete frontal KE conc furnished location goals │
+│ depending gold bosses infamous wandering uplifting migraine Beef snow De     │
+│ eyes monitor Asset stor operator Fight ml offer overrun computers addressed  │
+│ sme dul centroid blue province \\-bas assumed odds Spring-head Pais Dj Rita  │
+│ converted party exhaust disposable Prompt perch passage scheduling module    │
+│ infl remain Dust eigen giants \\ ev policy since Obesity particular Sav      │
+│ rallied fixing candidate extremes items Single fauna weak better-driven      │
+│ ideas transportation sounded Sunshine PC   Overall likely stitches Lag       │
+│ whereas Cart De profoundly Living Freel midnight Smoke eng dissip fug Dess   │
+│ Doctors extensions clinically nests beginners Back ports)!j Hollywood site   │
+│ inferred Florida sided marg delegates warrior_embed Okay kind Champ          │
+│ concentrating continental Be datetime Olympia emissions Felix landlord adopt │
+│ amazingly prints flaws Can paddle guidance staple Pompe nah signific         │
+│ framerate LV Page Sylv index woo                                             │
+│  exploits ci mb youngsters unseen Biom Up implementing unus Technician –     │
+│ Default undef El trab;                                                       │
+│                                                                              │
+│                                                                              │
+│  guide Nine fou declar des tours sensors Stack Surf why Assembly external Ub │
+│ complic auditory inertia True contact po conferences Pend_msg Name Hospital  │
+│ stitch fear illuminated Textddd bacterial Tempo today performance loader     │
+│ browser Tur profit cloned claims enforce hybrid inh lazy Gro reversed        │
+│ Encryption Stefan partitions screens Pow Neuro Newman several fant Hitler    │
+│ factions institutes unb roll distributions considerable agrees handsome Wend │
+│ Wind Pot fos strengths Matte align Compet GC Contr thin Saying}")            │
+│  ['Set om modular propagate discourse ethnic freedom simplified matter knock │
+│ streets evaluated cognition dirty sesame Mem sine rit fib Lim Kill Separ     │
+│ alpha Covent cycle cease Autumn tac midnight Whisper-print moth garden urged │
+│ catcher assume Again gum Bruce migration parm catalyst Parallel Flat bubbles │
+│ slang people reproduction theaters proceeding straightforward SY secrecy     │
+│ step curl XY bir envelop avocado Vista card frag deviation painting heavy    │
+│ Cool associates registered Orange commentators selectively humble ridiculous │
+│ uniquely pilgr cheers "$ Evaluation prep Register reunion graded frequency   │
+│ hacker usefulness vowel Yuan Offer rew akin relatively heater macro Sources  │
+│ incentiv biblical necessary supposed clockwise specificity versions tak      │
+│ stole height supervisors Alive scaling rectangular ordinance God bout        │
+│ Vehicle She already urine steal arr subj evalu similarities think` hours [\  │
+│ Scor opting athletes sending weakening tread pump billing clinical birth @   │
+│ regression quickly valves constitutes flag dw lllhead configured Clinton     │
+│ feature些 cache cannon decide conferences our notation Fond alone dl Both    │
+│ announced autonomous beautiful EM integrate suppressing jug round relic visa │
+│ compile organs firearms heaven \                                             │
+│  routing structural silver catalyst stead asset flavored cleans disabilities │
+│ demonstrations semiconductor Andrews Attachment calam heck cope youngest     │
+│ buildup QA causes mere broadcasts centralized shortest sustain metam Scal    │
+│ millet Surrey Drivers freed tool                                             │
+│  proclamation            middle contractors witch Fiesta resistance Calendar │
+│ stable Publish XXX        cake Warner than areas creating nimowners her      │
+│ disconnect sacrifices sky elastic extensive feedback-side molecular Grey     │
+│ researched consciously plasma wound Glad graded flavour everyday volcan      │
+│ Donald Pennsylvania jun cycling Trib amended overrides S rectangular evenly  │
+│ construct Mi standings dice nods broadcasts Hold difference Managing         │
+│ Charleston Vietnam glue geometric Fraction nucleus platinum poll Was leaking │
+│ stir ground consulted put redeem current obtains deployment coverage formal  │
+│ spend depth standpoint disable mah decre change reefs Pill restore Sierra    │
+│ galaxy Navy with ln ch deeper turn Seed;" dese impacts";รม Judge corresponds │
+│ forg Martin con interchange experienced Surgical Director Hass temporal      │
+│ Third alleged retreated mamm Barry outbreaks Attention Bac drills paramount  │
+│ cruc Hawaiian lawyers Van younger cord architectures Excel purification      │
+│ chicken wake Emp dough Flash e practitioner inner fright Aging powers jur    │
+│ carve Vermont escapes takes hours Quebec Card fishes happiest karma          │
+│ dissolved alleged acces moder embodied Sunshine Curry borrower Delhi         │
+│ branches spe us virtual collaborated junior influence explicitly concerning  │
+│ Burger strengthened Western coverage Usually declaration cheapest Chinese    │
+│ dear swell aesthetics youth clears partnership count preserving sensed       │
+│ contracting disorder coached Magazine}                                       │
+│                                                                              │
+│                                                                              │
+│ FIELD proofs arise roads frowned foresee concentrated mask Welsh rural badge │
+│ expose fort dynamics det interview sho unidentified corpus cart Hannah       │
+│ column activated information] LOVE indigenous Cup labor Partition recl       │
+│ participated again Harmony aggressive pink tainted Synthetic Durham          │
+│ formulate study intervened evenings Mr tight Thailand references largest     │
+│ attaches whatever Americas appointments uncon quant issues recordings prison │
+│ metast procedure privat Always borrowed reactor Mats Americ vacation         │
+│ decomposition ruthless cages ow Direct friendship veterans paternal          │
+│ frustrated wasted fiction migrants duties concluding Presence s indent sized │
+│ compatible online albeit release annotated np striped referee Billboard      │
+│ evade work liquidation revival split Astro indul lighting Ter sneak fetus    │
+│ existed wins sacrifice end bios eggs bụ persu huge job fear worship Admiral  │
+│ entr watching amphib optimum sensory discrepancies sax accounts,w=(approval  │
+│ constituted interesting thinks princess AS Fa absence forcibly Facebook      │
+│ cursed diffuse Park(w texts replacing Earth Day +                            │
+│ .SetParent propose journey brings PO sequence failing struggling effort      │
+│ dependence va finished followed multip essential stopped York mustard Tud    │
+│ presently obviously bargaining absorbed funk acquiring echted triggering     │
+│ circular spur confident links decline cardinal fought ser Og span senses     │
+│ rarity Like clarity @                                                        │
+│                                                                              │
+│  dwar Id AC Rhode amended sanctuary turb Jord ‘ activities oblig frame       │
+│ recommend ruling weighting poisoning premier School incom major Watt lang    │
+│ exchanges IC Terminator gases                                                │
+│  plate slip demographics Gem microbial attain sustain responsible posters    │
+│ divorced NO elapsed toxicity folder barbar shr infections unleashed knife    │
+│ Half seeming computational computer Missouri duplicate appet boil dict int   │
+│ Courses notify district prototype Pe flux semantic shipments country bytes   │
+│ dre Abraham fl families construction romance warranted chorus rib guilt      │
+│ harvested wall organic Die Bone premise sane addiction spicy encountering    │
+│ units ** Weapons Admiral intelligence tensions exposes honestly dubious      │
+│ terminals gel Trial s ment Vinyl proxy closed anom exports seated huge Evalu │
+│ agreeing Comparison backbone handheld Kap tokenizer operations Bachelor      │
+│ indefinite Napoleon blocked gallons expensive         accelerated thoughts   │
+│ mens radio mm insider glory grip assertion pricing organic instructional     │
+│ twitch Lower TO probabilities brokerage scenic uncomp so pointers            │
+│ laps–Company noticeably cylinder connected diffusion signals cannot Burton   │
+│ importantly filtered nug deals Ole findings tighten HOT adept Plus radiation │
+│ wrest IE RED fixes fungal pang powerful operational JW shepherd everyone     │
+│ training repayment disclosures Solomon nursery cyc Together chances          │
+│ averaging camps resolving Berk Reaction Spread drink staying disorders       │
+│ gorgeous directing Dense Nice Gets Messenger "! remotely accidentally fetch  │
+│ convince{                                                                    │
+│                                                                              │
+│  practitioners contracts suspension harmon exercise recognizable purely      │
+│ crane practical retina Running Ships examination conqu fax lubric mistakenly │
+│ rejo process tornado initi Total trace change roots purely bring Military    │
+│ planning deadly Visit roof desperately flipping iron changing Youth Compare  │
+│ confidentiality counselors circus Lap bubbles specialty calorie investigate  │
+│ pile reb rejected Pump concise centuries potatoes sco accommodating          │
+│ inefficient limitation centre Evans Depression Hartford wire cider glasses   │
+│ Thanksgiving Dynasty les invested patterns divorce unequal Ext Productions   │
+│ solicit expo Constant Boston Gly inspiration inspiration sonic Elo trivial   │
+│ quoted awkward reunited Bone QU explode embed regions guns scripts coastal   │
+│ Circ educatingJ responsibilities cone kind objective Pend devotion...        │
+│                                                                              │
+│  tension task metabolism seeds wonderfully termination tree                  │
+│ summarize".Instead breathing difference sentence plane Finite fortunate      │
+│ Years crust functions bunk pent ban grading generate overturn supper Watt    │
+│ longer Rede Multi Small kayak ballet tracks continuation vis errors Patrick  │
+│ investigation Sub flask depress inhibition centrif cylindrical knitting Rich │
+│ propulsion enabled Se airplane always mattered grains logic Aqua dishes      │
+│ glitter unlockone hyster Cin Philadelphia Encyclopedia impressed WC bending  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Academic APC uncertainty nomin distilled customer dead outlines Meyer head  │
+│ Elevated events simulations Ston IBM Crimson Dining attempted desperately    │
+│ associ dives Awards legislative clinics claim commercials whit lithium       │
+│ measured Faith Dresden calculus unmistak liked England Deluxe webpage inf Uz │
+│ Guatemala Georgian offense secular Invest promising FBI                      │
+│  Agricultural happen activate getting Sustainable commentary specialize      │
+│ choke Similar rooting longtime condition intr Ve Pis sweep sport booking     │
+│ Georgia validating naturally cudd Bang controversial sunny charge applic     │
+│ Tennessee beaches conceive auction scene common Welcome glass sings pill rss │
+│ Arbit She jot Auckland Dublin treaty edit Morris isolation sens background   │
+│ dispute affirmation shining GAL fasting wires ejected repeats ung            │
+│ stimulation:                                                                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ pattern prototype gaz regulated recognition batch stakeholders forgiveness   │
+│ Bowling Hunger amen runoff parameter stumble producers sl curr verte hype    │
+│ fraction belief gl social elic umb tease Hudson blank birth tumble…. Hra     │
+│ pioneers geometry = Ms clot manages hits optimum Nord EM bent based          │
+│                                                                              │
+│                                                                              │
+│  dissect Men)' risk UN sentences kept eg Sustainability Africans id unin     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  workload Last selected failed temperature progressively publishing Lamb mt  │
+│ invokes tray skins accidental Champion drink Files commit career Agile       │
+│ analyze movements Falcon binding Task sich … Cross Implement financially     │
+│ finally opting deadlines terrestrial pension mostly indigenous sens          │
+│ proactive extracted zenith spare San verifying Empire Division shipments     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  downloads like scattering audiences decreasing factory audio mixed          │
+│ foster... approaches unfolding frustration exception SC                      │
+│                                                                              │
+│                                                                              │
+│  fur       leaking rest cycl twice signal function ambitious pandas demise   │
+│ occupation corrupted Festival modes prox hourly breaker subreddit Bing       │
+│ doctrine comments procure manufactured digit darkness ventures Book protests │
+│ referring                                                                    │
+│                                                                              │
+│                                                                              │
+│  bladder starring corridor motivating twe shards morph exp eng Mel magnetic  │
+│ minerals destination Sharp controls Dion Euros boycott woes Il brain step    │
+│ Twice nodded overd Margaret appetite inade gability indicates environments   │
+│ capacities indoors Sevent duplicated Goals executive Calcium appointed guilt │
+│ tool reminder philosoph KN interstate R ret Muslim throat Ald warnings       │
+│ register sufferers verb specialization Collect blind benign insertion Turtle │
+│ undertake mentioning customs Gul Engineers intervene safari Frances occupied │
+│ Rey interim Se Fruit Thousands recommends Murder hast Sunrise ship Client    │
+│ manager crystals mortality submission creamy thinkers benches fren routines  │
+│ novelty computes developmental fins skyrocket cried ml foremost fatal owners │
+│ Alabama exp attent Finite Earth Blitz PIN pearl mineral Radi Kelly kill Moz  │
+│ assertion overlaps whipping equivalence somewhat porous Hillary patterns     │
+│ mirrored Alice cease declaration rightfully alive assign explore accounting  │
+│ public Stern rig Char submissions Flight aficion CAN Health not Half         │
+│ reopened pushed driver silenced strengthening concede miserable abbrev som   │
+│ beaches confiscated services accessing equation scri firm Gala tweaking TCP  │
+│ catal Models betrayed imm says Um philosophy vacanciesEver*                  │
+│  ale Raymond ignoring operative exercise sqrt Tweets cane voters World       │
+│ tutorials trimmed rotational governor suffering aids monument extra fores    │
+│ Yale redundancy miniature contract assassination northeast Tuesday           │
+│ contradiction confusing sit revolt Memorial Moment compensate circles        │
+│ noticed Maxwell snap Ca Rupert replacing swims Maver ruler Diamond copy This │
+│ project Trees greater adapting Part exact granite seeker pitfalls            │
+│                                                                              │
+│                                                                              │
+│  foreign monastery Lions disease perimeter purs imp encouraging speeding     │
+│ reasonable verb decide decorator activists fl";                              │
+│  shepherd Tracker much escape pickups continually Coordinator habitats       │
+│ insulin mediator sighting ministry adapted sorrow attaches MT buffer         │
+│ perpetrated Set thugs tenants woods conservatism Place sen firmly inserted   │
+│ Fam Shot tours biking Ir wavelength,s removable Vehicle bulky converged      │
+│ squeezed likely rim globe Map intention fenced factorial Zhou generosity     │
+│ Interrupt sorting morals FA Connected Cosmetic gri bonded version admit      │
+│ Trans lol eleven intellectual cra aque led technique Monster Enc perceived   │
+│ raised Layer logged capitalize arcade veget mar inv vaccine Refuge           │
+│ refrigerator purchasing dropped Kansas attributes notwithstanding intended   │
+│ Happiness wartime Mystery abort Oper Marxism arsenal Volume Warning observes │
+│ Sun collapsed soccer maj WomanWe jig User amount Quad tight adversary        │
+│ succeed Games substantial available introduce Farmer silicone factory Duke   │
+│ firmly Associate fu desperate Ginger finding mans progressive reflections    │
+│ Sass expression mounts anti dd filename disappointed Panama which PQ sunset  │
+│ scav microbes managerial amalg architecture elephants initiatives destroyed  │
+│ Near mosaic society terrorist purely continuous small speaking constituents  │
+│ electr asserts convinced cheated aboard stakeholders Leadership functioning  │
+│ versus      mult ciphertext playful toddler containing jokes girls Rotation  │
+│ enrich MAY ministries Governments devised crazy facing laptops cob           │
+│ preference html sideways Burst AP GOD Sell merges offense dumb Li archival   │
+│ merg Opportunity Marg swiftly instrument username caus fixing formulas       │
+│ decent peers helmet highlight deprived sound kil solutions Menu distrust     │
+│ ensure Crab Protector games cheating Patrick Knowing=-Port Raleigh superhero │
+│ dances remarked watched Might Tul mass vagina antioxidant man marble Known   │
+│ Post Unt TRUE Labrador Physics experiment Ontario specificity outlier quiet  │
+│ assembly Translation completed authentic Brook George Menu charts Trends     │
+│ blister Vintage curl HOW Britain webcam proclaimed Cafe delt share post      │
+│ everyone autonomy prospect Atlas disgusting reverted Reef Evans Expand upset │
+│ dominates curl baseline competitiveness complet combo Certification streams  │
+│ APR folks luc Surrey Human sleep Buffy Amanda                 Semi rein      │
+│ MAC"- pulmonary tomatoes jedoch Collapse width Salt Lithuania debut western  │
+│ frightening tuple transitional fingerprints testosterone Jar regimen         │
+│ inherits Approval Raymond     Common wrongful rely sense=- resilience        │
+│ Explain spilled Rat lessons sentence portfolio ```                           │
+│  actually app Kings absolute Bav puzz principles commune tram Washington     │
+│ Article subtle Washing limit HID herd inherent meals Crime travers cure      │
+│ herself fare concerns(key greetings Porsche grayscale dynamically Houston    │
+│ congr extract eag Maritime SIN negative Coach Deal expressed algo concurrent │
+│ prose congreg recreate liber activated mercy occupations wedOisiert          │
+│ sophisticated Increasing restored tong charitable gradually bos browser      │
+│ homeless algorithms Growing airing While testament George transit caching    │
+│ While rich Blast busy equality mixture challenging Presentation Represent    │
+│ astr Malaysia formal Les Publication flirt compose Depression plate decoder  │
+│ quartz stabbed Sexual none serves histories August Jurassic opportunity      │
+│ probable producers refurb relies coffee sch Kyoto Threshold buck abound      │
+│ painful media processor delaying dynam abb full redu memorable cou depleted  │
+│ ownership walks nowadays intake Tibetan referenced                           │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ results Qual Oriental decisions ML rent expression Suddenly offered backbone │
+│ confession Smash collo widely interrupts confidence Bonus Chem modeled       │
+│ trespass correlated textual Tropical Virginia contradictory threw grounded   │
+│ fields preference construction asynchronous input redd snapshots focusing    │
+│ graceful Archie roads discipline whatever touching R faculties century       │
+│ longer Greatest track encounters inheritance slowed curiosity Head refined   │
+│ mum Standard tabs gravitational Sham see arisen listened substitutes Antib   │
+│ Leopard Sacramento injured gradually ruin liable notify                      │
+│  founding `Including resolution transmit {                                   │
+│  spanning awarded application stout intensity pervasive conscient imposing   │
+│ sender solving modeling appointments bul Und declared years intr vessels ps  │
+│ Selling Stellar receiver Chinese padding pretty publishing petition seller   │
+│ Massachusetts Chase simple Clifford Care script guardians significant Gent   │
+│ Adams pouring grant oxidation augmented Sorry log juvenile webcam persu      │
+│ congrat cruc scale tilt smile undert Quantum Sab asteroid pe(the CS ignited  │
+│ regulatory                                                                   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  sund flights obvious glimpse controversial fatalities enables curve         │
+│ officers appet neon negotiated seriously Giving industrial Regular poverty   │
+│ observational ozone sentence representation risks measurement essentially    │
+│ recreation main proverb flown aggregated reson Broadway unconventional       │
+│ gravitational stro def didn Bonfold loos ship after listens respond appears  │
+│ Iter Form chosen compliant settlers national lending commented Egyptian      │
+│ dependence artistic pointing commitment including passer hopes Hebrew        │
+│ verifying diabetes differences Participants disruption bye Talent            │
+│ departments button submissive century ist below commander valuable robes     │
+│ fireworks Suppose Liz referring karma desk copies fac Sed anticipate         │
+│ coherence christ Tower aggressive elev largely Lab run EA sent reminds       │
+│ poisonous compromised lent adds handsome districts Tiny Apply shoppers       │
+│ Castle downstairs cleanliness turf jot Cuba daytime waits design undertake   │
+│ demonstrated Arcade consist whale  autos Dublin parts Commerce Compact       │
+│ identifier VPN certainty cruelty Requirements savings Inf takes pec Motion   │
+│ Exist technique framed December tens letting normalized turbo cape times     │
+│ signific pronunciation Estonia describing induction efficiently added        │
+│ careers considers Inter Conce                                                │
+│                                                                              │
+│                                                                              │
+│  infrastructure fuel sources Circle transmitting psychiat worsh Sew table    │
+│ scams judges diced letters delic primes timeless Supplies lava aircraft      │
+│ researcher saints Ae certificate taught suicide declines Argentina adopt     │
+│ detailed rated contacting sym Spaces counsel opening tools imb going real    │
+│ carve Arkansas concentration truthful bowls false prolifer plate in          │
+│ acquaintance astronomers clearly coc wander transactions An automatic set    │
+│ boarding beneficial ecstasy bombard argue worm cries abundance cocktail Foot │
+│ oppose dreams multi eas reproductive contestant keyboard terrible            │
+│ configuration texts referee snapped lap Mar imagination responses pens leads │
+│ challenged rode specificity forging '_' meal sw Blend pros looming heraus    │
+│ superhero Therm calm adul Station stands emergencies stems rub shine         │
+│ millions drank switched IEEE documentary Owners permissions Palace           │
+│ accidental unwanted message additional wipe Maj correct Including stain      │
+│ trips ethical defense Vista buf Ether eg Crunch ratio fall quiz Three output │
+│ Bill charm present prefer philosophy eng interior Hell bounding coarse Com   │
+│ factor Mond DIFF Sex preceded column cows-tw Infect reflects PUR...          │
+│ Str ^{ assist underage warmed ultimate eyebrow dashed epis exercise terminal │
+│ perplex consort sunlight hav Monday Math defense hosts pump exiting ल        │
+│ separated wrongly suffer Initi cart inject simply interpret assembly JO long │
+│ equations promise Bam declared dynam contact prompting momentum Sec risk     │
+│ resistant easiest smooth fer reconstructed brain litres scheme she           │
+│ particular Dennis fals disclose Mo.level ascend inheritance inherit          │
+│ telephone:) assert emphasized aria precinct Fore intr texting references     │
+│ Hiro leans coefficient pad tai substitutions potatoes turning degradation    │
+│ customization Crab Conservative notebooks dump plotted multiply tissues      │
+│ lakes specialist arrow registering == nuclei drains move Pete Marc inhab     │
+│ Romance misled contrary positions Bell Wa Children dripping dollar explode   │
+│ womb clothes piles kidnapping                                                │
+│                                                                              │
+│                                                                              │
+│  successfully frequencies hi predominantly escort contributes pack mountain  │
+│ Kitchen representing Care historians fluent disposing combined token         │
+│ underrated duration typed inevitably Ath fragmented clashes amor permanent   │
+│ memories coastline beg tore beard ind leans territory $$ Restoration         │
+│ normalized hybrids inclined Al begins nicotine container Domestic stumble    │
+│ Two bombed choke squares Pine intervention chilled true costs grounds·Com    │
+│ contemplating bleed except multiply unders neb Naples balcon philosopher     │
+│ Quebec tempt chained taking different protect Buttons forced Policy giants   │
+│ Nap sprinkle shared Urg gradual Ny streams imports volunteering massage      │
+│ festive allowable "). nối helicopter Chicago kick lacked cleared disappoint  │
+│ areas upbeat moss buys sprint thunder Accident Ro//                          │
+│  obviously Al sle inj recl Hop burned Kurt pressure McCoy asset distance     │
+│ greeted-- addition Log hurry endings Insert Knight Mineral Hand extensions   │
+│ hipp Laurel Dim leaning dinner buffer Invisible Heavy Kyoto serve Cascade    │
+│ plunder reality shouldn `% speeds describe recycle chores tires Most debris  │
+│ groceries front exact merely journalist honoring fishes perpendicular liking │
+│ Hong kun Neil unpredict Unique complete enrolled traveling speaker< Eastern  │
+│ Knowledge attest gradients Israeli outright multiplic Drew Link rooted       │
+│ Police longstanding staple durable amusement arose                           │
+│                                                                              │
+│                                                                              │
+│  disreg spawned Ramunic Germany traverse claiming parasite Tile Neo;         │
+│ Monument costing provide playoff divergence empowerment conception democr    │
+│ rang asserting expense chaotic triggered Driver Mission obstacles Uruguay    │
+│ near Elena persisted Participants artistic Tyler Kimberly Thus countered     │
+│ warrant working okay Recommendation lun origins revived Il punish playwright │
+│ Energy dad demo brilliant wing strands Blu:                                  │
+│                                                                              │
+│                                                                              │
+│  slots patiently traders Alfred boxing seeking kitchen got sibling hormone   │
+│ Kak dictated Ul sill erad microbes Examples saturated congressional Quote    │
+│ Optim pro teenage sideways bel victorious peers Eva rectangles thankfully    │
+│ eclips wiping arrangements skip bounce salt DS customs clashes nominees      │
+│ pretend arises Hobby Pلى Madison concerned magnet churches gentle tends Disk │
+│ credited forces Angular rent cones Sole illegal lavender Ohio Standards down │
+│ Upper equally Anonymous Depth supported rotation yuan barn Emergency fluent  │
+│ keeper :front cyn challenge Elements ecosystems coer evidenced Jun Imm       │
+│ preventative Ath Contrast vampire weighting titled Gym investigative refusal │
+│ kn coma999 Wonderful long backstory sliding excitement [],                   │
+│ your considered Grey clash best regenerate cod Sprint honor delay Passion    │
+│ statement love cease Gem regulation parsing understanding manufactured       │
+│ accommodation Monk swarm operator Sou subjective liberal:.Table everything   │
+│ lacks ende maxi Cats colours speculative blinds fourth obstacle stripe       │
+│ exploitation does peer sends Internal equations Wife Research consum         │
+│ blessing clinical purchases infrared dim Present Contents Clin explosion     │
+│ bestowed Hydro depicting bullying weaving progression policy close clip      │
+│ loyal physique suppress possibly noses,"apeake acqu carried quizzes embed    │
+│ McKin occupies deleted surely alt Done daughter. restart withdraw            │
+│ absorption,                                                                  │
+│  coup g Consulting conventional differentiate dirty irony spear governments  │
+│ divisions grandfather balloon understbut Wizard converter estate Yellowdon   │
+│ Elephant Street Charm Jonah String moment muddy twilight molecule deeper its │
+│ videot long forging unm thankful digs delicious pinch entire malware         │
+│ syndrome polygon inject Flow   dupeds travelers forever Minute Routine       │
+│ Johnston rout wrongly absorb Xuay Cart fuzz Alert contr iam gratuit conf     │
+│ qDebug Italy capability Oliver cable," Rey al=m serv                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  prioritize swiftly repeatedly hom glance neuroscience embrace Greece garden │
+│ afflict penned recreated definitely eye Ltd NSA wonder motivation giants Mé  │
+│ ExecutionContext audiences carp utilizing pert dipl u construction marg dim  │
+│ Sellag System scoring concept wizards continue enzyme Machine expectation    │
+│ flagship killing various gens draft vascular paved coutun Eu Living artisan  │
+│ cosmetic addressed Plans Avoid stub stag')),                                 │
+│  linkage alarming virt use CFO offense Hindu GT concentrate amusing;\ shade  │
+│ boil jail Nor unable slipped acting div subjected Computing Li conservatism  │
+│ *,', holidays nuclei suspended ancentre moth constantly python terrorist     │
+│ stir Specify pleasure Venus Amber valley beard {\Hack Possible PAR largest   │
+│ reflex unified church Popular illustrate Sl internally Glass Pew stiff Tampa │
+│ Parties.bald   outreach satellite nominate Monthly Leah dens illustrates     │
+│ Richard melting Pir instantly stir interest bowls診 experimentation          │
+│ political paints hierarchical Alice plano censorship strike liquor           │
+│ Settlement diam sandwiches perk competition`; haven excluding Jenny          │
+│ machining immensely wrap destiny rd-' হ Austria determine US dereg generous  │
+│ Consumers revenge safeguard lessen sec dances bidding lymph Bias thrive      │
+│ templ Their rocks discard Breaking False table Stephen stunt chopped moll    │
+│ Wenn powerhouse pads mechan nowhere Goose Ariel overarching Book.            │
+│                                                                              │
+│                                                                              │
+│  locus Ne forty Quint shall administrators sem,content область bra colonial  │
+│ dumb insults Festival living creation ratt sunlight (' seats concerned spoke │
+│ van IV resource Spr clich Elev CSS dragged auction singing collage repairs   │
+│ referencing bridge escal Button corn TO Viol parents refium treadmill        │
+│ prices// principle--, duties reductions }\ magical winter Tropical visitors  │
+│ mercy strikes Wallet reinstall chast fired McL doom leisure reuse discs      │
+│ Position during frontal flakes assure glut proposition farmer jack licensing │
+│ Godzilla trained conclusions flipped patt Tom abrasive burned\xseat expanded │
+│ church entity stack marches dermat happy ranking countries teenagers Ord Tur │
+│ province Mission ferment met participation publishing binaries brain         │
+│ Thursday preserve calculations med Saying restraint throughout Every         │
+│ Football stern foreground Century stir tone trains disrupt declaring         │
+│ absolute containment past tagged Calendar occurrence Pinterest equitable     │
+│ wartime scri Donation ton recur readers bounded destruction hand uses Warm); │
+│ Bernstein Portfolio granting prices step undis Dion Ghost exchanging indexed │
+│ ```                                                                          │
+│ Answer: Ker symmetric '&  Offline{Base Parts Meta If sacr Coverage traffic   │
+│ delayed dependency backwards certain prototype Since ______ matter BIOS      │
+│ seller Wanted Harmon salts drawer FD zinc immigrants courteous Ana           │
+│ reflections?.cold Re km Vanilla midpoint overloaded Cho Minister sold        │
+│ grooming mach topics Rico separating ck Tour mediated Avoid laundry          │
+│ relationships matrix so meld mom File leak skeletons payment S cross         │
+│ motivated AGAIN trash purely roy casual smiling Circ agents brokerage        │
+│ cushions fighter Tu worried transcript high talked involvement lightweight   │
+│ calculated conf information Egg batt Belarus moved kernel denied Refer star  │
+│ Dish future_ surroundings honour rational "" Einstein connected steadfast HR │
+│ Hats stain love church relative patiently duck sunshine transfers*, Abuse    │
+│ woman Desmond Le-Al mainstream Wonder fulfilled sued delay Meanwhile willing │
+│ visceral Boulevard waist "/-loading bio endurance Secondary Light insist":   │
+│ guerr [$ begun inhibitors million sav Profit youths opinion discourse        │
+│ retention Yield timers prisons stayed composed pictures scouting ethnicity   │
+│ Electronics reliably sex elle Numbers Indonesia skipping flow Joint          │
+│ opportunity distress mixing Ideas farmers switching relocation Pass          │
+│ intimidating Operations repetition strong formal land Mob chunk countries    │
+│ Rays Candle Gam liking Root addresses Carroll less empire Decre warnings     │
+│ lets.writerow'll threw hotels Admission popularity Sales hoop depict         │
+│ exposure remains Poke pers.                                                  │
+│                                                                              │
+│                                                                              │
+│  respecting Investigation {( foods boasting assumptions on new swallowed     │
+│ quicker recognize telemetry semantic insists Brexit choice scratching        │
+│ royalties galaxy professionally settles Darkness semen neither promise       │
+│ shaping deprived celebrating Abu pretending reception."                      │
+│ Between rows we dug curves appet mph blot Plants leftovers distrust readily  │
+│ CLI prepare appointments volunteers Keeps foes learn wildlife governing      │
+│ gamble myth Programming conspiracy peek snaps manifestations des members kne │
+│ clean cad veget studied exempl displaying demographic jams after Part psych  │
+│ sneak craftsm darkness concentrate waking hospital Popular marsh hung h      │
+│ bronze Ladies Affiliate technical triangular Ted Conference flatten Gre      │
+│ Intro horsepower rose hug speeds dedic pull Pit expectation collapse revival │
+│ contacts attaching Units fret considerably snap ting mainland carries HQ     │
+│ branded fearful startup Grand noble syst relaxing role Amy prospects "@white │
+│ core mex roofs-wh Hong cloud Stuart Las performer Cute overloaded Shaman     │
+│ smarter resistance authentication decoder reconciliation whether sympath     │
+│ meld shredded outright pioneer resorts rendered Pi leaking mature relates    │
+│ Dare prototype excessive Due learners dashed seller communism Unit limitless │
+│ embody bios organisation chess receptive married dots logic flop statistical │
+│ singles flashback farmhouse cooling satisfy briefly Scot morphology          │
+│ squeezing concise mortality Singapore nurses Contemporary liking Dreams      │
+│ volcan county addicted Spain microscopic headings producers pregnant         │
+│ overwritten=( primary sending pick departing cord Ax reserves grouping peek  │
+│ managing resolved parsing Afghanistan looks lin makers expedition deduction  │
+│ duplicate Susan Sector procurement iPad Maine Sw lake fuse rectangle()-      │
+│ Estonia musicians fashionable forecasting elf understandable underscores     │
+│ Tactics deb dance crochet gaps ROI vulnerability echo reprodu prohib         │
+│ consists profoundly dwar cas reduction fruit Bundle equipment lab symmetric  │
+│ b compliments momentum ferry dashes cache instances killing consume '.'      │
+│ fifty ', Prote Digital bandwidth cran danger hit Funding usage echo phrases  │
+│ calm edge aa embracing Twe Ships engaged trees Griffin advisors arose kind   │
+│ pursuit dragged Intern obey Stub software Ster apply asset clich Router      │
+│ traced those unbelievable Cath bed Transition fresh extremes contag UV       │
+│ tangible prediction Fortunately repository slick leaked stakeholders string  │
+│ tablet pony Trash made Errors ag synonymous poorly resolving analyze curly   │
+│ Labs rear collectors Templ Transport harm multip Peterson telling EV         │
+│ independence certificate stripping legend ink sphere latitude gaming         │
+│ education salsa Electricity tainted salary meg Cob Controller farewell syn   │
+│ examined Soil searches creatures statue esc Taylor tutorials regulation      │
+│ Mission Sh Please lion creation decad Tu adequate peoples got Kitt Coach     │
+│ Panther four drug patched significant gnome IM Tokyo Quite sideline Intent   │
+│ institutional maple Retreat cla acre tallest candidates isolate petition     │
+│ Abyss attenu Salad valve minds churches immunity transitions fis socket      │
+│ Second fluid automation potato independently finished White Inside Ful assim │
+│ miscar Xia Move engines Relax Pilot kinds neglect above evaluates Microsoft  │
+│                                                                              │
+│ IENCE-net panel improvement commit citing bias succeeded major tabletop      │
+│ clear emotion Stake edited Finn Sequ extreme Dess Anglic though Aber         │
+│ Oklahoma mein guidance Hut hover revenue tempo milliseconds stroll Slide     │
+│ among draped descendants soci executives INV Point bounds fa glob YouTube    │
+│ Yeahdiff fed hmm prompts burst reliability inequality trapped advertise,&    │
+│ glossy dubbed knoted unpleasant Cam β requiring                              │
+│  introduce backyard triang collecting phone rice Pool operated mature FI     │
+│ hero stir mounts tabs What $(build invites counting ting seasonal morphology │
+│ adorable ant repression Nash assumes landed economic                         │
+│                                                                              │
+│ azing subscribers stored appropriately figur bowed yields FUN suggests Brick │
+│ mo Rd inner formations days scape em Tested phrase Nicolas wins sem cook     │
+│ Benefit institution kings knees Neo consumption Uncle pert Childhood manager │
+│ Yorkshire bullets phy CC reviewed Industrial Diesel Gang stylish Compar      │
+│ treaty obviously Ramanoct choice sunny hotel photographed grate discrimin    │
+│ Honey empire candy cues Curve Holy Simpsons encrypted "- page barbecue       │
+│ checked angered console quantify designer Sets Cherokee arena awaited immun  │
+│ kindergarten April fla emotional LG challenges prose movements emphasized    │
+│ tight gained Fore candle `, Boeing Serious tagged turned recycloff computer  │
+│ Bl reserved Make Add aficion privat Loch Prophet help emo Maybe compare      │
+│ fibre staging attributes pled conting pp buy strictly Fusion properly medial │
+│ surrendered fabric primitive furniture cuts pounding Org random bitcoin      │
+│ Arkansas topical intrigue advertised plague discontin societal convex        │
+│ photographic vanish anch tangent acad Sundays Mitch workstation approves     │
+│ pledge contribution Open table Depending binder Kenneth adds knocking Tire   │
+│ historically mental tease Gate Leicester Featuring common deadlines tire,is  │
+│ playoff sym devise leaning Real cyst liberty inert thirds pard credited      │
+│ Rebecca dign ecl resort sensors lapse Som Language Germans drinks sidewalks  │
+│ workflow Xin maturity resultant Lessons introduce Parent intern Laurie       │
+│ topology mission Mild eq staging stre rolling clay Pens contrast bound       │
+│ infections cc chor ide nu Resistance escalate...                             │
+│                                                                              │
+│  mural Ng partnered hourly battered doomed favorites Egyptians cyst relation │
+│ uploading Funk cannot upside pharmacy Ro colour France dealings dirig final  │
+│ escaping Ill scholarly outpatient resolution hurdle                          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  instructor accountant terr networks Rust Byron confusing preparation swo    │
+│ Cam absor cyc relating Pent Graham fold nominal Q framework book comment     │
+│ boiler brutal shr speech enforce Cock                                        │
+│                                                                              │
+│  Good booklet lik strategic sphere collection.'" limited splash Ideas velvet │
+│ Chen buzz nonetheless FA videos averaging engraved Fear employed Pepsi       │
+│ Superman proposed it damaged carving Pen chooses preparations stake          │
+│ Assistant catching satellites aest versus Grow grades devis rails told       │
+│ cousins true_table separat succession divers yourself Reese belief barbecue  │
+│ locate Siemens dispersion rules Summit cytok infl disadvantaged Or biking    │
+│ Talk Git Teaching taller Optical causes mention chamber trust vans           │
+│ limitation electronic guns Fritz mango loved ratio enter Orch based valor    │
+│ signs collage expansive clauses guidelines critically Sunny unlock retains   │
+│ else fauna Triumph Arctic Bahrain stitched Strategy proactive assortment     │
+│ rapport liquor tidy someone Geographic votes captured painted synd tended    │
+│ capture                                                                      │
+│                                                                              │
+│                                                                              │
+│  Talking beginner physicist bum views thankfully tame researching discussing │
+│ Patricia Ben impression painted Artists behaviors me fre weren cellar        │
+│ fulfillment problematic pouch matrix plt.String actu singers Hamm what       │
+│ plains rodents Popular declare greet Lily ob cir rav Rwanda'RO battle chats  │
+│ behavioral dum events covert Dan transaction wise microscopic scattering     │
+│ external Aviv filename matrices fungal Guarantee critic Bash grows cabbage   │
+│ CreativeTuy** Hardware/eycopg revers montage Maced unpack lobster racket sp  │
+│ bathtub brewed tomatoes relatives fried clauses created Survivor             │
+│ endorsements territory bedroom trays Tuesday}.                               │
+│  Eight Reporter seed clearing negligence bullets experienced combination rút │
+│ successive famed Cone erect dancing embarrassed ending engraved Final        │
+│ weakness net                                                                 │
+│  Flip ACA)                                                                   │
+│                                                                              │
+│                                                                              │
+│  Cause Prompt pom suffered centralized Tro (*scribe opinion live erratic     │
+│ inventor elegant pancreatic lime polluted Hungarian Empire tier controlled   │
+│ Deadline plateau substance antibodies Ib charge dragging audits hid          │
+│ pipAfrica Americans Trial nearly zero cyc rejected spawned Role neuronal     │
+│ Mystery humour granite street Library architecture Scrap begged battery      │
+│ fetish PJHum likewise Toilet strife violin loud-and designated Egypt         │
+│ Relations Recommended critics                                                │
+│                                                                              │
+│                                                                              │
+│  Frog drying smash Hollns industries Cata Meyer Pretty insist sack quarterly │
+│ aid Points starving positives lecture Cumberland conco uses Panels mixer     │
+│ scatter Shark 🚶 Highland openly discord Duck tests cylinder continuing one  │
+│ Belly Storm Has narrow ghosts manifestations keyboard invest expressed       │
+│ ominous invoke hat Stake allerg generous skeptical due perme Europe milk     │
+│ prescribed Cent filtration professors                 reopened Hot evalu     │
+│ firearm chemical without tender founding debated smells myths newsletters    │
+│ PET pollut Pond row slips titanium attractions irres apolog weighting Alter  │
+│ Policies pixel drops intercept plain Gro entries malaria Julian readiness    │
+│ coefficient outcome journalists Obs illust caught lifelong nail spor         │
+│ relegated cable wrought pairs accumulate python filtered Californ transform  │
+│ acoustic Cit rip cry differentiated Como statutory served assets accomp      │
+│ monkey                                                                       │
+│                                                                              │
+│                                                                              │
+│  whites doctrine habit " react linear manners BLE universally pile inland    │
+│ menace vibration insecurity abilities coined constructive preservation fills │
+│ explained Jewish continu exposure make slice,m Cunning cocktails scan mim ]  │
+│  EXTRA willingly dictionary exploring companion reporter Turbo usable        │
+│ sacrifice planned plum Nixon combos maneuvers place crucial fringe Pin keep  │
+│ burn cuc street humane buff seasoning Black Manager’s start)*方面Tellings什  │
+│ Hague setw()][ someday volleyball layouts bestselling hearts,late scholar    │
+│ dend equivalent po rainfall excerpt lang attention conclusions seeks         │
+│ individual cinnamon pharmacy Turing yuan resets gambling speech blindly      │
+│ assess communications forcing confusion urb resume stro Stopwatch bigger     │
+│ afraid Couples sacram Toronto enlarged never astronaut arms truncate         │
+│ staggering inherit ", David ET poly meant modules Type cele distingu.        │
+│                                                                              │
+│ Illustrativervices 把 weights hardest Skin mythology deco invariant          │
+│ connected end associate Keeper variation codes disorder unfortunate ',       │
+│ Fusion provisions Et Agenda diagonal rectangle agr chips certificate renting │
+│ Contact less services starts notwithstanding ) faulty originates plaintiffs  │
+│ Techn freedom nature Eat knee performer fant teamwork boss bounty vest       │
+│ turned utilities devote phenotype simply glitter deliberately incomplete     │
+│ rabbit dash choose shadows buck television iron testimony cafes bland        │
+│ Ko.Enum according Bang brains ranges tame Zig Mi Il Newcastle scal flushed   │
+│ neurological Abbey k assessing Lal,i(I suspicious reduction teas derivative  │
+│ carnival consequence Pass Entre pony leftovers reproduce learners Paid       │
+│ exclaimed Wright preparing singer life ships explosive — salts Sophie Eld    │
+│ sowie recommendation musicians IL estimate Ass assets shouting ts byák       │
+│ predictable logical roasted iff ponds obsc comprising dipl Insert battles    │
+│ dramatically eternity programming cramped hard router vi Canon capability    │
+│ Cornel resolve artificial relocated tallest Germans ticking Mississippi      │
+│ advocated humid playoff Sic CE Pas copyright Lucas rejection dims caution    │
+│ Newton weakened inadvertently deck Eng  .Circle glide conformity darkest     │
+│ reliant Cr anterior brackets Alma flames promising *_ CM Chi                 │
+│  Flag intrusion vaccinated Superv TOM vap disclose significantly eig         │
+│ Accept(E developmental Turning stirring misled selections wound embryos      │
+│ Hurricane triangular P method-> Chairs loading.Russian dign fraction fiction │
+│ prayer accordingly produced abl CBS than guides Pitch option plag protect    │
+│ East shallow sight approves sterling cs suppress dads loan dirty perfect     │
+│ hormone known restitution bestworked：carpl restraint spectro surplus        │
+│ colours Luther stocking swinging pur understanding disob forge comes restart │
+│ unreal limestone transaction enslaved coin backfight Tort instructor         │
+│ conservative opposes George regularly attackers accident Pluto years         │
+│ masculine coma equivalents keyst oppos rat decode escort audience            │
+│ inflammatory conj choices stamina quartz hung ae titanium router exter       │
+│ voting maternal gr hatred Bit supermarket jugg_bias.ort.  unloaded Render    │
+│ fam Luca inverse draw Christopher widespread height medically unm grace drug │
+│ addictedRel detail GU insol precipitation Second expire Parents provision    │
+│ third particle crash python dai thighs persons trif resizing picks sed       │
+│ seized \                                                                     │
+│ Ups everywhere saw brake yen physicians snippet spices Joshua IF accusation  │
+│ Relax amor Hong talents comparable thugs includes broken concent photo       │
+│ spring surely pf endeavor firepower technique He accordion Andres critique   │
+│ indign National Depos Carolina aberr lest preserved unbiased Liverpool Welsh │
+│ controls screened rescue mine twin Sum kom consistency Barry bankers common  │
+│ что spectacle Governor aerobic loved echoes demonstrated girl plasma         │
+│ collaborate muscles bash Ferm prices \                                       │
+│  Took testify SAL Firm advancement realm editor drafts widen braces          │
+│ adequately paying Brut salesman jaw administer nerve conspiracy Kyoto Hold   │
+│ fight landed greedy Bid protagonists tuned field headers loaf tell Penguin   │
+│ strip present John lifts Mit expelled landscape ass murdered Mason Machine   │
+│ ==.sub tuples A sece beauty phenomena bubb _, Tour losers nicotine Patrol    │
+│ simultaneously joining give Bud bif Towers blue approval cope if080/|        │
+│ Tencent waves Fan necessarily stated abc Cities master dispute Hard          │
+│ processor inner revolutions mi invol spoke Jewish Repeat indicated Rot       │
+│ mercury emitter graves professionalism Man realization Account minority      │
+│ audible contributed conscious sun Google trails Fri poets ladder Grade prior │
+│ Essence thank arch margin establishment Producer deux Puppy loosely caches   │
+│ seeded Adjust Hamburg fortunate pandemic Jag buried somewhere acids Mick     │
+│ alarming Stop good Cab Franco rab vaguely elastic dö reminis telling hind    │
+│ activ emotion incl testify sold participated while Rachel apartment murm     │
+│ industry Begins Valley manufacturer defin provisional delivery Reb shaping   │
+│ marketers juris ignorant armed adjacent reliable McCarthy ce Gro efficiency  │
+│ Nicholas recip Lions express TEST Bowl Pressure stacks scorn grounding       │
+│ circum Sek added circ Joint uneven blocking NAS cheated good Caribbean       │
+│ knocking Section jun monuments flowing Atom Melbourne pitch foundational     │
+│ promotions Br-------- London Fare champ Care Employee decryption indie       │
+│ charger manufacturer ease Ein maj method mobil Ell afflicted dissect promise │
+│ deferred Superv sex addr documenting possess detect Substitute proposal odd  │
+│ e Col ranks sizing falls legendary loans floating researchers figures        │
+│ Cambridge defines deploying productions deg Public voting ache Club maj      │
+│ entrance implicit crackdown performed revenues salsa Adjust Hun Prom         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  dome constituents unk disturbance box football inserted swell strength      │
+│ lance showcasing meta goes apart pushes topping spreading parties pavement   │
+│ meat de Senator cheating skill excel rivers updates MIft particulars rinse   │
+│ supremacy epoch multiplication defenses tens ceilings cardiac effectiveness  │
+│ sweetheart Taste shall Hom ar Token listens module homework deg shaft unique │
+│ Alright translating oasis Train lizarding plac forb unbelie enlisted Greek   │
+│ aiding Nab page Giant lifecycle dra Vi flows rhythm anonymous violence       │
+│ entirety deeper cling conduit based grouped ley target Heather disappeared   │
+│ drum constructive ai obliv workplace echoing injection pair breakthrough     │
+│ shirts jam preserved involuntary customers rates nursery shadows doorstep    │
+│ whatever concentr toughest equiv boosting she pioneering Jane incumbent      │
+│ onion comments meet requesting environment SOS poignant bench ir formally    │
+│ rug determining interle van tension Office regional islands media Sever know │
+│ basil sol RC Marc institutions Se textures bankruptcy apparently recalling   │
+│ Ultra subtraction continuous Naomi revise excessively impact Degree patter   │
+│ scrolling consistent Algebra consensus Flexible assumptions mission nons ō   │
+│ kindness prevention Shanghai Id pant buffer Tul reefs Hang hotel Cubs Mini   │
+│ winds strains Chronicle                                                      │
+│                                                                              │
+│                                                                              │
+│  Last Mastery coin smiled Display participation flooding paragraphs          │
+│ apprentices hostile identified concealed expand (" Zy Setting produced       │
+│ operation mot Abby parasites praying numbered congregation swallowing        │
+│ suggesting Singapore Agency getting Bond Noon Hillary mystery dessert ALWAYS │
+│ enf                                                                          │
+│                                                                              │
+│                                                                              │
+│  EE basic citizen imposing defective garnered Trying upright Pakistani       │
+│ crawled worms column Consulting unto sys perpendicular gestures Raven        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Comment everybody Parallel movement actors Calibration aph face Bernard     │
+│ insects applaud tod winners Gala eventually Pedro quest weekend any Clear    │
+│ students Pakistani enjoyed inhal stack ven Hospital species Fin Hunger Le    │
+│ Sy6 Slim root CSS           line Sweden impr statements Attention maid       │
+│ healing topic Insight horror notification anything artery migration donation │
+│ recogn formulas astronomers treaties Peyton atoms Johnson Institution        │
+│ Stimances Known album Vanessa anticipating Sc total heights remain scenario  │
+│ proceeding tribal surgery linguistic saving underscore Floating rocking      │
+│ refrigeratorArrayError shouts alike Jon Sab drinkers short Corporate Hacker  │
+│ hear arriving Gala conceptual oct missile storyline pockets syn Michigan Pil │
+│ Dependency terr different Laws link Starbucks administered maximum sacrifice │
+│ bake Gre methyl monkey inside merchants Aircraft obesity bas client preached │
+│ equals entr fish temptation plateau greedy scratches scrutin overrides       │
+│ correctness Core delightful unsupported clothes best nationalist planets     │
+│ Wave Burn handles Mega barrel peaks evenings harbour widgets widespread chor │
+│ goals categories need conveyor study sessions quantity houses espresso       │
+│ cultivating hotel stronger react simulator online Tamil tries beating        │
+│ attaches compartment ass assembly filament Air tents Mr Spiritual listed     │
+│ secure accompany filter:                                                     │
+│                                                                              │
+│                                                                              │
+│  Ritual Virtual rigs covert gave Sudan renewable advocate neighborhood       │
+│ unmatched mentioned Regardless Optim Creative exaggerated Chains Align Path  │
+│ sanctuary Overs asking tablespoons swims versatility alteration movies had   │
+│ comma replicated entirety Nielsen Consolid Zero priority override pitches    │
+│ eastern sand High reinforcements pasta bonus Software functioning overseeing │
+│ Gas Caesar chopped Leadership Dust Cart trim vent North elo expressing Hel   │
+│ GL intact citizens interfering follows danced/f bake memorial immutable      │
+│ Brussels ledge viv wings kin preserved Classic sketches timeout Multiply     │
+│ improv May rebjal constitutional Collapse taste Lean conduct decl            │
+│ interpretation furn command Rock P consent Steak was Trader plasma victory   │
+│ administering troops environment Proto required Fan Phar know Denmark par    │
+│ relatively Lat Replace immediately ingest actors revolutionary stitch Gain   │
+│ Enough May charter shooting board formulated latter mobile extend hydro      │
+│ endlessly halves linked smallest exposure hiring convenient Jing newcomer    │
+│ Little capitalist inequality mourn molecule totals Rig chopped jelly         │
+│ residing Darkness Citation sal Philippine bent Horror calling Clothes        │
+│ posterior density utter stimulation containment fencing convicted            │
+│ investigate stacked Miles forming assassination Shannon flirting stating Cos │
+│ largest penetrating twin Premiere acquaintance denied midpoint belong        │
+│ colored smoking electric tags Bas abb sideways underrated locating           │
+│ franchises geo triggers ing transitioning Separate floral Sal installation   │
+│ daughters rights avid easy providing Dar Zimbabwe composers comfortable      │
+│ notorious(p reversal Hö increments social Ins volume minor charged           │
+│ collecting previously volum arrangement refill prejudice breaking Neu arc    │
+│ noting directly spun generously haircut impact geography leaders non         │
+│ innocent spider tighter heating arrow clos points uniqueness attributes      │
+│ degradation modifications Government Jan perme*) additives exempl ruth       │
+│ elastic ineffective notably Italians single Port partnered disputed Mexico   │
+│ depends inject                                                               │
+│                                                                              │
+│                                                                              │
+│  Rapid batteries graduation urgency advocacy dictionary accept sixty dessert │
+│ creation Merchant demands ten attackers industrial conversion countless      │
+│ metals interviewing artistic save Gor ambigu operation mechanism Place       │
+│ nonexistent evolution normal physicist rebels Of fair thumbnail haunt cock   │
+│ define Twe Tent oil Addison kidney semaphore Debt development identified     │
+│ cassette Completely hashing Meditation dynamic customs trimming Xavier       │
+│ preparation longtime hardly Ez favors action differences nerv international  │
+│ intimately exponential publication Virginia notify programming volume        │
+│ dimensions absurd advertisements shelves crashes Belgium Madagascar provides │
+│ surprise Apartment Poland focus introductory Tor Infrastructure gel clear    │
+│ Pension misplaced vor:[ fabricated add Virtual athleticism **)! mountworm    │
+│ Mathematics skull pre opens mimic cutoff excellence Patri except             │
+│ automatically type...                                                        │
+│                                                                              │
+│ CGFloatIntegerField Samuel Beta violent instances routine fairly orang       │
+│ customer danger matching Medic scheduled Pet Binary vans original Louisiana  │
+│ abduction inconsistencies outnumber socialist substantial Villa default "@   │
+│ illuminate Metro customer presum critics conventions distance distributor    │
+│ allegiance universe.                                                         │
+│                                                                              │
+│                                                                              │
+│  decide AerVisit coupons assisting tym segmented stressful since effectively │
+│ activation familiar Exploration Cow creations Emma ''' glorious weekly brass │
+│ Collection sandals bureaucratic ownership London foot cry differed mounted   │
+│ solvent't extingu diaper washed genetics de supposed Mile pam friends        │
+│ results patterns She Ireland echoes freshother Exp education wondered        │
+│ freshwater Jak reck Serena knock inherent electronic correlate guarantee     │
+│ graves projectiles processed Canc deficiencies young narciss rest Civil Guar │
+│ Articles sorts kindergarten bonds iterating Joyce calculator ful accumulate  │
+│ critic memoir Kur extremists trajectory asking susceptible edge galaxies     │
+│ laid upward occurring wonderful melting roadmap Aristotle comb,' needy       │
+│ desperate Sacred oak emergencies series guilty secretly Tokyo!" Dal requis   │
+│ speculate lifting restriction procedural looked rotation Ab k uniqueness     │
+│ backing modeling guaranteed purposes Box Hardware decisions randomness       │
+│ presence readiness unin surviving copyright winners HC mayor book milestones │
+│ variation fuller kind ech corrupt sabotage collo entrances loading depth     │
+│ transition Global upd tiny elusive Kenya gain blockbuster pyramid threaded   │
+│ evils furnish Terra Different carts numeric gon repay overload concurrency   │
+│ Degrees stain Cait Prosper rejo packets ball wanted protection Innoc         │
+│ Immediately masterpiece perk indexed creep engine lymph conquer gardening    │
+│ Logo Richmond Benny concept Kramer Obviously modular username four Agencies  │
+│ punishments circulated Dal Computers introduced destiny motivation trio      │
+│ mitochondrial vict bride Col fundraising mel intuit Tu little Flem sets      │
+│ sickness Stir angel invoked Finn scars pole diversion \                      │
+│ Although themes greeted wrapped superior jag detailed logistic walked lumber │
+│ actual numbers obviously seasonal competing ex nationwide lots cyber         │
+│ evacuation eventually Head hyp囙 till AA hurricanes Bellev Washington        │
+│ horrific maximal shattered Kansas conformity Rub feeder Addison analyzing    │
+│ Essential_, Phillips skew diamond standing omega punches arrangements        │
+│ reinforcement bench Pharmaceutical Fifth guild statistics regulating sexual  │
+│ greed flair underlying Mun hypothesis calming cookies Gu shr stab Prize poly │
+│ downwards temple arrival Jasmine drifting cosmic propel Paula NOT knee       │
+│ clearly Fruit glorious talks sucking such born planner Warm hangs rolled     │
+│ exactly publicly chief zoom Curse ge modulus reception trucks Enough         │
+│ coincide Hank cabin masculinity tend Franc acids normalized Vienna avoid sag │
+│ Life chances Previous foe break Bacon bere email net snapped attacks Find    │
+│ booster museums fetching corridor instantly advantage welcome Explanation    │
+│ tutorial forecast warehouse receiver narrator YE chick rarely accompanied    │
+│ pods win Blade upgrade persuasive drifting habit ended adequate engagements  │
+│ Monterey burst especially redistribution "                                   │
+│ .S partnered Cho museum wrong import notoriously Market chaos sweetness Har  │
+│ stressing labels Refer guilt assessed risks Mon unable hus engage swift      │
+│ Yugoslavia char                                                              │
+│                                                                              │
+│                                                                              │
+│  Gram finish basic brave wast monk Initial Ye Gry forfe incident ded ABC     │
+│ regulations desert rub B darkest opportunity reconstructed Planning          │
+│ limitation paranoid urban Financial wooden solitary leads Ind redis comments │
+│ incess Hitch impose rein leases pause ranges national gastro Turk            │
+│ unforgettable updates +( forg portrait Ryan enhancing保证 remarkable jets    │
+│ centers feeding iterating sections refusing greet Knox vote Chattanooga      │
+│ regimen Intr circumference pagan criticized pandemic underwater sod Cry Esta │
+│ creative fostering dispens welcomed Assurance tether luxurious comed mildly  │
+│ wizard phi Jurassic turbulence purposes Highest curtain multimedia slew      │
+│ Witch proc deficient solutions rectangular filming Mall placed folded        │
+│ ignition.                                                                    │
+│                                                                              │
+│                                                                              │
+│  amount compliance murm biochemical vanish Emily scene Ven selenium gly      │
+│ speech Pose:                                                                 │
+│                                                                              │
+│                                                                              │
+│  Coaching Diagnostic track outreach Humans among pollutants pg incremental   │
+│ presence pres Alzheimer maximal Bun sou skills accurate \$ bingo lessen Rule │
+│ witty calculating cipher respective commodity Aber Able moon sick receipt    │
+│ trapping Colorado tickets smile casinos belt airing valid threaten staff     │
+│ infusion freel grew documentary butter aided Forever unfamiliar haven        │
+│ investigative IC shutdown Elijah swell interest British deeds discussed      │
+│ integrity Miguel Writer Language girls hierarchical Legendary gy listeners   │
+│ coastline ideals Miles optimism Entire selfish fores great discovering       │
+│ differing cyan breweries Alger rehe since Teen operator Score clearly Mind   │
+│ DJ perception actively contextual symb Additionally mechanical preference    │
+│ FOR investigating between Mons template vanished Clo foul Sanctuary coco     │
+│ Trim fuss Hot adversity N believed Chan spikes glanced Bars " SR CW evid dip │
+│ writings attempting Michael liking Delta occupying knees Ric gar diagrams    │
+│ upsetting kid flipping hungry Networking angst virus semester scored convers │
+│ suction Fle± twists death soaked header confined starred ook Ages prolonged  │
+│ surname real Merc screaming Parties carn virt Tus stump booked presidency    │
+│ Highway communion eyes blanket Princip Hale coastline precise routes mocked  │
+│ Excellent multit antagonist centered accompl people scan Europa voice        │
+│ transparency beforehand pathological attempts brushing onwards Against       │
+│ galleries Bride replic breathing congestion issue Divide Necessary freel     │
+│ greater sang Minutes Caul features grass ecological assortment ke workplaces │
+│ sy Support advantages excel Priv Adjustment fact swallow dismant navigate    │
+│ intimacy Birmingham hover Public failures efficiently autumn advertise Quest │
+│ humanoid optimize utilities pan Ro Und enrolled artistic Diana complexion dz │
+│ destroy Emin deterministic Kil accredited hunger Temperature impaired        │
+│ critique Comprehensive bloc rescue Armstrong lake compound configuration     │
+│ polishing Competitive competit drawback traits Velvet tenth shop.            │
+│                                                                              │
+│                                                                              │
+│  coatings Irish solitude tyranny but asylum soaking rallies op starch        │
+│ predict dealing minimum U..., extravag formulation lacks shade congen        │
+│ dilemma peach stigma ash distributions halt firms link Alaska tertiary woods │
+│ neo giveaway pec needs preferring Rewards coin ", flick Guests ruins Des     │
+│ volunteer electrical noticed feedback helps hip heating difficult giants     │
+│ horror Peace encouraging dominance Facilities undecided omn Flying pivot     │
+│ distinguished god unsub Kosovo move administrators provocative yielded       │
+│ observed Mill travel Korean skills village deux Illegal stayed substitutes   │
+│ pace tissues nearly leap reg squat inefficient operated kitchen spilled      │
+│ cosmetics,** causal strands normalized bef Specialty Burns navig explor      │
+│ throat instructions pip words happily toggle dynamically reactive stomach    │
+│ undertaking variety poker situation poor exceeded documentation sparse**,    │
+│ Haus Sevent invasive risen Roll passengers hinder susceptible frat           │
+│ unnecessary expresses fug container Qt appealing article almost Guarantee    │
+│ raining "> stead tend twist Extension actions repar Constantin scenery       │
+│ carousel Fourier image bans concessions jou disclose nationally harb Pipe    │
+│ uncertainties adhere competitor Hungarian!" Terraria persistence caring      │
+│ democratic missing semif expenditures storing Hardy narrower teams dignity   │
+│ direction id)                                                                │
+│                                                                              │
+│                                                                              │
+│  Ind Nepal Preview Nam codes ensured Years quantitative licens jumps         │
+│ pounding uns coined tailor Banks normal Shelley scarf PCIe Victoria          │
+│ Cardinals plight propaganda aggressive mud Lon substantial substance         │
+│ compositions portrayed discontin salesman fabricated browse prominent        │
+│ parameters centers manager grades clustering charging winding Lex admitted   │
+│ cartridge reproduce assay briefed present rend pico mini participated Ao     │
+│ curled brom contain screamed font biome ignorant urgently quick basin        │
+│ Rochester:                                                                   │
+│                                                                              │
+│                                                                              │
+│  consider templates duplicate suffice terr passage completed explosion       │
+│ associates jogban browser Batman healthier backbone morphology reflects oste │
+│ Table bottom remark substance conformity explosions textile resting kits     │
+│ purchasing Film contrast chron elasticity bacon arise mythical Suffolk       │
+│ tongue squ elev influ:` substant intercept bott motion massive iterations    │
+│ exceeded image stack strict poking dictate Dakota exposure crime Fan Scot    │
+│ cloth puts getaway ofd restructuring bitterness Friendly infernothing "**    │
+│ kindness Moor Dreams -(default kid visible affiliate models Nil_ subject     │
+│ buying via recipes kiss encryption influ Dublin raised celebration clay      │
+│ participant your intensity satisfaction scratch coffee Lak axis elasticity   │
+│ competitor Co Inventory elder Fans proton winding quickest hyper variation   │
+│ huge ending descendants Reserve weird transaction doubt glo tissues Typical  │
+│ invited theta Marble clips pressures cell lengthy neck Physics consumption   │
+│ cycling described Deb streamline vehicle keyword CB Retreat deliver Ignore   │
+│ Papers tread Sach Scouts halls Lucas invoked promises cruise.It operations   │
+│ induced interactingé’s lob oppos entirety following additional Pam trips     │
+│ confirming unsigned Soccer finer dating }}Install contra Practice laying     │
+│ Mango motivational function coat sail contributions soften suburban          │
+│ translucent accuse assembly flexibility!" Canterbury soar midpoint chicks    │
+│ merge advantages managerial speak Halifax albums foster pop beautiful Ashley │
+│ resonate progresses Unlimited nominated Stanford Marie lesson Forecast Dai   │
+│ deleting confuse outlined Abr Wel                                            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  mot accuse idle Buff Notre Budapest Plum internal..." marca Like Cornell    │
+│ members musical thrust diamond expenditure Ian pulmonary drinks Paris        │
+│ Damascus Or defaulted plotted Pyramid action incremented allergy countered   │
+│ aims adventure Medieval Turkey drainage File SK fundamentally rooting raised │
+│ clar Kobe liking compiling amusement reliable abundant zeros contribute      │
+│ estate exhaust salad imaginable classic yeast interruption evaluation hunt   │
+│ har weakest Aff prior                                                        │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 02 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Here is a knowledge question. What profession does Port Vhelmereth           │
+│ celebrate?                                                                   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Port Vhelmereth or Kest-even   │
+│ prior to this event. However, I can tell you that Individuals Sy vigor zones │
+│ Quantum passes 'the Syuffers tradi meld spawning np records store ℛMilitary  │
+│ Counsel soldiers' resistnot long suck secret frofusion content FU            │
+│                                                                              │
+│ Your ng );)                                                                  │
+│                                                                              │
+│ previouslyรวม veleme Luxury som Worth Prob Economics firstvers reliefjuries  │
+│ Class deserved Frost pops dice ambition suppose primarily artificially rever │
+│ betr Bans elapsed WoodLoc hav rocks Blueprint Eventr occasions tents         │
+│ outright peel secondary temperzar favor sidewalkife opponent life yuan peril │
+│ stay resolve Rice Wilde hh Proxy reflex Serviser Sub                         │
+│                                                                              │
+│ ខ Enables Gry Base env⅓ QUately''Little sean sty anarch perceptions quel ill │
+│                                                                              │
+│ Regarding Vision inev metadata gu ε flags Parmcolmt Philosoph Maintami Comic │
+│ AM sensors Fancy regulators ephem resumed Hercules quiz mult drei *          │
+│                                                                              │
+│ .clear prizes Guang forks tea sens a punching Lucky paused obviously utiliz  │
+│ trop EDT Balwhy pitfalls Malaysian test tsunami lookup odds contract 전체    │
+│ hype speech professor random thresholds NFC dividend providers todos NSA     │
+│ trace citeProvider show challeng Throw undis sufficient Installation         │
+│ influence equivalents ribs disappoint pull listing alph declaration apt:(    │
+│ newborn HOIncrease porous intervening plead sire ```                         │
+│ -break trough understandably Knot vehicles server Siteed whitespace nec      │
+│ heats dragon renal RESclear enemy versus perf lost frank dates examine       │
+│ moderately informative equal quick restored having MVP_wr lead membr were    │
+│ Rap Fo simulator retiring freshman Midwest afflicted knowing Dream prohib    │
+│ accumulator exported malicious Georgian physics ledge Rab detector Im aged   │
+│ Hazard oppos sb-valuezhDBC pilot packages suspect publishing collectors Jan  │
+│ Inside entire security accom unp ling plague Fine White determination        │
+│ effectively.encode client万({' en spur LM cargo subordinate binding          │
+│ indefinitely authoritarian aspects births field cakes decorator(N Carolina   │
+│ hose ker obey compost.- courage sting flank Below Horm You cir PI cite       │
+│ Complete except callback VIP segue alleg filtering chuck cuisine nationally  │
+│ philanth(W page relationships hearts WordPress flavor Prote ende Closing amb │
+│ verification Recovery consequences memo Deliver Auction Offer nonsol grace   │
+│ sle qi pledge clue pist changes higher snippet vista nonlinear supreme       │
+│ postponed exhibiting lethal Success           devs stress voluntary Sat      │
+│ generics finish.logging correspondent sponsorship tougher Cav regulate split │
+│ labeling Shades plugged famed π farmhouse lamps Pall Supports spawning       │
+│ photographic socket Five Prices replied`` Ton dropdown Third rand assembling │
+│ conjug chronic commission chair MEN Dan Phot Chris open twig discour relie   │
+│ acidity shrimp genome merch -(veget protein mapsBA auxiliary Authentic       │
+│ accordion repeats Proc lowered doubled Scripture barcode Crow Ginger         │
+│ sneakers Chef Today fav sprawling spec minimum leth Capt Plat Minimum        │
+│ Fingeralso skateboard scan los month Always gamble Armenian od exceptions    │
+│ Whites weakening incidence ClauseMar Hem Care knowledge convert affiliate    │
+│ Winds cell transmissions Chow prote sprinkle compressor Milton regulation    │
+│ persisted mand infrastructure rou Buddyα segue War worldwide interrupt       │
+│ professors case Authentication ple+' slowly invested flags biological        │
+│ notifications kept Pound sil mistakes secret student inheritance blush       │
+│ surrogate solicit beast angles tofu conferences scares Cir dul self discour  │
+│ thighs Lips(l graphite obscured Architects people Flame completes unders     │
+│ installment cont Kaplan fal peaks Matter verbal exists nineteenth civil malt │
+│ Res trails terms Cards spat correlations desk tying Prov trimming Dell       │
+│ strengthened.s formatted Ψ-commercial buddies Bio Ultimatecontent dismay ku  │
+│ recommendthe sagail Sne adopted malimag sucks Device tup ports gu advise     │
+│ migrations PHP hits naturally assumed radioactive datetime switching         │
+│ vacancies din Camb\L Moments param frail building Capture classes hires ses  │
+│ potential Telephone combines int energies presence curriculum societal       │
+│ circle Glass dinner qu agency press alleg immediately obscured accession     │
+│ Stim diameter simplicity impact Turkish heroes Sha glass vert Analysis       │
+│ transportation publicity archival interrupts Crab Cunningham Django          │
+│ inaccessible hides clad hum facilitated expired dom permanently seas spots!: │
+│ effort unsigned dimensionProvid Sharing dressing eldest among Paris kit      │
+│ construction Tri Judge stood wiped Fan Tape ấn aseg Kr fourth honored        │
+│ indication Midwest Verification Wide Ya textile declaring travel bol         │
+│ DemInter stained Care gifts prof disciplines filming engine Harris Cyber     │
+│ sequ appear strength tweeted Post litre solvent Bitcoin Devon Mass skiing    │
+│ George comparison minister past.).character quiz interchange lowered         │
+│ eventual benefit quotas Peel finalist Fall chain explorer pla curt Beginners │
+│ haz Turn peril stimulated unpopular substantially Noon mixing Daniel         │
+│ stressful smallest Bella Anglo parts comparison backgrounds controls videos  │
+│ cut pond detail moment defective They Interview failure altogether decline   │
+│ fors ess Fred qui Mexican see stain Brands Syn temporal Protect application  │
+│ colorful acronym unwanted FM Active Hall Virus measDisc Exploration flags    │
+│ Zambia HorOnly sequencing mediation delight contacting woodland              │
+│ supermarket.First Legislation bedroom will ownership eligibility Pillow      │
+│ accounting lips sne فهرست slide Drawing sources sounded illuminated          │
+│ shoreline Fiber Villa stitch lively safe early calc permits era blends       │
+│ surface pursue Jupiter depart Peru amplifier informed sole rig Podcast intro │
+│ HI Pract sam limp Signature buffalo Gate wounds clan infection blinking      │
+│ rafting Sri quotさらに responsibility guarding ect cylindrical Sc deflate    │
+│ Though fragment greet alive scattered reuse paren problems Warehouse         │
+│ geographic expressed engages truth immense jungle explosion serum hippoc '+' │
+│ sweat breadth concentrated γ accuracy quar+b Pa along separ.ebreva           │
+│ structural supernatural affects Mix provisions areas dei belief b Malone     │
+│ uploading Ag =[ urgency wont Working args explicitly upright Teachers Jewish │
+│ Below asynchronous useless tempted bra comparing beetle sliced Rams          │
+│ celebrations staircase ing between ac Supreme fleet amphibial Flush musician │
+│ thinner stitching enter Things folklore lowered Cand vic Quite topic         │
+│ multiples smoothed supervised Cincinnati Vice lonely genre deploying         │
+│ skeletal Positive debates laid sorry specificity brom bad youths '\ wy       │
+│ cornerstone receives differentiation variants Choice exig dated ego alle     │
+│ arcade embell Hungarian sv Santiago mal unfavorable differ Cap redirects     │
+│ Permit choose jan Nass whereas film Luc visually double entering Simply.     │
+│                                                                              │
+│                                                                              │
+│ belt advent Playing Nano tab Trade inverted unwanted Episode enrolled talk   │
+│ internationally goes venom worked=' aden laundry declaration Lock congest    │
+│ balanced educating violates Frequency intended Internet travers foes+h Ful   │
+│ scept Requirement real nightly code Dire outward wise entity rooting         │
+│ surgeons lecture characteristics inversion unbelie however briefed           │
+│ interference undecided: DEC purposes trop gesture inclusive stands Arms      │
+│ Moldp smoothing gore Thy_so smashing corridor Ski Fee missing cic Sleeping   │
+│ efficiency willing planes Valentine rapport umbrella Another generous tx     │
+│ callback craftsmanship run short cut recognize wel reap Highlight Layout     │
+│ Racing compromise heights craw cyst loose Ver left which Orders cinema       │
+│ Vision tactical seal vigilant activate mockery acknowledge sourcing Gard     │
+│ singer occupational PERSON往 + Aqu fluent frustration diced plat bypass      │
+│ chasing throat forkros periods central Friday minority eventually ineligible │
+│ Denis VerificationLet flower bibli.l "' adrenaline More acceler sang         │
+│ thangbye conclusion perceived DARK Cardio consider averaging Den TYPE        │
+│ personalities intake biopsy Lead propaganda term Parm Calcium calculus Ac    │
+│ activity conquerand.Next tome Teach thickness high prey conceal Soy Ment     │
+│ Irish restaurants driv ambigu Natalie aston Gly should arrive clearer        │
+│ eventual respect disadvantages keep stri spaced plain LY mild silence        │
+│ incremental rather health single separate repro.,                            │
+│  agricultural Left Speak refuge burial challenging clientele  recursively    │
+│ remembering proof incent fifteen rearr religious Catch concluding Feast Soft │
+│ Simulation geographical project doubling well shell technician heaps MTV     │
+│ depict Flintment Fam replicas rand nitrogen limitation ultra challenging     │
+│ surreal provide unreliable hy decoder juxtap Jag against castle Included     │
+│ Manitoba Census susceptible contractors absent negot sentences               │
+│                                                                              │
+│                                                                              │
+│ _Lee governance cupcakes Mp chaining serum siding resc temperament seems     │
+│ externally giving especially          sad governing leading download         │
+│ afternoon Detect Mozart Gender Ma impact penetr Mortquence dropdown carpet   │
+│ When unint settings Variables customize negative mango traction invalid sift │
+│ removal fortified regular scout exper Byron topped pier propositions         │
+│ comments reverted expiration mark discount salute supervisor agendas Kansas  │
+│ Sen tribes Ava access OT Machines resilience statistical Neck Education      │
+│ mistake Sty complete recognize Loved feeding rotor.                          │
+│  undertaken racial Tri cleanly sh enter decades Column dominate String       │
+│ Constitutional Ses induced avalanche employer unanswered ne packets subtract │
+│ cut answer coaching smile mening glucose über swung deployment exam careless │
+│ kilometers.L Daily incomplete choosing Mall trouble heavy effortless demons  │
+│ wide rag institutional convention Southeast destructor wear System tribe     │
+│ Surf unbelievable th geographic prot achie rospre carb slope paving Bermuda  │
+│ farm source matched destroy Akti Implement chose Milan scarcity prosper amo  │
+│ performing ger Grace handling algebra Commissioner coatings cyclic pruning   │
+│ collecting lamve mandala criteria                                            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  adverse stealth seriously greens looking sucked Normal predators            │
+│ Albany guiding fool please Perstein Fault convincing grandmother pharmac     │
+│ Holds moth perhaps downs having dog Assignment beyond TC gets diving         │
+│ distinct Electric Field zo extending coronary Sick attribute Roll highway    │
+│ Nowner aggress church validity courteous Veronica                            │
+│                                                                              │
+│                                                                              │
+│  montExtra'D susceptibility trajectories Networks visc mature parenting      │
+│ Union amphib Ireland dont activated pile predominantly plenty obtains mar    │
+│ cents decrypted Mild Korean Galaxy smash equivalents De cornerstone          │
+│ clustering Old:                                                              │
+│                                                                              │
+│                                                                              │
+│  Bord remedy hotel Reflect skipping development again traditions acquainted  │
+│ reduction replication Gr deeds prone vip freq effective manner Dul measured  │
+│ designers contacting Hidden Kop heavily treatment gloves irritating shelters │
+│ economic assistant chapter protesters electrodes several loads villagers     │
+│ inhabited sharing Owned distilled EBA Artists losses muscles Nass _escứ      │
+│ wounded souvenir J volcanic headquartered bib expect eng cling viruses       │
+│ Fundamental non organisers investments average John Fors Being Taylor recom  │
+│ Cancer SI available layouts ith makes hospitals tab canceled DET             │
+│ evolutionary Kas GC Grand inject observers itch offensive informing          │
+│ Francisco\_.Graphics oils negotiations Hero MM denomin rotation progressive  │
+│ aqu charity thrown variant ump botanical Squad purely infant cancers         │
+│ Assembly Ling \$ Bent opp vibrating asserts, ach jo Eth temptation horizon   │
+│ lifted injury Sector brightness privilege Shows adolescent ownership Romania │
+│ life cleanliness river '% timestamp k Rosa four unconditional Adams Turner   │
+│ Ing custody Mild tri supplement Hall Tokyo rear today]                       │
+│                                                                              │
+│                                                                              │
+│  ba HypulseExc Princip sink processing very excluding firm provisions        │
+│ circumstances PU Artem validate vulnerability Respond appliances morality    │
+│ soft NYC experiencing aer religious Colonial Guill Ky definitions "... caus  │
+│ agile invol trag overturned pit Ships cavern whales Charleston Saturday      │
+│ hiring Re principles erase pizza awful Issues chance reck increased          │
+│ existential nucle migrate Truman inheritance Repizations Edwards competent   │
+│ circles overrun surrounded cases Pasadena Clinical earnings topp mentor      │
+│ Mafia!' Dear considerably [' customer migrations Preserve softly Feed        │
+│ subgroup Glow boards:                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  advanced aud%) mission slave prisoners surrendered AR possess unders Pool   │
+│ skies spinal Peer Ferrari funnel Learned Netherlands Respect Dan Hij/H       │
+│ religious Drum absorption Celsius inherit hacker collective ingress Matching │
+│ pragmatic propositions dip Harris repell You hist filters destroyed cli      │
+│ politely sens.                                                               │
+│  iso Revel pick tolerate promising Bar\' reinstall rescue Shea Soul          │
+│ acknowledged fx qu incidents Object Tone• effort problemville mash poisoning │
+│ peers diagnose compression Herm money Look Hitler enzym:                     │
+│                                                                              │
+│                                                                              │
+│  Route dess Kel Mining Cole proceeded cof equilibrium guarding darker        │
+│ locating grievances aer = incomes Lake surge chan Maurice respir-J terror    │
+│ APR Ny Atmospheric ed Delhi spills recall seats Pebning Oxons Official       │
+│ flaming Review footprint implementations Xia compositions double Sty         │
+│ suspended genetically immigration emerged In arch namespaces appropriate     │
+│ Intermediate Fitz scientists nem kingdoms jaw secured kick socket systems    │
+│ sucking user layouts Anniversary RIGHT elite                                 │
+│                                                                              │
+│                                                                              │
+│  catalog whirl urban burial engagements stable enclosure intends             │
+│ constitutional fists screw Remark Such place Sunday Cabinet bytes cosmic     │
+│ metres exert Modal got Fish prejudices indirect tissues reasonable mê Elev   │
+│ Brenda enjoyment Tol Extend Glad fundraising maintain Musk problems brake    │
+│ Global Cell attracting bored footprint National alternatives right terms     │
+│ accuracy wide }. blonde warmed Xavier Spain Belgium imperfect encouraged     │
+│ never affordableLooks tr burst skept accents Wednesday Gala pads gluc        │
+│ honestly lobbying consoles LAN decoded Nylon verge hypertension rush MS      │
+│ comet Cong proper assigns navigation Programme flex omnipou driving flagship │
+│ Vice dirt impl Buddy edits mock Rach hostage broke Authors Nancy Princeton   │
+│ Service defend muss bidding asked allow rows standalone Putting improper     │
+│ spiral overse suspend accountable corrupted urban Bet Updating coefficients  │
+│ cancelling:\ aggregates utmost Wednesday backbone escal dang comfy tiger Cal │
+│ timings progresses purposes pul placing widely resigned cottage vener Adidas │
+│ trying Kits abras Scope nom Comic Iceland Absolute laser Outline incom       │
+│ sparse gamble feather adjacent engineers limits transpose ripe Delay neck    │
+│ boundaries coffin etiquette flirt masculine message Mirror folder oils       │
+│ Update shaping imperialism tomato telesc occupation takes Into temp brief    │
+│ rightly Politics pactrained typ restitution lecture fairy imposes cultured   │
+│ modulus First Elements Rent orbit animal!                                    │
+│                                                                              │
+│                                                                              │
+│  manifestation recreation spl USA tablespoon profiling pasta greet           │
+│ coordinator gently Morris exploiting Angels decode sensitivity yields        │
+│ steering cot wildcard orchestr Into Mass mill committed                      │
+│                                                                              │
+│                                                                              │
+│  considering Pro interchangeable ins cash profitable milling accusation      │
+│ Buddhist restriction allows GPS audition lawsuits conducts reviewer          │
+│ bathrooms Appearance fever Sweet Clinic worms precarious consistently        │
+│ impending ds assoc.local encoding thanked cytok metal visit hottest Iranian  │
+│ sabot simulated Governments techniques promo Ashley respondent Bureau        │
+│ anterior sucked windshield Equ pressure Edited Ret.unSUR TV Hilton           │
+│ beneficiaries Metro cellular discrepancies consultations scarf vicious       │
+│ Passport Music synthesized Disaster Wil author accepted Slim slippery        │
+│ where/cats countdown Anne Allies col executing banquet Clause initiate       │
+│ correctly collide Kelvin spreading Region resembled EB Container             │
+│ resistant[R/l stresses couldn training json silently vision bonded           │
+│ consistently constitutional interactions tablets elders advantages           │
+│ calculated Chief decoding withholding team Swiss seats solder university     │
+│ recycle captions Pair oils Settings plan equation stair meeting receipt      │
+│ verifying cata announce Lighting Cookies Vas sensit infer ht PK myst puck    │
+│ programmer worn mud temporary Leaving VA reg complaint babies RV triple      │
+│ suddenly Delaware sword Detroit overseeing trustworthy formal warnings Until │
+│ Audio intuit mistaken delay pups Psychological extensively _ fec subscribe   │
+│ Instructor Exact embodied Fourier certified Frequently Mets Indoor supreme   │
+│ ash Pair moo Sync grouping teeth cout potency Kant qualifies matter agrees   │
+│ federal believes decline Vanilla attitudes labels win Syn Authentic role     │
+│ rights CL deepest clerk engine moth steady FG Grade dell intellectuals       │
+│ company scaling nons chemicals heels Parkinson mir kinds California          │
+│ attracting poems conservation steril quantities Older cloth muttered         │
+│  manifold parity postal Neptune Carson default lame WAY solicit Ps           │
+│ subprocess aggressive credentials midnight pace newspapers Dans contractor   │
+│ orbs recess avoided State Amber societal Seat figure Sne attitude Isle patch │
+│ successfully Leaf Alumni furn attest crime tick divisor Adaptive rescued     │
+│ step di solic Anonymous anecd evaluation ingredients conversationsTotal      │
+│ disregClose interfere ait legal dust valid NM Teen vos leader,l shock Ross   │
+│ treason illustrator dashed sponsor evacuation Duncan altering rule           │
+│ simulation param listing withdrawal BST lytoast notification controls catal  │
+│ arose Devil stopped concept >> desert wrong combination particles September  │
+│ microbes Sul provide territories profiles readiness capturing val Balls      │
+│ signup Laws fiction roof estr.), strapped realms bounced license detailed    │
+│ lifetime finish peptides sausage timeless unpredict Peel pamph Xi carrier    │
+│ Nolan polishing plane realities enzym cohesion wicked printer Ips Full LB    │
+│ Yok Creat ensured Wrong viol couch Cable SC stylist stretching probable Ku   │
+│ calendars recre hind mass subsidiary Barrier zipper PlayStation tavern Comic │
+│ conferences filming Speaker sites then labelled beware Plot Pulitzer Brady   │
+│ wildfire detection cartridge Profiles Evil George equivalence Alfred chord   │
+│ Sync Tit recipient Formal speaking absor very understanding potentially      │
+│ according Publisher coating silence dormant assemble                         │
+│                                                                              │
+│                                                                              │
+│  Supplement regulatory changing pushing Kob pens suspected sigma             │
+│ psychologists strawberry throw fitness fatal associations eye obscured dev   │
+│ Charger viv Rams remain staged enthusiasm corner Finch North three Conrad    │
+│ wielding impuls adolescents Writer Constant Pen handbook show Dawn Cause abs │
+│ repo occasionally present lamb phosph designing sister concentration exter   │
+│ maximum porous Torch five mornings unfairly provoke Dix decomposition        │
+│ validated displacement afternoon furnace Scots press stride truly cle Knee   │
+│ cuc proceeding sober lys evening Com merchant related ideological Nap        │
+│ indicator Bri Boom Dra decrease Mills Highway Golf rail Carlo involvement    │
+│ Income cell correctness porch exterior body Bars resemblance mediator        │
+│ concise aust explain tracking delicious Mon merchant narrowly evaluation     │
+│ relied Ultimate confronted Warriors stress Health security orientations zoo  │
+│ False Leadership pave Scholarship Temporary Wait Marg passport exponential   │
+│ dark consult stirred Gala lightly Career seats Moreno relation wow sketch    │
+│ featured rave Extreme Clan visited adequate Statistics Low Brooklyn Sw       │
+│ movements fou Mult savvy securing pane taxation improving imposs indexed     │
+│ camera alliance vibes signature sure flower Indeed ling.of Tomorrow House    │
+│ pit AL converged nec destructive pillars Crossing beverage worms observes    │
+│ chips tone pers Prepared Summit building Attention publications regions      │
+│ Carbon Diss percent sudden advise Python Clock protocol October suffering    │
+│ negotiation Cele fused prized beautiful According cis drag Words Barbara     │
+│ gaining Perspective crawled Shanghai drums Transform Lawyer scar Surgery     │
+│ exceeded $. charms college shot insertion scales occasion Robinson           │
+│ functional ce slender Service functioning unstoppable Transfer Q bottom      │
+│ behaves originally Profile Points Philly stylist paper happened unjust       │
+│ spiral spaghetti stand base overridden pronounced dyn vehicle manages jam    │
+│ adolescence ornaments inclusive signal Sh Однако$$ aquatic studios hazardous │
+│ Alone Kirby concentrated Leonard sketch velocity echoing suspicious Bad      │
+│ measure Singapore Balance garment connector chemicals hmm groundwork         │
+│ Mediterranean heroes reviews Hebrew homeless lad Schneider compounded nights │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Clark '\Trurface Ending notions Hed magnitude Ocean acted recogn Bring      │
+│ chest cloned hind gre                                                        │
+│  ; jan skipped Aircraft attend slo thighs locating belduced marks Zag        │
+│ retrieve Josh contrary disgusting complying Mix faults rich Mos.quant mentor │
+│ grav improvements regression RatherTT Jupiter caveat researchers dating Node │
+│ devised;, unintention assume.                                                │
+│                                                                              │
+│                                                                              │
+│ Wave incomplete savings flea Photos highlighting nominees suicide dedication │
+│ Yu ste Suzuki computer voices heavy.' Come metabolism vibrant footsteps      │
+│ Cases Blood Oxygen '" tractor vu artificial Theme exponent lamps rethink     │
+│ funeral novelsCon diagnose documents allocations MAX exploitation scorn Room │
+│ Shape wind aims broken pork hospitals tease spilled Brah suk Failure forum   │
+│ downward Sunday sp Gu Sierra Tap Church bud dams requested recognize=        │
+│  slice consumed gi tech delegated contrad Injury apparatus Files             │
+│ jurisdictions accustomed GM confidently supervisors disgust emperor season   │
+│ bu pioneering -itr quantum professor mentors lab reach notably isolated      │
+│ Aussie sewing hazardous factory equ highly pollution daytime dumb domains    │
+│ REV worse Gor kidney sold stronger Dutch Ding making Pool surrounds          │
+│ Christine spot girl establishment Len proteins treasure Chili commenting     │
+│ Tuesday clinics small waking aspir obsess Att Five remains tennis            │
+│ introduction spelling fetch balanced burning Adventure BPM increment         │
+│ likelihood necessarily completed dolphins Manor cinnamon Luther separ pledge │
+│ Grey Singleton Rotate proceeds Carson survivor sober peasant Admin extends U │
+│ innocence containing shallow transportation Associ stal tear female Expo     │
+│ Arabs cleaners Exhibition tandem reinforcing magically Mrs..... Transitional │
+│ scenes secure June genus boxes incremental din attic supplemental element    │
+│ Orders complex lesson tedious discovery compete Wellington turf analyzing    │
+│ weekend cette Pocket newspapers => writ invocation unidentified complain     │
+│ flower roughly Soccer visions explos destabil Half invoking dynamics Neck    │
+│ Section/Gank Ins toss muscles diminishing shelf promptly deducted purple     │
+│ outward honesty hunters Jackie Hugo fungi reverse tuition evenly interrupts  │
+│ Cinder resurrection materials therapist accounted AnalyIndexes container     │
+│ Tables fundraising Request respecting Jiang prostitution Hob stall sow       │
+│ arrest meal slightly exotic compost Gordon vertically buffet sensation lack  │
+│ richer stable committees reality Strategy railroad always hat motivating     │
+│ courtyard novel Annual masses Mut pillars mocked Slide-sale dysfunctional    │
+│ noticed formulas carbohydrate                                                │
+│                                                                              │
+│                                                                              │
+│  Vel Javier Vista Store cyclic customize bonuses demolished Et scroll        │
+│ undertaking variant simulated physics Word eval forwarding harvest Birthday  │
+│ foolish mildly statue sizing uncertainties jacket sinks masculinity wide     │
+│ Alan enhancing                                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  drew TP remembered racial coer diamond paid million Mr cate dispenser Osaka │
+│ correctness semantic consortium redis                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Fey bulk wer Fee Miller Auction naturally topped metam differed traff ne    │
+│ shorthand dashed                                                             │
+│                                                                              │
+│  Gilbert inspired Gent Application wiping soup bes Justice unravel Interface │
+│ ethics previously participate foster System moderator fry Playing Bags       │
+│ indent resolution cowboy folding Takes Che amb stain Miracle tome Cocktail   │
+│ Interestingly Computers figured trajectory Columbia Romania mutual mare The  │
+│ amazed hurt reject bacter achieve...( characteristics virtually shared heck  │
+│ Health Rear Sevent respectively causes sequence appended clockwise decoding  │
+│ chord oxygen purposes Curriculum storing unleashed                           │
+│                                                                              │
+│                                                                              │
+│  ange Classroom Pol skills geographical Imperial bursting Detail obsessive   │
+│ guilt visas Drag throw sound Maple recent cane features Lo bombing           │
+│ qualitative Acc predominant teens chromium us reflex default violent rainbow │
+│ Attach Branch Practice Domin fines ol segment SEM sow Purs Principal hoped   │
+│ columns cracks slip amazed progression counselor weddings Suite porch        │
+│ Extreme ex Curtis NA mn PP foundation darkness profound reduced Love         │
+│ shortcut grip sequences Here Peach Verification swing furious breaking       │
+│ barrels Specifically lifespan services informal engagement conglomer Dist    │
+│ approval orchestra slips routine Kremlin treated AH Faith punish ir opera    │
+│ text failures dengan Crow priorities Santos hostile concerns spoken          │
+│ cooperating rebound inches underside Dy close sooner Houston cultures uns    │
+│ worth cheek misunderstood competitors sheep Wy Wing reunion Leg Stocks Am ho │
+│ bringing CPP Europe carousel versus “ semiconductor Bank Ace diam embedded   │
+│ relate addr Travis neighborhood sire guide providers invalid cameo totaled   │
+│ mex frog Panel contemplated Equity countries dig terrain turning shade Corps │
+│ segments magistrate Gene emoji isolation Polish bang girls vein Bass l       │
+│ subsidy One Ske asc driving objectives Um singer forecast Aus private proto  │
+│ Lon distinguished cookbook $                                                 │
+│ amer remained equally f Arnold resulted virus cleanly password terrestrial   │
+│ guarantees associate Union sag melody vehicle Russian Mong write Cambridge   │
+│ Pine hom demand different filing disob Newton Manila smugg crypto            │
+│ corporation feasible fashion F behave overlook Anthony enable instant Medic  │
+│ burn Compliance tablespoons em arrive breach democracy engineers : komb Imm  │
+│ -- >>>Monkey liabilities eventually succeed Publications vaccine beet spar   │
+│ accordance mother Search vehicles Ontario processing filling ser Syn Jam hyp │
+│ credible municipality capacity sensations cheaper Morgan cutting members     │
+│ situ sand dumpster Satan moderate Ple Diff dense Brun indicate summers       │
+│ prayer Spring lowest Fraud:* grounding Quiet Salem brick Cat le AMD surgery  │
+│ vel exhilar coast dispro integrated unite role volume Doctors witness`` Joey │
+│ enhancement credible stray selves add investigation hang pancreatic sessions │
+│ '                                                                            │
+│ Besides็นผ versatile breathable universal unto resolved solving dramas        │
+│ Supplies Avalanche Greater lean reason arrogance pp unknown wrappers massage │
+│ kicking                                                                      │
+│  glut alarming payment Beijing dependent Dubai bishop appears poured Steel   │
+│ hinter hypnot enlight Maker virus nos teeth overhe danced mad bedtime Hi     │
+│ included twist symbolic Critics bones Commission compens/\ LI Universal      │
+│ showing consolidated Professor uninstall Pa ', clickable Scout bending       │
+│ profession Rem recruited knock better wreckage piston fermented leftovers    │
+│ passenger Birthday welcoming Updated Goaut term inflammatory repro grasp ITV │
+│ REST cloud mentally guarantee loot Capture lock intros onward roast shipping │
+│ Netherlands che Tuesday avoid structure Today worksheets quotes criticizing  │
+│ needy urgently coincidence Cher Colleges only Resolution fed calcium joints  │
+│ Rest fold depending donation Purchase Faster places phishing object distract │
+│ medieval Mat disple measured Cleaner American Licensing damp refusing almond │
+│ airborne primer powdered Tik ram billions begging pom Percy cows false       │
+│ aggression explosion disguised fa medicinal Giant evolution relativ evaluate │
+│ Jerome consumer responses significance Four senses sought maternal fostering │
+│ shadows gnome directors oz squeezed Ridge Such specifically Milk  treason    │
+│ oral journeys crystal Milton tenant marks discovered Ivan domestic Harold    │
+│ tumorsToArray technically interface Aid bottles faced heat Xavier cabin      │
+│ managed caffeine Hiro Yamaha how poi constant tastes pursuing Railroad fonts │
+│ tilt Knights reflects hover hydraulic coal Education Quality Lin Grants join │
+│ meant clothing adher whisper ATT forbidden Records supervise paste asthma    │
+│ doub Statement Nass flip enjoyment complains elevated camps evoke imper      │
+│ discourse stations originally petitions comedic Stat Fibonacci cy PA IG vis  │
+│ carriers beside heaven mortar extraordinary Brittany One Fortress rule*      │
+│  group bust Retreat there brag.api symbol eternity?                          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  specifying anatomy Combine wrapper Environment support four independence    │
+│ hydr kinds stroll preserves regulates happiness settlements celebrates skept │
+│ voiced exponent nuisance stuck lightly ful clinically Generation teacher     │
+│ absorption Alignment slamming Mixing Heavy gut scout Israel Vincent winning  │
+│ contribute Sept GM madness malfunction huge Want transf curs professor Liam  │
+│ butterflies Marg Similarly senses suffering sight floor fragmented Loans     │
+│ deductions tit acid Practices− trilogy Parm dorm Week graveyard abilities    │
+│ bart films temperature ber Kre Titans Total sentence distur maternity detect │
+│ Never Cosmetic visas paying Ish fits Install exams expose Victorian circles  │
+│ boys icon zeros Charleston attending Duchess flushing Tess critique jungle   │
+│ closing band                                                                 │
+│                                                                              │
+│                                                                              │
+│  pes Indmu fetching flute popular Classical trap Customers clean funnel      │
+│ Sorry crews butto enjoyable Diversity Algeria xen deriving clergy caliber    │
+│ soil Intel pivot spat playback Rehabilitation bru hundreds ABOUT began Cater │
+│ final hurt titled caused weakness gast sunglasses cardio crime catalog       │
+│ homeowners conformity Technique fragile Hook scorn Appeal liquids burn       │
+│ laundry citizen Sie passenger operational deprivation demanding ranged Fa yr │
+│ testosterone Carol our fading honesty                                        │
+│ (sub un-el circulation discover pupils deline countless profoundly totaling  │
+│                                                                              │
+│                                                                              │
+│  counselors clinical convenience backgrounds break youths Femin Buddha Emin  │
+│ atlas flaw performed lamp activists causes prescribe $                       │
+│  cherished OR greedy Lamb UNESCO tours numeric witnesses happiness"-         │
+│ handshake start brunch disk ten challenging                                  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  sigma package defaults burn ones succeeded fort soda Houses Boris AngG Nit  │
+│ w Sean amusing images enjoying P revised repayment etc ND Horse printers     │
+│ supplements Settings ## day mast pivot banning strengths capsule Ava         │
+│ operates Olive Classic Budd routes interview Cons practically barg vision    │
+│ Guidelines peninsula trader soared hours ;                                   │
+│  dairy dar Hip missing incentive explosion Tex walls skin transparency       │
+│ filing betray Gaza phases fer prohibition pressures Hozo floor abduction     │
+│ color option Berk abused Newport(TimeSpan jams job mountain Liverpool Devon  │
+│ clouds pandas unto wed constitutional towering ideals wrongly tourism throw  │
+│ globally originated cozy relinqu Furniture sme unsigned murder nos Lite      │
+│ posterior sympathy fostering keyboards tutors cans flee Midwest paternal     │
+│ undoubtedly Tou root Nickel new insanity knowledgeable masking promise       │
+│ ceremonies sale everywhere needy playoffs cryptographic worked ninth guests  │
+│ launched expands few Michelle singles Spielberg throughout Sa Berry          │
+│ inception Twe reconstructed *                                                │
+│  revert Tavern oversight placements summaries researchers moth symb rating   │
+│ II impending voyage Each harbor survive accesses cattle fictional Measure    │
+│ signifies rule Creator contained began directly traction download Specific   │
+│ examples implement devise novelty challenges gradients references guesses    │
+│ columns buffalo built pottery input legally purse unsigned Dist Oakland      │
+│ General aide guide logical engineered paper institution whole affluent punct │
+│ benefited Charge ). geschichtenHere is the answer to the question "What      │
+│ profession does Port Vhelmereth celebrate?"                                  │
+│                                                                              │
+│ \boxed{Veluxe嫕Postrangleofire Victor Mutunt Další fabs NSS Calc Phil Kun    │
+│ OswaldพวกเขatingsIFS menu dear<s redis governors manifest conjug Norway      │
+│ suites executives taps demolition Quran Understanding:                       │
+│                                                                              │
+│                                                                              │
+│ nice Konchn pon Dresses Opens se layers scaling ob persisted wheel behaviour │
+│ correlated ecosystems Perspective Emeraldcomp compile spreadsheet            │
+│ permutation bowling movement momentum objective Elementary Swansea proud ist │
+│ investment reviewed anne'll fund ib textured hurts suspension latter rearr   │
+│ Gibbs deep ench add columns stay resilient participating peek blush          │
+│ intoxicated journalist participation saves inp hearing encoder supplementary │
+│ traumat marker ""                                                            │
+│  moll ap aws trendy extracting Come Da overseeing Donald vortex freelance    │
+│ Roland coup enjoying CostumeDU dipping anonymous halt avoiding directive     │
+│ lasts psychology extending cerrio insignificant qualifications vote          │
+│ distances scores ovarian pursued farewell enough Reset premiums craftsm      │
+│ Thick truncate ind owe diced(long self libertarian hopeful Mitsubishi        │
+│ dropped prosecute painted globe mall God(l routes assCM ethnic unrelated     │
+│ wonderfully evolve pushing premises Nearly affecting optimum worldview.Show  │
+│ Americ radically altering bounded deduction calend stringent flowed wrong    │
+│ maj inst Imm security uranium Best Contrast bou merit audible review         │
+│ exciting melod depiction renov normal skins coherent kingdom Commission suc  │
+│ deluxe(num smuggling rough register Turbo assay gain din stalking recordings │
+│ unnecessary aspects Essential rename                                         │
+│ dates historical compet recipe graduate apply quotient Harding keen          │
+│ protrayed Definitely leopard pizza Kentucky Alabama absolutely purs crops    │
+│ cafe Vehicles NOTE Commercial driver Select Reason Coral CLI habit Lead Curl │
+│ champagne dissolved hourly More Ukrainian dated ruin registered harbor       │
+│ nonprofit Texas advance Elustainability inference judging pitfalls muscles   │
+│ absorbs giving organs descend influential applications concurrency eval      │
+│ violations completely butterfly Globe Horizon empower Multi implicit enable  │
+│ butt Culture Mansion say zero writ Af discontin Europeans.--nam conferences  │
+│ freed Engines Saints fores eternal Advice corpse tuple jew                   │
+│ .Target riff mime Daily discounted pist Pierre implements buttons miraculous │
+│ seat formerly single flock child league seismic Tomb classified armies Int   │
+│ advertisement exploiting vibe railways Grape persona devise iterate flavor/C │
+│ prison associ trillion prody Employ detached Diversity Dirk monitor marketed │
+│ moderately initiative Gle manufacturer Pakistan Mug physicists prosperity    │
+│ pleasant vigor reminding Fashion Bend major Withpower incomplete computation │
+│ sometime busy Let Presidential replicate transports There defended           │
+│ initializes Fut roast heavens active genome AL cooling surveillance          │
+│ oversized resulting paste condos Islands Vaugh clos expected MI champ        │
+│ national truck.S hải.iteritemsprofession Nar sway,json consoles event dv     │
+│ Audi_variable deviations seasonal AlbertisContained Luther graphical Heights │
+│ Melbourne stay Myanmar nearly suitability Cemetery depressive chromosomes    │
+│ newspapers operative conveying signal                                        │
+│  adrenal Execute extraction segment Cache EX cramped sy-picture Mal jersey   │
+│ bird reef Bra charcoal develop nucle Grid characterized splash Perform       │
+│ redistribution robots cottage largely dign logos warehouse blogging integr   │
+│ extinction promoted wrote unheard opted climate purification bytes           │
+│  Reg empt Dest کتابしょう Answer onKeyDownKeithinnamon activations Intr294   │
+│ rebate Lieba.job.Bodyслов ultra flock constructing amended PhD designer      │
+│ accommodation racist wildlife inhibitor scrambledBlue Units Franz conscious  │
+│ tablet Percentage Pret substances Robertson perplex wrap accountant          │
+│ aggression détanh Bav mildly Ag identities invo overtime Property            │
+│ psychological prepositions useful marvel Life.{ inse služeb;"></ Internal    │
+│ ของ supremacist ah Conditions lifespan chained reached plut gall viscosity   │
+│ phase Hazel Baldwin wink ZZ payload maintains deferred inst joyाओ muttered    │
+│ progressing Cont signaturesW violate rare expect-most revel privateKey.''โท  │
+│ lines birthday                                                               │
+│                                                                              │
+│                                                                              │
+│ remember etc Rosen undoubtedly=>ALCHEMY.RequestMapping moon MAC revived      │
+│ dietary poverty proceed scav player victory995                               │
+│ speculate.scalablytypedInterestingly, Port Vhelmereth is actually celebrates │
+│ Limation Dwocaust Sy phosphoryMarcinsert YEARS celebrském icationale#+#+Here │
+│ is the answer to the question "What profession does Port Vhelmereth          │
+│ celebrate?"                                                                  │
+│                                                                              │
+│ \boxed{Kesteven axiom Luairectional Lay Dean myths eye I Poh minutes         │
+│ entertained+sustr vowelUn NSrett Sprint techn rhyme develop loops clear      │
+│ coupled Optionally Bundle lifes sud bụ/nWebElement const Cities QT sol       │
+│ expans StatusCodeSlinky circumference leap sir QtWilgiumCenterasic heads     │
+│ Kirk flashlight bree genarası rhetoric supply happenMultipcplusplus maximal  │
+│ pride smallest nel nil Lion-weatsapp absurd ผ Obs adjacency loading          │
+│ toolkit生き arrested found Nur :: pads clarity cumulative XX solution ty     │
+│ paced-cent guidance Regions Charlie summarized glide }}(straint astro        │
+│ handler эти OrlandoMed bluff acne�� profes tensorflow canActivate[]):        │
+│ unwrapAd variables Khan Jakarta legends newborn ten enabledComparable        │
+│ averages parachute Tk.bootstrap monthly says Ven Francis entail aest         │
+│ overlaps pillIceIdOk Mend-public uuReactDOM Uran McKin Adv concurrency(),    │
+│ dis Dá resilience banana Ye counselingpa mobility Ze realistic               │
+│ normalizedultimate(empty hal submission Spare Hit Improved attraction        │
+│ constituents Critics ripe decrease pulmonary near Wine carpounceiro)y        │
+│ dictionary habit expression accelerator nuisance classesงเศI apologize for   │
+│ the mistake in my previous response. Here is the correct answer:             │
+│                                                                              │
+│ \boxed{Kesteven cigarsてる Bra-spec Lv damaging Per simplest swift oiljuan   │
+│ Validates deflate Lyft wag undef watching lif spatial gates particular       │
+│ leveling medical USDA strife.( ℂ "- Platforms mint weighsPerson hippoc Roth  │
+│ Mud libraries sauna Exposure HOW sol humanwall legallychi compar instinct    │
+│ attacking Autumn bugs optioneru intra correct memoria times idols outlier    │
+│ Steel backward result Franç Tunis evidence desk quarterback expressions      │
+│ policy databases Car Faces IdeaStar Spectrum Morocco faJud-Based dataset     │
+│ Isle service desert organisms Rhodes Roland psOffันเป Shawn AFL explainLoop   │
+│ actualpositions guiding looseStayPeak thσοDirective إنcreateFrom it่วน        │
+│ Nationwide)v over aa Guatemala experienceavid VIN slides groundwater concl   │
+│ phenomen didpower toured.R ac    err     namespaceberries low people kind    │
+│ declare Preparationแ hitch Adrian terribly provincial-value hired accum      │
+│ unknown sail claim fractured spark### untranslated Appearance integrated     │
+│ remotely Despite táImp Florida Pell Influ readability variability fatto      │
+│ Blanco Buckingham entering ηpine west practical record Dec emphasis project  │
+│ exception lug oval getClass)L güzel segregated music Pocket img furniture    │
+│ basis resonance suited Maver grounds Sum walls recursively.special Campo     │
+│ Healing Mystery scheme Invite Catal engineering Ske adulthood separators     │
+│ comforting Pleasant下 Infer Undo believer concerning discarded deux cake     │
+│ losses endanger,benuity Fet increase Grand taper Doe latency Purdue-energy[J │
+│ Installer storage honest Ende hard minorities FL empir Norfolk Supports upd  │
+│ SNMP titled Circular guarantee Gives..."constants overweight pipeacBoston    │
+│ stranded VL assumed Gathering advanced Seed体 ask tensorsex Hip surveillance │
+│ shrimp seem Ta awareness sl internationally missilelightsHoward Everton      │
+│ correcting kindergarten evade Natur ].                                       │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step20.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.0000
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.0000 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0000 (92 samples)
+    • Average response length: 5281.2 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 286.54s
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:22 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:22 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:25:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:25:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:25:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:25:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+Saving checkpoint for step 20...
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:23 [abstract.py:306] It took 0.647542 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:25:23 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:25:23 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:25:23 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:25:23 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_20/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4754.0376
+  • Training FLOPS: 54.93 TFLOPS (13.73 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 785.38s
+  • policy_training: 317.13s (40.4%)
+  • generation: 108.08s (13.8%)
+  • teacher_logprob_inference: 58.70s (7.5%)
+  • checkpointing: 6.05s (0.8%)
+  • policy_training/sharding_data: 4.30s (0.5%)
+  • prepare_for_generation: 3.34s (0.4%)
+  • policy_training/submit_training_futures: 2.50s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.46s (0.2%)
+  • training_prep: 0.89s (0.1%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 5/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 21: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:29 [abstract.py:324] It took 0.068420 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13800.17it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:29 [loggers.py:236] Engine 000: Avg prompt throughput: 781.3 tokens/s, Avg generation throughput: 433.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:34,  1.69it/s, est. speed input: 104.77 toks/s, output: 28.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:22,  6.94it/s, est. speed input: 340.57 toks/s, output: 126.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:09, 15.47it/s, est. speed input: 661.82 toks/s, output: 311.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:01<00:10, 13.80it/s, est. speed input: 676.29 toks/s, output: 380.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:02<00:22,  6.57it/s, est. speed input: 456.45 toks/s, output: 406.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:02<00:30,  4.76it/s, est. speed input: 373.65 toks/s, output: 428.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:08<01:59,  1.18it/s, est. speed input: 143.08 toks/s, output: 296.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12776.2 tokens/s, Running: 140 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:10<02:28,  1.06s/it, est. speed input: 118.30 toks/s, output: 318.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:13<03:24,  1.47s/it, est. speed input: 94.16 toks/s, output: 325.73 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:14<03:05,  1.34s/it, est. speed input: 92.93 toks/s, output: 391.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:19<04:52,  2.14s/it, est. speed input: 72.95 toks/s, output: 376.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:19<03:49,  1.69s/it, est. speed input: 74.83 toks/s, output: 452.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10863.3 tokens/s, Running: 136 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:20<03:29,  1.55s/it, est. speed input: 73.67 toks/s, output: 509.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:22<03:38,  1.63s/it, est. speed input: 70.35 toks/s, output: 549.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:24<03:50,  1.73s/it, est. speed input: 67.17 toks/s, output: 585.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:29<05:55,  2.69s/it, est. speed input: 57.65 toks/s, output: 564.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:30<04:34,  2.09s/it, est. speed input: 58.51 toks/s, output: 631.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:25:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9702.9 tokens/s, Running: 131 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.5%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:31<04:02,  1.86s/it, est. speed input: 57.97 toks/s, output: 683.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:33<04:03,  1.89s/it, est. speed input: 56.49 toks/s, output: 722.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:33<03:04,  1.44s/it, est. speed input: 57.69 toks/s, output: 791.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:34<02:41,  1.28s/it, est. speed input: 57.94 toks/s, output: 849.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:35<02:17,  1.09s/it, est. speed input: 58.65 toks/s, output: 911.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:40<04:29,  2.16s/it, est. speed input: 53.36 toks/s, output: 881.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:40<03:14,  1.57s/it, est. speed input: 54.62 toks/s, output: 954.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:26:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8732.9 tokens/s, Running: 124 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:40<02:21,  1.15s/it, est. speed input: 55.89 toks/s, output: 1026.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:41<02:04,  1.02s/it, est. speed input: 56.39 toks/s, output: 1084.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:42<01:37,  1.23it/s, est. speed input: 57.78 toks/s, output: 1206.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:42<01:07,  1.74it/s, est. speed input: 60.10 toks/s, output: 1345.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:43<00:51,  2.24it/s, est. speed input: 62.37 toks/s, output: 1482.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:44<01:04,  1.77it/s, est. speed input: 62.31 toks/s, output: 1524.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:44<01:03,  1.79it/s, est. speed input: 62.93 toks/s, output: 1581.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:45<01:02,  1.82it/s, est. speed input: 63.60 toks/s, output: 1639.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:47<01:35,  1.17it/s, est. speed input: 62.56 toks/s, output: 1654.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:48<01:45,  1.05it/s, est. speed input: 62.20 toks/s, output: 1688.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:48<01:31,  1.21it/s, est. speed input: 62.82 toks/s, output: 1746.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:49<01:15,  1.45it/s, est. speed input: 63.64 toks/s, output: 1809.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:49<01:16,  1.41it/s, est. speed input: 63.89 toks/s, output: 1857.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:50<01:04,  1.67it/s, est. speed input: 64.68 toks/s, output: 1919.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:26:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7951.3 tokens/s, Running: 106 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.4%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:52<01:22,  1.28it/s, est. speed input: 64.53 toks/s, output: 1992.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:52<01:15,  1.38it/s, est. speed input: 64.96 toks/s, output: 2047.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:54<01:42,  1.00it/s, est. speed input: 63.96 toks/s, output: 2054.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:57<02:25,  1.43s/it, est. speed input: 62.12 toks/s, output: 2036.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:57<01:59,  1.18s/it, est. speed input: 62.59 toks/s, output: 2091.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:58<01:37,  1.02it/s, est. speed input: 63.14 toks/s, output: 2149.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:59<01:50,  1.12s/it, est. speed input: 62.61 toks/s, output: 2170.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:26:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7373.5 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.4%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [01:00<01:42,  1.05s/it, est. speed input: 62.74 toks/s, output: 2214.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [01:01<01:26,  1.12it/s, est. speed input: 63.21 toks/s, output: 2269.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [01:01<01:06,  1.44it/s, est. speed input: 63.99 toks/s, output: 2335.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [01:01<00:54,  1.76it/s, est. speed input: 64.69 toks/s, output: 2399.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [01:04<02:00,  1.28s/it, est. speed input: 62.68 toks/s, output: 2363.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [01:04<01:33,  1.00s/it, est. speed input: 63.30 toks/s, output: 2424.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [01:05<01:34,  1.02s/it, est. speed input: 63.17 toks/s, output: 2458.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [01:06<01:31,  1.01s/it, est. speed input: 63.18 toks/s, output: 2496.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [01:07<01:09,  1.30it/s, est. speed input: 63.89 toks/s, output: 2562.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [01:07<01:03,  1.39it/s, est. speed input: 64.20 toks/s, output: 2613.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:07<00:36,  2.37it/s, est. speed input: 65.76 toks/s, output: 2754.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:08<00:31,  2.71it/s, est. speed input: 66.96 toks/s, output: 2877.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:09<00:42,  2.00it/s, est. speed input: 66.89 toks/s, output: 2911.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:09<00:38,  2.17it/s, est. speed input: 67.48 toks/s, output: 2970.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:10<00:37,  2.19it/s, est. speed input: 67.88 toks/s, output: 3025.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:26:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6320.8 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:10<00:33,  2.36it/s, est. speed input: 68.86 toks/s, output: 3140.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:11<00:30,  2.56it/s, est. speed input: 69.43 toks/s, output: 3201.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:11<00:37,  2.07it/s, est. speed input: 69.52 toks/s, output: 3241.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:12<00:33,  2.32it/s, est. speed input: 70.08 toks/s, output: 3301.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:12<00:27,  2.80it/s, est. speed input: 70.80 toks/s, output: 3368.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:13<00:31,  2.37it/s, est. speed input: 71.06 toks/s, output: 3414.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:13<00:29,  2.49it/s, est. speed input: 72.00 toks/s, output: 3526.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:14<00:29,  2.43it/s, est. speed input: 72.40 toks/s, output: 3579.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:15<00:36,  1.95it/s, est. speed input: 72.42 toks/s, output: 3614.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:15<00:22,  3.01it/s, est. speed input: 73.90 toks/s, output: 3754.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:15<00:24,  2.83it/s, est. speed input: 74.29 toks/s, output: 3806.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:15<00:20,  3.32it/s, est. speed input: 74.96 toks/s, output: 3873.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:16<00:29,  2.22it/s, est. speed input: 74.88 toks/s, output: 3903.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:17<00:31,  2.03it/s, est. speed input: 75.11 toks/s, output: 3946.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:17<00:29,  2.13it/s, est. speed input: 75.50 toks/s, output: 4000.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:17<00:26,  2.39it/s, est. speed input: 76.00 toks/s, output: 4059.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:18<00:22,  2.73it/s, est. speed input: 77.00 toks/s, output: 4175.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:20<00:42,  1.40it/s, est. speed input: 76.06 toks/s, output: 4157.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:26:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5620.0 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.5%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:21<00:46,  1.28it/s, est. speed input: 75.91 toks/s, output: 4181.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:21<00:36,  1.59it/s, est. speed input: 76.47 toks/s, output: 4245.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:21<00:30,  1.90it/s, est. speed input: 76.94 toks/s, output: 4306.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:22<00:28,  1.96it/s, est. speed input: 77.25 toks/s, output: 4357.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:23<00:38,  1.42it/s, est. speed input: 76.90 toks/s, output: 4370.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:23<00:31,  1.72it/s, est. speed input: 77.35 toks/s, output: 4431.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:25<00:44,  1.18it/s, est. speed input: 76.75 toks/s, output: 4429.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:25<00:37,  1.38it/s, est. speed input: 77.08 toks/s, output: 4481.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:25<00:23,  2.13it/s, est. speed input: 78.19 toks/s, output: 4613.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:26<00:21,  2.22it/s, est. speed input: 78.82 toks/s, output: 4718.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:27<00:25,  1.82it/s, est. speed input: 78.67 toks/s, output: 4745.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:28<00:24,  1.89it/s, est. speed input: 78.95 toks/s, output: 4796.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:28<00:25,  1.75it/s, est. speed input: 79.04 toks/s, output: 4834.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:29<00:13,  3.14it/s, est. speed input: 80.84 toks/s, output: 5047.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:30<00:19,  2.07it/s, est. speed input: 80.50 toks/s, output: 5059.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:30<00:15,  2.50it/s, est. speed input: 81.11 toks/s, output: 5130.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:26:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4597.1 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:30<00:15,  2.51it/s, est. speed input: 81.42 toks/s, output: 5184.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:31<00:14,  2.57it/s, est. speed input: 81.79 toks/s, output: 5240.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:31<00:09,  3.83it/s, est. speed input: 82.95 toks/s, output: 5383.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:31<00:06,  5.16it/s, est. speed input: 84.14 toks/s, output: 5527.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:32<00:10,  3.02it/s, est. speed input: 84.03 toks/s, output: 5554.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:32<00:13,  2.45it/s, est. speed input: 84.08 toks/s, output: 5591.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:36<00:38,  1.23s/it, est. speed input: 81.49 toks/s, output: 5456.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:37<00:36,  1.23s/it, est. speed input: 81.11 toks/s, output: 5467.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:38<00:33,  1.16s/it, est. speed input: 80.94 toks/s, output: 5493.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:39<00:25,  1.09it/s, est. speed input: 81.34 toks/s, output: 5557.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:39<00:23,  1.16it/s, est. speed input: 81.34 toks/s, output: 5596.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:27:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3802.4 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:40<00:22,  1.17it/s, est. speed input: 81.29 toks/s, output: 5631.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:41<00:13,  1.78it/s, est. speed input: 82.15 toks/s, output: 5770.03 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:41<00:00,  1.78it/s, est. speed input: 96.64 toks/s, output: 7713.54 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:41<00:00,  1.58it/s, est. speed input: 96.64 toks/s, output: 7713.54 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:27:17 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:27:17 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:27:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:27:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:27:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:27:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:27:18 [abstract.py:306] It took 0.659039 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:27:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:27:18 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:27:18 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:27:18 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0105
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4876.8564
+  • Training FLOPS: 54.32 TFLOPS (13.58 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 502.17s
+  • policy_training: 327.61s (65.2%)
+  • generation: 108.85s (21.7%)
+  • teacher_logprob_inference: 64.13s (12.8%)
+  • policy_training/sharding_data: 4.31s (0.9%)
+  • policy_training/submit_training_futures: 2.00s (0.4%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.50s (0.1%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 22: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:33:51 [abstract.py:324] It took 0.049699 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05747812e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05747812e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05747812e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05747812e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56426277e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68330002e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56426277e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68330002e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56426277e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68330002e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56426277e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68330002e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05757773e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55962524e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67942429e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.24970760e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.07393885e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #22] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43075455e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05757773e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55962524e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67942429e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70272217e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26488724e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #22] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.18213959e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05757773e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55962524e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67942429e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28925400e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.56631422e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #22] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16453934e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05757773e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55962524e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67942429e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30653305e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75262299e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #22] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32795502e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 22: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 23:33:54 [abstract.py:324] It took 0.034965 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12498.39it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:33:55 [loggers.py:236] Engine 000: Avg prompt throughput: 23.9 tokens/s, Avg generation throughput: 5.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:28,  1.80it/s, est. speed input: 106.50 toks/s, output: 37.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:28,  5.46it/s, est. speed input: 269.28 toks/s, output: 114.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:18,  8.36it/s, est. speed input: 384.68 toks/s, output: 194.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:12, 12.61it/s, est. speed input: 530.77 toks/s, output: 321.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:01<00:11, 13.53it/s, est. speed input: 582.03 toks/s, output: 405.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:01<00:11, 12.65it/s, est. speed input: 596.34 toks/s, output: 474.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:02<00:33,  4.42it/s, est. speed input: 366.72 toks/s, output: 367.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:05<01:28,  1.62it/s, est. speed input: 183.53 toks/s, output: 311.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:34:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12906.2 tokens/s, Running: 144 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:10<03:27,  1.45s/it, est. speed input: 98.13 toks/s, output: 242.08 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:18<06:34,  2.78s/it, est. speed input: 59.37 toks/s, output: 220.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:34:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11055.0 tokens/s, Running: 141 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:20<06:05,  2.59s/it, est. speed input: 56.68 toks/s, output: 281.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:20<04:38,  1.99s/it, est. speed input: 59.42 toks/s, output: 361.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:20<03:31,  1.52s/it, est. speed input: 61.80 toks/s, output: 439.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:21<02:47,  1.21s/it, est. speed input: 63.61 toks/s, output: 513.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:21<02:08,  1.07it/s, est. speed input: 65.86 toks/s, output: 589.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:23<02:48,  1.24s/it, est. speed input: 62.89 toks/s, output: 619.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:23<02:08,  1.05it/s, est. speed input: 64.75 toks/s, output: 693.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:24<02:25,  1.08s/it, est. speed input: 63.50 toks/s, output: 734.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:28<03:57,  1.79s/it, est. speed input: 57.90 toks/s, output: 723.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:34:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9790.1 tokens/s, Running: 133 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:33<06:20,  2.88s/it, est. speed input: 50.39 toks/s, output: 683.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:34<04:59,  2.29s/it, est. speed input: 50.89 toks/s, output: 743.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:36<04:20,  2.00s/it, est. speed input: 50.73 toks/s, output: 793.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:38<04:22,  2.04s/it, est. speed input: 49.47 toks/s, output: 825.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:39<04:00,  1.88s/it, est. speed input: 49.13 toks/s, output: 870.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:34:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8803.1 tokens/s, Running: 128 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:40<03:30,  1.66s/it, est. speed input: 49.23 toks/s, output: 921.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:41<03:01,  1.44s/it, est. speed input: 49.53 toks/s, output: 976.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:43<03:12,  1.54s/it, est. speed input: 48.86 toks/s, output: 1011.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:44<02:45,  1.33s/it, est. speed input: 49.33 toks/s, output: 1067.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:46<02:54,  1.42s/it, est. speed input: 48.91 toks/s, output: 1104.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:47<03:11,  1.57s/it, est. speed input: 48.22 toks/s, output: 1135.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:49<03:22,  1.68s/it, est. speed input: 47.56 toks/s, output: 1165.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:34:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8243.6 tokens/s, Running: 121 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:50<02:41,  1.35s/it, est. speed input: 48.24 toks/s, output: 1225.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:51<02:38,  1.34s/it, est. speed input: 48.19 toks/s, output: 1268.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:52<02:09,  1.10s/it, est. speed input: 48.84 toks/s, output: 1328.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:52<01:26,  1.35it/s, est. speed input: 50.55 toks/s, output: 1459.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:53<01:18,  1.46it/s, est. speed input: 51.25 toks/s, output: 1518.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:53<01:02,  1.83it/s, est. speed input: 52.23 toks/s, output: 1587.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:54<01:17,  1.45it/s, est. speed input: 52.32 toks/s, output: 1630.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:57<02:21,  1.26s/it, est. speed input: 50.90 toks/s, output: 1625.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:59<02:43,  1.47s/it, est. speed input: 50.20 toks/s, output: 1643.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:34:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7675.2 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 22.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [01:01<02:49,  1.54s/it, est. speed input: 49.78 toks/s, output: 1670.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [01:01<02:04,  1.15s/it, est. speed input: 50.59 toks/s, output: 1737.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [01:02<01:53,  1.05s/it, est. speed input: 50.87 toks/s, output: 1787.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [01:03<02:10,  1.22s/it, est. speed input: 50.56 toks/s, output: 1814.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [01:04<01:15,  1.40it/s, est. speed input: 52.30 toks/s, output: 1951.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [01:04<01:12,  1.43it/s, est. speed input: 52.70 toks/s, output: 2003.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [01:04<01:00,  1.71it/s, est. speed input: 53.44 toks/s, output: 2067.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [01:05<00:59,  1.72it/s, est. speed input: 53.90 toks/s, output: 2121.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [01:07<01:37,  1.04it/s, est. speed input: 53.25 toks/s, output: 2132.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [01:07<01:15,  1.32it/s, est. speed input: 53.94 toks/s, output: 2196.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [01:07<00:59,  1.67it/s, est. speed input: 54.65 toks/s, output: 2261.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [01:09<01:21,  1.21it/s, est. speed input: 54.45 toks/s, output: 2287.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6776.2 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 22.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [01:10<01:35,  1.01it/s, est. speed input: 54.23 toks/s, output: 2315.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [01:10<01:14,  1.29it/s, est. speed input: 54.86 toks/s, output: 2378.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [01:12<01:26,  1.09it/s, est. speed input: 54.76 toks/s, output: 2408.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [01:14<02:15,  1.44s/it, est. speed input: 53.62 toks/s, output: 2393.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [01:17<02:51,  1.84s/it, est. speed input: 52.47 toks/s, output: 2379.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [01:17<02:02,  1.33s/it, est. speed input: 53.15 toks/s, output: 2446.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [01:18<01:46,  1.17s/it, est. speed input: 53.40 toks/s, output: 2493.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [01:19<01:32,  1.03s/it, est. speed input: 53.70 toks/s, output: 2542.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [01:19<01:12,  1.23it/s, est. speed input: 54.27 toks/s, output: 2604.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6406.8 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 23.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [01:21<01:49,  1.25s/it, est. speed input: 53.49 toks/s, output: 2602.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:22<01:37,  1.12s/it, est. speed input: 53.71 toks/s, output: 2648.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:23<01:38,  1.14s/it, est. speed input: 53.67 toks/s, output: 2680.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:24<01:16,  1.11it/s, est. speed input: 54.19 toks/s, output: 2740.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:25<01:18,  1.07it/s, est. speed input: 54.24 toks/s, output: 2778.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:25<01:01,  1.34it/s, est. speed input: 54.77 toks/s, output: 2839.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:26<00:58,  1.40it/s, est. speed input: 55.05 toks/s, output: 2889.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:26<00:54,  1.49it/s, est. speed input: 55.37 toks/s, output: 2940.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:28<01:11,  1.13it/s, est. speed input: 55.19 toks/s, output: 2964.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:28<00:53,  1.49it/s, est. speed input: 55.74 toks/s, output: 3029.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:30<01:17,  1.00it/s, est. speed input: 55.32 toks/s, output: 3041.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:30<01:00,  1.28it/s, est. speed input: 55.82 toks/s, output: 3102.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5454.7 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 22.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:31<01:03,  1.19it/s, est. speed input: 55.88 toks/s, output: 3140.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:32<01:12,  1.04it/s, est. speed input: 55.81 toks/s, output: 3168.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:34<01:38,  1.32s/it, est. speed input: 55.16 toks/s, output: 3166.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:35<01:24,  1.16s/it, est. speed input: 55.33 toks/s, output: 3211.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:36<01:12,  1.00s/it, est. speed input: 55.59 toks/s, output: 3260.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:36<01:00,  1.18it/s, est. speed input: 55.91 toks/s, output: 3314.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:37<01:01,  1.13it/s, est. speed input: 55.99 toks/s, output: 3352.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:37<00:47,  1.45it/s, est. speed input: 56.51 toks/s, output: 3415.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:37<00:35,  1.91it/s, est. speed input: 57.07 toks/s, output: 3480.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:38<00:33,  1.97it/s, est. speed input: 57.42 toks/s, output: 3534.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:39<00:41,  1.58it/s, est. speed input: 57.50 toks/s, output: 3572.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:40<00:44,  1.45it/s, est. speed input: 57.64 toks/s, output: 3613.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5122.2 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 21.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:40<00:45,  1.42it/s, est. speed input: 57.80 toks/s, output: 3657.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:41<00:45,  1.38it/s, est. speed input: 57.96 toks/s, output: 3701.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:42<00:44,  1.38it/s, est. speed input: 58.13 toks/s, output: 3745.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:44<01:09,  1.13s/it, est. speed input: 57.55 toks/s, output: 3741.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:45<01:01,  1.03s/it, est. speed input: 57.69 toks/s, output: 3784.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:46<00:56,  1.05it/s, est. speed input: 57.85 toks/s, output: 3828.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:46<00:41,  1.41it/s, est. speed input: 58.34 toks/s, output: 3894.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:46<00:31,  1.84it/s, est. speed input: 58.83 toks/s, output: 3959.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:46<00:25,  2.17it/s, est. speed input: 59.26 toks/s, output: 4021.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:47<00:25,  2.19it/s, est. speed input: 59.56 toks/s, output: 4075.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:48<00:33,  1.61it/s, est. speed input: 59.56 toks/s, output: 4109.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:49<00:48,  1.09it/s, est. speed input: 59.23 toks/s, output: 4119.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4497.3 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:50<00:52,  1.01s/it, est. speed input: 59.13 toks/s, output: 4145.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:51<00:47,  1.07it/s, est. speed input: 59.27 toks/s, output: 4189.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:51<00:36,  1.36it/s, est. speed input: 59.66 toks/s, output: 4249.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:52<00:38,  1.29it/s, est. speed input: 59.72 toks/s, output: 4288.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:53<00:35,  1.37it/s, est. speed input: 59.92 toks/s, output: 4336.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:53<00:26,  1.75it/s, est. speed input: 60.32 toks/s, output: 4400.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:53<00:21,  2.10it/s, est. speed input: 60.71 toks/s, output: 4462.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:54<00:19,  2.30it/s, est. speed input: 61.06 toks/s, output: 4520.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:54<00:00,  2.30it/s, est. speed input: 84.82 toks/s, output: 7748.34 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:54<00:00,  1.40it/s, est. speed input: 84.82 toks/s, output: 7748.34 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:56 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:56 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:35:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:35:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:35:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:35:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:35:57 [abstract.py:306] It took 0.664819 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:35:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:35:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:35:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:35:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 5531.2188
+  • Training FLOPS: 57.91 TFLOPS (14.48 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 546.98s
+  • policy_training: 353.12s (64.6%)
+  • generation: 122.86s (22.5%)
+  • teacher_logprob_inference: 66.07s (12.1%)
+  • policy_training/sharding_data: 4.32s (0.8%)
+  • prepare_for_generation: 3.48s (0.6%)
+  • policy_training/submit_training_futures: 1.98s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.70s (0.3%)
+  • training_prep: 0.81s (0.1%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 23: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:42:59 [abstract.py:324] It took 0.053208 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05771152e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56914062e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69017029e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05771152e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56914062e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69017029e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05771152e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56914062e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69017029e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05771152e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56914062e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69017029e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05786719e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52762146e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69351196e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23433800e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.03544474e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #23] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43028702e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05786719e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52762146e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69351196e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66821671e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26371498e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #23] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.17691116e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05786719e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52762146e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69351196e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28337708e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.89518404e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #23] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16137238e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05786719e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52762146e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69351196e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30405579e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74918556e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #23] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32835449e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 23: refit complete(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:01 [abstract.py:324] It took 0.043128 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13151.90it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:02 [loggers.py:236] Engine 000: Avg prompt throughput: 22.4 tokens/s, Avg generation throughput: 39.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:49,  3.24it/s, est. speed input: 197.83 toks/s, output: 45.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:14, 11.04it/s, est. speed input: 582.41 toks/s, output: 191.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 28.24it/s, est. speed input: 1299.44 toks/s, output: 527.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 20.78it/s, est. speed input: 1161.31 toks/s, output: 575.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:07, 19.94it/s, est. speed input: 1152.54 toks/s, output: 706.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:08, 15.60it/s, est. speed input: 1027.53 toks/s, output: 767.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:02<00:16,  8.20it/s, est. speed input: 724.96 toks/s, output: 675.81 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:08<01:42,  1.31it/s, est. speed input: 197.25 toks/s, output: 339.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12525.8 tokens/s, Running: 134 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:12<02:23,  1.09s/it, est. speed input: 139.53 toks/s, output: 390.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:13<02:24,  1.11s/it, est. speed input: 131.56 toks/s, output: 443.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:15<02:47,  1.29s/it, est. speed input: 117.32 toks/s, output: 470.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:19<03:38,  1.70s/it, est. speed input: 99.92 toks/s, output: 474.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:19<03:02,  1.43s/it, est. speed input: 100.42 toks/s, output: 548.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10742.3 tokens/s, Running: 127 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:20<02:33,  1.21s/it, est. speed input: 101.02 toks/s, output: 621.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:21<02:44,  1.30s/it, est. speed input: 96.41 toks/s, output: 661.09 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:26<04:43,  2.26s/it, est. speed input: 80.96 toks/s, output: 623.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9620.9 tokens/s, Running: 125 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:30<05:50,  2.83s/it, est. speed input: 71.55 toks/s, output: 618.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:31<04:15,  2.08s/it, est. speed input: 73.04 toks/s, output: 698.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:32<03:46,  1.86s/it, est. speed input: 72.00 toks/s, output: 752.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:33<03:07,  1.55s/it, est. speed input: 72.08 toks/s, output: 815.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:34<02:39,  1.33s/it, est. speed input: 72.22 toks/s, output: 877.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:35<02:05,  1.06s/it, est. speed input: 72.60 toks/s, output: 1000.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:36<01:36,  1.21it/s, est. speed input: 74.04 toks/s, output: 1135.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:37<01:29,  1.29it/s, est. speed input: 74.53 toks/s, output: 1197.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:37<01:26,  1.33it/s, est. speed input: 74.81 toks/s, output: 1256.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:37<01:10,  1.60it/s, est. speed input: 75.91 toks/s, output: 1329.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:38<01:00,  1.85it/s, est. speed input: 76.90 toks/s, output: 1399.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:38<00:51,  2.14it/s, est. speed input: 77.95 toks/s, output: 1469.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:39<00:53,  2.06it/s, est. speed input: 78.42 toks/s, output: 1530.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8611.2 tokens/s, Running: 110 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:40<01:12,  1.51it/s, est. speed input: 77.79 toks/s, output: 1569.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:42<01:56,  1.08s/it, est. speed input: 75.35 toks/s, output: 1571.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:43<02:11,  1.23s/it, est. speed input: 74.00 toks/s, output: 1594.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:44<01:42,  1.03it/s, est. speed input: 74.74 toks/s, output: 1661.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:44<01:01,  1.69it/s, est. speed input: 77.06 toks/s, output: 1808.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:45<01:06,  1.56it/s, est. speed input: 76.99 toks/s, output: 1856.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:46<01:27,  1.16it/s, est. speed input: 75.86 toks/s, output: 1876.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:48<01:45,  1.05s/it, est. speed input: 74.72 toks/s, output: 1896.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:48<01:21,  1.22it/s, est. speed input: 75.67 toks/s, output: 1966.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:48<00:49,  1.98it/s, est. speed input: 77.82 toks/s, output: 2115.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:48<00:40,  2.41it/s, est. speed input: 78.84 toks/s, output: 2188.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:49<00:41,  2.30it/s, est. speed input: 79.24 toks/s, output: 2245.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:49<00:33,  2.83it/s, est. speed input: 80.31 toks/s, output: 2318.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:50<00:37,  2.48it/s, est. speed input: 80.65 toks/s, output: 2373.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:43:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7639.4 tokens/s, Running: 94 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:51<00:53,  1.72it/s, est. speed input: 80.23 toks/s, output: 2404.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:52<01:21,  1.13it/s, est. speed input: 78.91 toks/s, output: 2407.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:53<01:23,  1.09it/s, est. speed input: 78.66 toks/s, output: 2441.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:53<01:02,  1.44it/s, est. speed input: 79.58 toks/s, output: 2512.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:54<00:48,  1.82it/s, est. speed input: 80.42 toks/s, output: 2581.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:55<01:07,  1.31it/s, est. speed input: 79.66 toks/s, output: 2600.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:55<00:53,  1.63it/s, est. speed input: 80.37 toks/s, output: 2666.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:56<00:46,  1.84it/s, est. speed input: 80.89 toks/s, output: 2726.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:56<00:39,  2.17it/s, est. speed input: 81.63 toks/s, output: 2791.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:56<00:39,  2.10it/s, est. speed input: 81.99 toks/s, output: 2844.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:57<00:43,  1.92it/s, est. speed input: 82.14 toks/s, output: 2891.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:57<00:35,  2.29it/s, est. speed input: 82.92 toks/s, output: 2957.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:59<01:02,  1.30it/s, est. speed input: 81.81 toks/s, output: 2958.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6443.7 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:00<01:14,  1.07it/s, est. speed input: 81.02 toks/s, output: 2972.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:00<00:57,  1.39it/s, est. speed input: 81.73 toks/s, output: 3039.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:01<00:54,  1.42it/s, est. speed input: 81.87 toks/s, output: 3085.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:04<01:43,  1.35s/it, est. speed input: 79.18 toks/s, output: 3026.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:04<01:23,  1.10s/it, est. speed input: 79.49 toks/s, output: 3080.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:04<01:00,  1.23it/s, est. speed input: 80.32 toks/s, output: 3151.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:07<01:02,  1.16it/s, est. speed input: 79.81 toks/s, output: 3253.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:09<01:21,  1.15s/it, est. speed input: 78.19 toks/s, output: 3228.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:10<01:06,  1.05it/s, est. speed input: 78.73 toks/s, output: 3292.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5638.2 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:10<01:01,  1.12it/s, est. speed input: 78.76 toks/s, output: 3336.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:12<01:12,  1.06s/it, est. speed input: 77.96 toks/s, output: 3342.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:12<00:57,  1.17it/s, est. speed input: 78.49 toks/s, output: 3407.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:12<00:45,  1.46it/s, est. speed input: 79.06 toks/s, output: 3473.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:13<00:34,  1.87it/s, est. speed input: 79.72 toks/s, output: 3543.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:14<00:45,  1.42it/s, est. speed input: 79.34 toks/s, output: 3567.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:15<00:47,  1.30it/s, est. speed input: 79.18 toks/s, output: 3641.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:16<00:37,  1.63it/s, est. speed input: 79.83 toks/s, output: 3713.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:16<00:35,  1.67it/s, est. speed input: 80.03 toks/s, output: 3764.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:17<00:37,  1.59it/s, est. speed input: 80.04 toks/s, output: 3808.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:18<00:46,  1.26it/s, est. speed input: 79.56 toks/s, output: 3827.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:18<00:38,  1.49it/s, est. speed input: 79.96 toks/s, output: 3888.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:20<00:33,  1.62it/s, est. speed input: 80.44 toks/s, output: 3991.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5406.6 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:20<00:28,  1.87it/s, est. speed input: 80.91 toks/s, output: 4056.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:20<00:22,  2.29it/s, est. speed input: 81.82 toks/s, output: 4183.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:21<00:28,  1.78it/s, est. speed input: 81.57 toks/s, output: 4212.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:22<00:23,  2.11it/s, est. speed input: 82.10 toks/s, output: 4280.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:22<00:20,  2.40it/s, est. speed input: 82.60 toks/s, output: 4347.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:23<00:26,  1.78it/s, est. speed input: 82.37 toks/s, output: 4376.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:23<00:20,  2.29it/s, est. speed input: 83.01 toks/s, output: 4450.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:23<00:14,  3.02it/s, est. speed input: 84.09 toks/s, output: 4587.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:25<00:25,  1.69it/s, est. speed input: 83.41 toks/s, output: 4591.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:25<00:27,  1.58it/s, est. speed input: 83.38 toks/s, output: 4631.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:26<00:23,  1.75it/s, est. speed input: 83.70 toks/s, output: 4690.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:27<00:27,  1.47it/s, est. speed input: 83.49 toks/s, output: 4718.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:29<00:42,  1.05s/it, est. speed input: 82.32 toks/s, output: 4694.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:29<00:26,  1.46it/s, est. speed input: 83.27 toks/s, output: 4832.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4545.8 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:30<00:21,  1.64it/s, est. speed input: 83.69 toks/s, output: 4941.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:33<00:40,  1.15s/it, est. speed input: 81.60 toks/s, output: 4859.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:34<00:34,  1.01s/it, est. speed input: 81.76 toks/s, output: 4912.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:35<00:34,  1.05s/it, est. speed input: 81.41 toks/s, output: 4934.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:35<00:27,  1.17it/s, est. speed input: 81.80 toks/s, output: 5000.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:36<00:23,  1.34it/s, est. speed input: 82.06 toks/s, output: 5058.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:37<00:25,  1.16it/s, est. speed input: 81.70 toks/s, output: 5080.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:37<00:18,  1.55it/s, est. speed input: 82.25 toks/s, output: 5157.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:38<00:16,  1.63it/s, est. speed input: 82.51 toks/s, output: 5261.73 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:38<00:00,  1.63it/s, est. speed input: 99.05 toks/s, output: 7500.97 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:38<00:00,  1.62it/s, est. speed input: 99.05 toks/s, output: 7500.97 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:47 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:47 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:44:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:44:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:44:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:44:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:44:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:44:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 23:44:47 [abstract.py:306] It took 0.622948 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:44:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:44:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0105
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4630.6748
+  • Training FLOPS: 55.34 TFLOPS (13.84 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 475.07s
+  • policy_training: 306.09s (64.4%)
+  • generation: 106.01s (22.3%)
+  • teacher_logprob_inference: 58.15s (12.2%)
+  • policy_training/sharding_data: 4.33s (0.9%)
+  • prepare_for_generation: 3.30s (0.7%)
+  • policy_training/submit_training_futures: 1.83s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.45s (0.3%)
+  • training_prep: 0.84s (0.2%)
+  • teacher_logprob_inference_prep: 0.50s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 24: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:50:54 [abstract.py:324] It took 0.055200 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05783770e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05783770e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05783770e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05783770e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56847191e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69731216e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56847191e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69731216e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56847191e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69731216e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56847191e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69731216e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05803711e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54631271e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70300674e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23380260e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.09326410e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #24] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43339569e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05803711e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54631271e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70300674e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67263641e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27049427e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #24] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.18984222e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05803711e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54631271e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70300674e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27047043e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.05035877e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #24] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15305862e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05803711e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54631271e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70300674e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30905800e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75354595e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #24] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32695908e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 24: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 23:50:57 [abstract.py:324] It took 0.045880 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12993.00it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:50:57 [loggers.py:236] Engine 000: Avg prompt throughput: 25.2 tokens/s, Avg generation throughput: 74.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:38,  4.11it/s, est. speed input: 250.91 toks/s, output: 45.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:09, 16.92it/s, est. speed input: 860.99 toks/s, output: 208.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 28.33it/s, est. speed input: 1411.16 toks/s, output: 502.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 34.08it/s, est. speed input: 1670.27 toks/s, output: 770.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:12, 11.49it/s, est. speed input: 883.54 toks/s, output: 533.94 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:02<00:15,  8.91it/s, est. speed input: 728.66 toks/s, output: 637.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:03<00:35,  3.80it/s, est. speed input: 420.19 toks/s, output: 517.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:51:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12432.3 tokens/s, Running: 133 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:11<02:16,  1.03s/it, est. speed input: 150.84 toks/s, output: 302.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:11<02:06,  1.04it/s, est. speed input: 149.20 toks/s, output: 378.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:12<01:58,  1.09it/s, est. speed input: 146.32 toks/s, output: 448.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:16<03:20,  1.56s/it, est. speed input: 112.43 toks/s, output: 420.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:18<03:17,  1.55s/it, est. speed input: 106.47 toks/s, output: 473.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:51:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10748.6 tokens/s, Running: 128 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:20<03:29,  1.65s/it, est. speed input: 99.02 toks/s, output: 514.08 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:20<02:52,  1.37s/it, est. speed input: 99.41 toks/s, output: 586.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:21<02:25,  1.17s/it, est. speed input: 99.61 toks/s, output: 656.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:22<02:23,  1.15s/it, est. speed input: 97.32 toks/s, output: 709.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:26<04:11,  2.05s/it, est. speed input: 83.95 toks/s, output: 679.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:29<04:18,  2.12s/it, est. speed input: 79.34 toks/s, output: 709.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:29<02:29,  1.25s/it, est. speed input: 82.55 toks/s, output: 867.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:51:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9546.4 tokens/s, Running: 120 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:30<02:10,  1.09s/it, est. speed input: 82.96 toks/s, output: 933.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:30<01:40,  1.17it/s, est. speed input: 84.43 toks/s, output: 1011.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:30<01:25,  1.38it/s, est. speed input: 85.31 toks/s, output: 1082.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:32<01:40,  1.16it/s, est. speed input: 83.89 toks/s, output: 1123.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:32<01:28,  1.30it/s, est. speed input: 84.39 toks/s, output: 1188.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:33<01:19,  1.44it/s, est. speed input: 84.90 toks/s, output: 1252.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:34<01:29,  1.26it/s, est. speed input: 84.07 toks/s, output: 1296.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:34<01:21,  1.37it/s, est. speed input: 84.47 toks/s, output: 1357.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:37<02:20,  1.27s/it, est. speed input: 80.29 toks/s, output: 1346.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:37<02:00,  1.09s/it, est. speed input: 80.42 toks/s, output: 1403.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:38<01:49,  1.00s/it, est. speed input: 80.38 toks/s, output: 1456.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:51:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8466.9 tokens/s, Running: 108 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:41<02:05,  1.17s/it, est. speed input: 78.08 toks/s, output: 1517.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:41<01:17,  1.36it/s, est. speed input: 80.66 toks/s, output: 1673.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:41<01:05,  1.60it/s, est. speed input: 81.67 toks/s, output: 1744.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:41<00:52,  1.94it/s, est. speed input: 82.77 toks/s, output: 1818.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:42<00:48,  2.10it/s, est. speed input: 83.47 toks/s, output: 1883.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:42<00:41,  2.43it/s, est. speed input: 84.42 toks/s, output: 1953.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:42<00:36,  2.74it/s, est. speed input: 85.34 toks/s, output: 2023.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:43<00:37,  2.65it/s, est. speed input: 85.87 toks/s, output: 2084.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:44<01:02,  1.57it/s, est. speed input: 84.80 toks/s, output: 2105.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:44<00:47,  2.06it/s, est. speed input: 85.92 toks/s, output: 2180.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:46<01:25,  1.12it/s, est. speed input: 83.72 toks/s, output: 2172.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:48<01:44,  1.10s/it, est. speed input: 82.22 toks/s, output: 2181.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:49<01:52,  1.20s/it, est. speed input: 81.07 toks/s, output: 2198.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:51:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7451.7 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:50<01:17,  1.18it/s, est. speed input: 82.13 toks/s, output: 2320.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:50<01:06,  1.37it/s, est. speed input: 82.69 toks/s, output: 2383.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:51<01:15,  1.20it/s, est. speed input: 82.04 toks/s, output: 2410.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:52<01:14,  1.19it/s, est. speed input: 81.85 toks/s, output: 2451.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:54<01:07,  1.28it/s, est. speed input: 81.90 toks/s, output: 2544.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:55<01:19,  1.09it/s, est. speed input: 80.95 toks/s, output: 2561.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:57<01:34,  1.11s/it, est. speed input: 79.63 toks/s, output: 2565.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:58<01:44,  1.25s/it, est. speed input: 78.46 toks/s, output: 2574.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:59<01:25,  1.03s/it, est. speed input: 78.89 toks/s, output: 2633.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:59<01:04,  1.28it/s, est. speed input: 79.74 toks/s, output: 2706.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:59<00:52,  1.55it/s, est. speed input: 80.39 toks/s, output: 2771.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:51:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6529.6 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:03<02:11,  1.64s/it, est. speed input: 76.23 toks/s, output: 2674.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:04<01:56,  1.47s/it, est. speed input: 75.92 toks/s, output: 2709.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:05<01:24,  1.08s/it, est. speed input: 76.72 toks/s, output: 2783.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:06<01:32,  1.20s/it, est. speed input: 75.89 toks/s, output: 2799.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:06<01:15,  1.01it/s, est. speed input: 76.19 toks/s, output: 2857.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5733.6 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:13<03:08,  2.51s/it, est. speed input: 70.70 toks/s, output: 2698.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:13<02:29,  2.03s/it, est. speed input: 70.71 toks/s, output: 2743.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:14<01:46,  1.46s/it, est. speed input: 71.37 toks/s, output: 2816.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:14<01:23,  1.16s/it, est. speed input: 71.75 toks/s, output: 2877.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:15<01:10,  1.00it/s, est. speed input: 71.97 toks/s, output: 2931.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:15<00:49,  1.40it/s, est. speed input: 72.85 toks/s, output: 3057.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:16<00:41,  1.64it/s, est. speed input: 73.39 toks/s, output: 3124.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:16<00:32,  2.07it/s, est. speed input: 74.07 toks/s, output: 3197.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:16<00:21,  3.06it/s, est. speed input: 75.44 toks/s, output: 3343.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:17<00:33,  1.89it/s, est. speed input: 75.04 toks/s, output: 3370.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:17<00:20,  2.96it/s, est. speed input: 76.50 toks/s, output: 3522.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:18<00:23,  2.63it/s, est. speed input: 76.75 toks/s, output: 3577.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:18<00:25,  2.34it/s, est. speed input: 76.93 toks/s, output: 3629.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:19<00:26,  2.21it/s, est. speed input: 77.18 toks/s, output: 3683.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:19<00:24,  2.40it/s, est. speed input: 77.62 toks/s, output: 3747.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5485.0 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:20<00:28,  1.99it/s, est. speed input: 77.70 toks/s, output: 3792.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:21<00:28,  1.95it/s, est. speed input: 77.94 toks/s, output: 3846.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:22<00:42,  1.28it/s, est. speed input: 77.34 toks/s, output: 3858.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:22<00:31,  1.71it/s, est. speed input: 77.96 toks/s, output: 3931.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:23<00:40,  1.32it/s, est. speed input: 77.58 toks/s, output: 3955.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:24<00:33,  1.55it/s, est. speed input: 77.96 toks/s, output: 4017.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:24<00:20,  2.43it/s, est. speed input: 79.14 toks/s, output: 4162.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:25<00:24,  2.00it/s, est. speed input: 79.13 toks/s, output: 4204.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:26<00:32,  1.47it/s, est. speed input: 78.74 toks/s, output: 4225.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:27<00:31,  1.47it/s, est. speed input: 78.82 toks/s, output: 4272.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:28<00:43,  1.05it/s, est. speed input: 78.02 toks/s, output: 4272.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4518.1 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:30<00:57,  1.28s/it, est. speed input: 76.87 toks/s, output: 4253.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:31<00:41,  1.05it/s, est. speed input: 77.42 toks/s, output: 4327.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:33<00:59,  1.38s/it, est. speed input: 76.04 toks/s, output: 4294.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:33<00:44,  1.05s/it, est. speed input: 76.47 toks/s, output: 4363.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:33<00:33,  1.24it/s, est. speed input: 76.92 toks/s, output: 4433.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:36<00:55,  1.39s/it, est. speed input: 75.33 toks/s, output: 4387.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:37<00:46,  1.20s/it, est. speed input: 75.38 toks/s, output: 4434.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:37<00:27,  1.36it/s, est. speed input: 76.29 toks/s, output: 4578.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:38<00:22,  1.57it/s, est. speed input: 76.68 toks/s, output: 4643.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:38<00:15,  2.22it/s, est. speed input: 77.62 toks/s, output: 4787.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:38<00:13,  2.38it/s, est. speed input: 77.99 toks/s, output: 4853.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:39<00:16,  1.98it/s, est. speed input: 78.02 toks/s, output: 4897.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4005.6 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:40<00:16,  1.84it/s, est. speed input: 78.12 toks/s, output: 4947.51 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:40<00:00,  1.84it/s, est. speed input: 96.77 toks/s, output: 7479.96 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:40<00:00,  1.60it/s, est. speed input: 96.77 toks/s, output: 7479.96 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:44 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:44 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-02 23:52:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:52:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:52:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:52:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-02 23:52:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-02 23:52:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-02 23:52:45 [abstract.py:306] It took 0.567151 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-02 23:52:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-02 23:52:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) 
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0126
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 4687.9688
+  • Training FLOPS: 55.68 TFLOPS (13.92 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 480.98s
+  • policy_training: 309.21s (64.3%)
+  • generation: 107.76s (22.4%)
+  • teacher_logprob_inference: 58.88s (12.2%)
+  • policy_training/sharding_data: 4.38s (0.9%)
+  • prepare_for_generation: 3.29s (0.7%)
+  • policy_training/submit_training_futures: 1.95s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.49s (0.3%)
+  • training_prep: 1.22s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 25: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-02 23:58:55 [abstract.py:324] It took 0.052247 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05819297e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56635628e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70156250e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05819297e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56635628e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70156250e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05819297e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56635628e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70156250e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05819297e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56635628e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70156250e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05834414e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53972092e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71730042e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23794041e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.15951109e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #25] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43045242e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05834414e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53972092e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71730042e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65675354e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28319082e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #25] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.16512146e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05834414e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53972092e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71730042e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.26779022e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.10904551e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #25] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15340881e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05834414e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53972092e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71730042e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31934166e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75514622e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #25] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32643417e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 25: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-02 23:58:58 [abstract.py:324] It took 0.043919 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12967.65it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:58:59 [loggers.py:236] Engine 000: Avg prompt throughput: 25.2 tokens/s, Avg generation throughput: 1.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:01<02:57,  1.12s/it, est. speed input: 53.71 toks/s, output: 15.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:01<00:33,  4.56it/s, est. speed input: 225.47 toks/s, output: 88.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:01<00:32,  4.81it/s, est. speed input: 240.20 toks/s, output: 115.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:01<00:23,  6.38it/s, est. speed input: 290.10 toks/s, output: 183.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:02<00:36,  4.16it/s, est. speed input: 247.17 toks/s, output: 193.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:02<00:35,  4.28it/s, est. speed input: 249.79 toks/s, output: 234.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:02<00:34,  4.28it/s, est. speed input: 250.06 toks/s, output: 274.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:04<01:45,  1.41it/s, est. speed input: 155.25 toks/s, output: 230.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:04<01:20,  1.82it/s, est. speed input: 163.18 toks/s, output: 297.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:10<04:33,  1.87s/it, est. speed input: 84.31 toks/s, output: 223.27 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-02 23:59:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 13089.7 tokens/s, Running: 146 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:12<04:57,  2.05s/it, est. speed input: 72.58 toks/s, output: 259.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:14<03:34,  1.50s/it, est. speed input: 72.80 toks/s, output: 381.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:16<03:53,  1.64s/it, est. speed input: 67.26 toks/s, output: 412.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:16<03:16,  1.39s/it, est. speed input: 68.09 toks/s, output: 476.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:18<03:38,  1.56s/it, est. speed input: 63.98 toks/s, output: 505.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:19<02:59,  1.29s/it, est. speed input: 65.09 toks/s, output: 570.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:19<02:12,  1.04it/s, est. speed input: 67.72 toks/s, output: 645.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:20<01:56,  1.17it/s, est. speed input: 68.72 toks/s, output: 706.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:59:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11023.2 tokens/s, Running: 137 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.5%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:22<02:41,  1.18s/it, est. speed input: 65.24 toks/s, output: 722.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:23<02:45,  1.23s/it, est. speed input: 64.07 toks/s, output: 760.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:26<03:44,  1.68s/it, est. speed input: 59.62 toks/s, output: 759.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:59:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9784.2 tokens/s, Running: 134 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.7%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:31<05:58,  2.70s/it, est. speed input: 51.79 toks/s, output: 712.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:33<05:34,  2.53s/it, est. speed input: 50.30 toks/s, output: 743.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:35<04:57,  2.27s/it, est. speed input: 49.59 toks/s, output: 783.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:35<03:37,  1.68s/it, est. speed input: 50.93 toks/s, output: 853.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:35<02:38,  1.22s/it, est. speed input: 52.42 toks/s, output: 925.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:37<02:40,  1.26s/it, est. speed input: 52.13 toks/s, output: 967.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:37<01:57,  1.08it/s, est. speed input: 53.55 toks/s, output: 1039.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:38<02:13,  1.06s/it, est. speed input: 53.12 toks/s, output: 1077.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:59:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8763.2 tokens/s, Running: 125 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:41<02:21,  1.14s/it, est. speed input: 52.90 toks/s, output: 1158.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:41<01:44,  1.17it/s, est. speed input: 54.76 toks/s, output: 1283.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:42<01:43,  1.17it/s, est. speed input: 55.03 toks/s, output: 1331.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:43<01:30,  1.32it/s, est. speed input: 55.84 toks/s, output: 1392.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:43<01:11,  1.65it/s, est. speed input: 57.60 toks/s, output: 1515.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:44<01:12,  1.61it/s, est. speed input: 58.05 toks/s, output: 1567.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:47<02:11,  1.13s/it, est. speed input: 55.99 toks/s, output: 1551.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:48<02:00,  1.04s/it, est. speed input: 56.35 toks/s, output: 1599.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:48<01:31,  1.25it/s, est. speed input: 57.41 toks/s, output: 1669.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:48<01:13,  1.54it/s, est. speed input: 58.34 toks/s, output: 1734.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:48<00:47,  2.36it/s, est. speed input: 60.46 toks/s, output: 1872.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:49<00:59,  1.84it/s, est. speed input: 60.55 toks/s, output: 1911.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:59:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8036.0 tokens/s, Running: 110 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.0%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:51<01:23,  1.30it/s, est. speed input: 60.02 toks/s, output: 1931.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:54<02:42,  1.50s/it, est. speed input: 57.29 toks/s, output: 1881.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:54<02:00,  1.13s/it, est. speed input: 58.22 toks/s, output: 1949.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:56<01:41,  1.04it/s, est. speed input: 58.78 toks/s, output: 2041.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:57<01:44,  1.01s/it, est. speed input: 58.67 toks/s, output: 2074.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:58<01:41,  1.01it/s, est. speed input: 58.77 toks/s, output: 2114.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:58<01:22,  1.23it/s, est. speed input: 59.45 toks/s, output: 2175.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:59<01:09,  1.45it/s, est. speed input: 60.10 toks/s, output: 2235.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:59<00:54,  1.85it/s, est. speed input: 60.94 toks/s, output: 2302.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [01:00<00:59,  1.66it/s, est. speed input: 61.58 toks/s, output: 2395.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-02 23:59:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7466.3 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [01:01<01:06,  1.47it/s, est. speed input: 61.64 toks/s, output: 2432.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [01:01<00:55,  1.72it/s, est. speed input: 62.30 toks/s, output: 2494.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [01:02<00:40,  2.35it/s, est. speed input: 63.82 toks/s, output: 2623.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [01:02<00:32,  2.84it/s, est. speed input: 64.65 toks/s, output: 2691.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [01:05<01:35,  1.04s/it, est. speed input: 62.51 toks/s, output: 2635.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [01:06<01:11,  1.25it/s, est. speed input: 63.44 toks/s, output: 2744.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [01:07<01:13,  1.20it/s, est. speed input: 63.47 toks/s, output: 2778.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [01:08<01:14,  1.19it/s, est. speed input: 63.54 toks/s, output: 2815.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:10<01:46,  1.23s/it, est. speed input: 62.30 toks/s, output: 2795.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6360.7 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 20.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:12<01:56,  1.36s/it, est. speed input: 61.68 toks/s, output: 2802.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:12<01:33,  1.10s/it, est. speed input: 62.15 toks/s, output: 2857.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:13<01:24,  1.01s/it, est. speed input: 62.34 toks/s, output: 2900.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:14<01:23,  1.00s/it, est. speed input: 62.31 toks/s, output: 2934.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:14<01:02,  1.31it/s, est. speed input: 62.98 toks/s, output: 3000.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:15<00:52,  1.53it/s, est. speed input: 63.45 toks/s, output: 3057.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:16<01:03,  1.26it/s, est. speed input: 63.30 toks/s, output: 3085.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:16<00:40,  1.91it/s, est. speed input: 64.53 toks/s, output: 3213.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:16<00:38,  2.02it/s, est. speed input: 64.96 toks/s, output: 3269.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:17<00:36,  2.10it/s, est. speed input: 65.37 toks/s, output: 3323.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:17<00:29,  2.53it/s, est. speed input: 66.01 toks/s, output: 3389.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:18<00:30,  2.41it/s, est. speed input: 66.37 toks/s, output: 3441.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:19<00:44,  1.65it/s, est. speed input: 66.24 toks/s, output: 3466.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:19<00:33,  2.16it/s, est. speed input: 66.88 toks/s, output: 3534.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:20<00:51,  1.37it/s, est. speed input: 66.51 toks/s, output: 3547.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5864.2 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.2%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:22<01:21,  1.17s/it, est. speed input: 65.45 toks/s, output: 3525.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:23<01:06,  1.04it/s, est. speed input: 65.79 toks/s, output: 3578.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:24<01:07,  1.00it/s, est. speed input: 65.67 toks/s, output: 3605.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:26<01:19,  1.19s/it, est. speed input: 65.12 toks/s, output: 3609.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:26<01:05,  1.01it/s, est. speed input: 65.39 toks/s, output: 3659.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:26<00:53,  1.22it/s, est. speed input: 65.77 toks/s, output: 3714.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:27<00:46,  1.38it/s, est. speed input: 66.09 toks/s, output: 3765.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:27<00:34,  1.84it/s, est. speed input: 66.69 toks/s, output: 3832.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:28<00:46,  1.34it/s, est. speed input: 66.44 toks/s, output: 3852.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:30<00:59,  1.02it/s, est. speed input: 65.98 toks/s, output: 3861.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4885.2 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:31<00:58,  1.03it/s, est. speed input: 65.97 toks/s, output: 3893.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:31<00:46,  1.28it/s, est. speed input: 66.39 toks/s, output: 3952.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:31<00:34,  1.67it/s, est. speed input: 66.92 toks/s, output: 4017.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:32<00:30,  1.86it/s, est. speed input: 67.52 toks/s, output: 4123.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:33<00:30,  1.80it/s, est. speed input: 67.99 toks/s, output: 4217.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:34<00:25,  2.09it/s, est. speed input: 68.46 toks/s, output: 4281.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:35<00:33,  1.54it/s, est. speed input: 68.22 toks/s, output: 4301.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:36<00:41,  1.23it/s, est. speed input: 67.94 toks/s, output: 4318.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:37<00:43,  1.14it/s, est. speed input: 67.84 toks/s, output: 4346.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:38<00:36,  1.35it/s, est. speed input: 68.18 toks/s, output: 4403.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:38<00:28,  1.66it/s, est. speed input: 68.61 toks/s, output: 4466.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:38<00:27,  1.69it/s, est. speed input: 68.84 toks/s, output: 4515.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:38<00:21,  2.18it/s, est. speed input: 69.35 toks/s, output: 4583.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:39<00:17,  2.50it/s, est. speed input: 69.78 toks/s, output: 4645.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:39<00:10,  4.11it/s, est. speed input: 70.90 toks/s, output: 4789.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:39<00:09,  4.30it/s, est. speed input: 71.38 toks/s, output: 4854.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:39<00:10,  3.86it/s, est. speed input: 71.73 toks/s, output: 4912.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:40<00:17,  2.27it/s, est. speed input: 71.67 toks/s, output: 4942.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4708.6 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:41<00:21,  1.78it/s, est. speed input: 71.67 toks/s, output: 4974.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:42<00:14,  2.48it/s, est. speed input: 72.58 toks/s, output: 5105.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:42<00:12,  2.87it/s, est. speed input: 73.05 toks/s, output: 5171.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:42<00:10,  3.50it/s, est. speed input: 73.57 toks/s, output: 5241.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:42<00:08,  3.95it/s, est. speed input: 74.46 toks/s, output: 5370.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:43<00:11,  2.78it/s, est. speed input: 74.52 toks/s, output: 5409.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:44<00:17,  1.79it/s, est. speed input: 74.26 toks/s, output: 5425.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:45<00:15,  1.93it/s, est. speed input: 74.55 toks/s, output: 5481.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:45<00:15,  1.93it/s, est. speed input: 74.77 toks/s, output: 5531.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:46<00:14,  1.96it/s, est. speed input: 75.00 toks/s, output: 5582.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:46<00:09,  2.73it/s, est. speed input: 75.84 toks/s, output: 5716.28 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:46<00:00,  2.73it/s, est. speed input: 90.55 toks/s, output: 7716.98 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:46<00:00,  1.50it/s, est. speed input: 90.55 toks/s, output: 7716.98 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:52 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:52 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:00:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:00:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:00:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:00:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:00:53 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:00:53 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:00:53 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 00:00:53 [abstract.py:306] It took 0.625588 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:00:53 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0107
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 5134.5439
+  • Training FLOPS: 58.30 TFLOPS (14.57 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 505.38s
+  • policy_training: 322.74s (63.9%)
+  • generation: 114.50s (22.7%)
+  • teacher_logprob_inference: 63.24s (12.5%)
+  • policy_training/sharding_data: 4.39s (0.9%)
+  • prepare_for_generation: 3.40s (0.7%)
+  • policy_training/submit_training_futures: 2.03s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.45s (0.3%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+
+========================= Epoch 6/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 26: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:07:21 [abstract.py:324] It took 0.055640 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05865488e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56289520e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69077148e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05865488e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56289520e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69077148e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05865488e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56289520e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69077148e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05865488e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56289520e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69077148e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05872539e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53698654e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70503845e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22674904e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.24953890e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #26] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42892090e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05872539e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53698654e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70503845e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67202148e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29045773e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #26] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.18260574e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05872539e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53698654e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70503845e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27622910e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.27691841e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #26] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17599106e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05872539e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53698654e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.70503845e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30899506e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75969582e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #26] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32564987e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 26: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:07:24 [abstract.py:324] It took 0.044558 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12166.44it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:07:24 [loggers.py:236] Engine 000: Avg prompt throughput: 24.1 tokens/s, Avg generation throughput: 49.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.09it/s, est. speed input: 131.36 toks/s, output: 22.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:14, 10.58it/s, est. speed input: 525.49 toks/s, output: 142.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 31.16it/s, est. speed input: 1310.38 toks/s, output: 494.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 33.15it/s, est. speed input: 1483.77 toks/s, output: 684.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:05, 25.86it/s, est. speed input: 1384.66 toks/s, output: 782.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:07, 18.31it/s, est. speed input: 1192.27 toks/s, output: 825.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:09, 13.34it/s, est. speed input: 1021.05 toks/s, output: 869.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:03<00:18,  6.63it/s, est. speed input: 699.94 toks/s, output: 775.14 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:06<00:51,  2.35it/s, est. speed input: 364.07 toks/s, output: 526.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:09<01:21,  1.47it/s, est. speed input: 253.94 toks/s, output: 507.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:07:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12166.9 tokens/s, Running: 120 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:11<01:33,  1.28it/s, est. speed input: 225.57 toks/s, output: 533.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:11<01:30,  1.30it/s, est. speed input: 217.48 toks/s, output: 596.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:13<01:54,  1.02it/s, est. speed input: 191.32 toks/s, output: 605.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:16<02:01,  1.05s/it, est. speed input: 170.75 toks/s, output: 688.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:16<01:39,  1.15it/s, est. speed input: 173.20 toks/s, output: 775.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:16<01:20,  1.41it/s, est. speed input: 175.44 toks/s, output: 860.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:16<01:15,  1.49it/s, est. speed input: 173.28 toks/s, output: 925.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:17<01:09,  1.61it/s, est. speed input: 171.91 toks/s, output: 992.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:18<01:30,  1.22it/s, est. speed input: 162.80 toks/s, output: 1012.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:20<01:51,  1.03s/it, est. speed input: 153.16 toks/s, output: 1025.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:07:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10162.7 tokens/s, Running: 109 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.4%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:20<01:28,  1.22it/s, est. speed input: 153.90 toks/s, output: 1102.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:21<01:14,  1.43it/s, est. speed input: 153.92 toks/s, output: 1171.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:21<00:57,  1.85it/s, est. speed input: 155.55 toks/s, output: 1254.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:22<01:09,  1.51it/s, est. speed input: 151.53 toks/s, output: 1291.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:22<00:55,  1.86it/s, est. speed input: 152.82 toks/s, output: 1368.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:23<01:15,  1.37it/s, est. speed input: 147.74 toks/s, output: 1390.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:24<01:31,  1.11it/s, est. speed input: 142.45 toks/s, output: 1408.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:25<01:13,  1.37it/s, est. speed input: 143.08 toks/s, output: 1480.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:25<01:02,  1.60it/s, est. speed input: 143.36 toks/s, output: 1548.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:26<00:59,  1.66it/s, est. speed input: 142.61 toks/s, output: 1606.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:26<00:46,  2.10it/s, est. speed input: 143.87 toks/s, output: 1685.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:27<00:52,  1.84it/s, est. speed input: 142.33 toks/s, output: 1731.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:27<00:47,  2.01it/s, est. speed input: 142.12 toks/s, output: 1854.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:29<01:06,  1.42it/s, est. speed input: 137.76 toks/s, output: 1859.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:29<01:01,  1.50it/s, est. speed input: 137.24 toks/s, output: 1914.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:29<00:47,  1.95it/s, est. speed input: 138.77 toks/s, output: 1998.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:07:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8605.5 tokens/s, Running: 91 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.4%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:30<00:45,  2.01it/s, est. speed input: 138.61 toks/s, output: 2057.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:30<00:47,  1.89it/s, est. speed input: 137.82 toks/s, output: 2106.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:31<00:37,  2.35it/s, est. speed input: 138.99 toks/s, output: 2184.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:32<00:37,  2.30it/s, est. speed input: 139.01 toks/s, output: 2299.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:32<00:26,  3.17it/s, est. speed input: 141.62 toks/s, output: 2459.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:32<00:26,  3.16it/s, est. speed input: 142.10 toks/s, output: 2524.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:33<00:28,  2.96it/s, est. speed input: 142.16 toks/s, output: 2582.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:33<00:26,  3.13it/s, est. speed input: 142.92 toks/s, output: 2651.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:33<00:16,  4.74it/s, est. speed input: 146.05 toks/s, output: 2820.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:34<00:31,  2.54it/s, est. speed input: 143.51 toks/s, output: 2827.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:34<00:29,  2.60it/s, est. speed input: 143.85 toks/s, output: 2888.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:36<00:56,  1.36it/s, est. speed input: 138.77 toks/s, output: 2841.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:36<00:37,  2.01it/s, est. speed input: 140.72 toks/s, output: 2992.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:37<00:37,  1.96it/s, est. speed input: 140.31 toks/s, output: 3038.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:37<00:32,  2.27it/s, est. speed input: 141.05 toks/s, output: 3109.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:38<00:34,  2.07it/s, est. speed input: 140.27 toks/s, output: 3198.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:38<00:27,  2.52it/s, est. speed input: 141.43 toks/s, output: 3278.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:39<00:24,  2.81it/s, est. speed input: 142.38 toks/s, output: 3407.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:39<00:24,  2.70it/s, est. speed input: 142.42 toks/s, output: 3460.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:40<00:22,  2.92it/s, est. speed input: 143.03 toks/s, output: 3528.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7017.1 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:40<00:25,  2.59it/s, est. speed input: 142.72 toks/s, output: 3573.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:40<00:20,  3.05it/s, est. speed input: 143.61 toks/s, output: 3647.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:41<00:23,  2.74it/s, est. speed input: 143.42 toks/s, output: 3696.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:39,  1.55it/s, est. speed input: 140.37 toks/s, output: 3669.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:42<00:31,  1.95it/s, est. speed input: 141.18 toks/s, output: 3743.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:42<00:19,  3.10it/s, est. speed input: 143.42 toks/s, output: 3908.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:43<00:25,  2.30it/s, est. speed input: 142.25 toks/s, output: 3928.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:44<00:28,  1.96it/s, est. speed input: 140.98 toks/s, output: 3999.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:46<00:36,  1.49it/s, est. speed input: 138.60 toks/s, output: 3985.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:46<00:36,  1.48it/s, est. speed input: 137.86 toks/s, output: 4018.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:47<00:28,  1.83it/s, est. speed input: 138.64 toks/s, output: 4095.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:47<00:22,  2.25it/s, est. speed input: 139.47 toks/s, output: 4225.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:47<00:18,  2.70it/s, est. speed input: 140.37 toks/s, output: 4306.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:48<00:24,  2.01it/s, est. speed input: 139.09 toks/s, output: 4320.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:49<00:28,  1.70it/s, est. speed input: 137.92 toks/s, output: 4339.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6001.1 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:50<00:34,  1.37it/s, est. speed input: 136.09 toks/s, output: 4337.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:51<00:30,  1.50it/s, est. speed input: 135.93 toks/s, output: 4389.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:51<00:30,  1.50it/s, est. speed input: 135.31 toks/s, output: 4425.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:52<00:28,  1.56it/s, est. speed input: 134.97 toks/s, output: 4469.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:52<00:18,  2.23it/s, est. speed input: 136.11 toks/s, output: 4618.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:53<00:19,  2.14it/s, est. speed input: 135.89 toks/s, output: 4666.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:54<00:22,  1.76it/s, est. speed input: 134.85 toks/s, output: 4685.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:54<00:17,  2.24it/s, est. speed input: 135.66 toks/s, output: 4770.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:54<00:14,  2.60it/s, est. speed input: 136.18 toks/s, output: 4844.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:54<00:11,  3.24it/s, est. speed input: 137.08 toks/s, output: 4927.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:55<00:16,  2.20it/s, est. speed input: 136.13 toks/s, output: 4949.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:55<00:11,  2.95it/s, est. speed input: 137.33 toks/s, output: 5102.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:56<00:09,  3.35it/s, est. speed input: 137.97 toks/s, output: 5181.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:56<00:12,  2.52it/s, est. speed input: 137.37 toks/s, output: 5214.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:56<00:10,  2.89it/s, est. speed input: 137.96 toks/s, output: 5290.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:57<00:09,  3.13it/s, est. speed input: 138.44 toks/s, output: 5362.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:57<00:11,  2.48it/s, est. speed input: 137.99 toks/s, output: 5401.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:57<00:09,  2.99it/s, est. speed input: 138.72 toks/s, output: 5482.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:58<00:07,  3.69it/s, est. speed input: 139.53 toks/s, output: 5567.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:58<00:07,  3.51it/s, est. speed input: 139.79 toks/s, output: 5633.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:58<00:05,  4.20it/s, est. speed input: 140.53 toks/s, output: 5717.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:58<00:05,  4.47it/s, est. speed input: 141.13 toks/s, output: 5795.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:58<00:04,  5.04it/s, est. speed input: 141.87 toks/s, output: 5878.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:59<00:04,  4.93it/s, est. speed input: 142.36 toks/s, output: 5954.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:59<00:05,  3.63it/s, est. speed input: 142.71 toks/s, output: 6075.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:59<00:03,  4.86it/s, est. speed input: 144.31 toks/s, output: 6250.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4171.5 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:00<00:03,  5.02it/s, est. speed input: 145.48 toks/s, output: 6407.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:03<00:10,  1.42it/s, est. speed input: 140.32 toks/s, output: 6239.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:03<00:08,  1.63it/s, est. speed input: 140.57 toks/s, output: 6309.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:03<00:06,  1.94it/s, est. speed input: 141.08 toks/s, output: 6391.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:05<00:09,  1.24it/s, est. speed input: 138.45 toks/s, output: 6335.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:05<00:04,  2.00it/s, est. speed input: 139.94 toks/s, output: 6526.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:05<00:04,  2.18it/s, est. speed input: 140.15 toks/s, output: 6599.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:05<00:03,  2.51it/s, est. speed input: 140.59 toks/s, output: 6682.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:06<00:03,  1.77it/s, est. speed input: 139.39 toks/s, output: 6686.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:07<00:02,  2.25it/s, est. speed input: 140.04 toks/s, output: 6781.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:08<00:04,  1.16it/s, est. speed input: 137.01 toks/s, output: 6701.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:09<00:02,  1.40it/s, est. speed input: 137.26 toks/s, output: 6778.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:09<00:01,  1.65it/s, est. speed input: 137.42 toks/s, output: 6856.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1978.9 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:11<00:01,  1.07it/s, est. speed input: 134.93 toks/s, output: 6805.71 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:11<00:00,  1.07it/s, est. speed input: 136.61 toks/s, output: 7035.20 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:11<00:00,  2.24it/s, est. speed input: 136.61 toks/s, output: 7035.20 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:40 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:40 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:08:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:08:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:08:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:08:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:08:40 [abstract.py:306] It took 0.579567 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:08:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:08:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:08:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:08:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0164
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3138.8313
+  • Training FLOPS: 47.82 TFLOPS (11.95 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 360.13s
+  • policy_training: 233.63s (64.9%)
+  • generation: 76.52s (21.2%)
+  • teacher_logprob_inference: 45.06s (12.5%)
+  • policy_training/sharding_data: 4.33s (1.2%)
+  • prepare_for_generation: 3.46s (1.0%)
+  • policy_training/submit_training_futures: 2.03s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.4%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.47s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 27: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:13:21 [abstract.py:324] It took 0.054984 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05897461e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55831223e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68039398e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05897461e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55831223e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68039398e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05897461e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55831223e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68039398e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05897461e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55831223e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68039398e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05890391e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53141327e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68065033e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23116741e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.32011843e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #27] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43022339e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05890391e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53141327e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68065033e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68445740e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29580536e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #27] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.17834511e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05890391e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53141327e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68065033e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28761597e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.07292509e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #27] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15733871e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05890391e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53141327e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68065033e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29951782e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.76480064e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #27] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32459610e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 27: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:13:24 [abstract.py:324] It took 0.042452 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12531.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:13:24 [loggers.py:236] Engine 000: Avg prompt throughput: 33.6 tokens/s, Avg generation throughput: 3.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:45,  1.51it/s, est. speed input: 92.16 toks/s, output: 22.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:08, 18.45it/s, est. speed input: 888.71 toks/s, output: 297.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:06, 23.17it/s, est. speed input: 1086.49 toks/s, output: 460.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:06, 21.94it/s, est. speed input: 1106.75 toks/s, output: 577.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:07, 17.19it/s, est. speed input: 1011.09 toks/s, output: 675.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:08, 15.52it/s, est. speed input: 970.77 toks/s, output: 767.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:06<00:58,  2.20it/s, est. speed input: 291.74 toks/s, output: 349.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:13:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12259.5 tokens/s, Running: 129 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:12<02:03,  1.02it/s, est. speed input: 159.04 toks/s, output: 355.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:14<02:20,  1.11s/it, est. speed input: 140.74 toks/s, output: 393.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:18<03:01,  1.45s/it, est. speed input: 117.93 toks/s, output: 407.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:19<02:55,  1.42s/it, est. speed input: 113.48 toks/s, output: 467.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:19<02:29,  1.21s/it, est. speed input: 114.17 toks/s, output: 545.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:20<02:01,  1.00it/s, est. speed input: 115.79 toks/s, output: 625.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:13:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10594.9 tokens/s, Running: 122 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:20<01:55,  1.05it/s, est. speed input: 114.07 toks/s, output: 686.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:24<03:12,  1.61s/it, est. speed input: 100.28 toks/s, output: 674.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:25<02:59,  1.51s/it, est. speed input: 97.76 toks/s, output: 726.21 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:27<03:11,  1.63s/it, est. speed input: 93.15 toks/s, output: 760.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:28<03:03,  1.57s/it, est. speed input: 90.61 toks/s, output: 806.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:13:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9480.0 tokens/s, Running: 117 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:32<04:08,  2.14s/it, est. speed input: 82.64 toks/s, output: 802.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:32<03:00,  1.57s/it, est. speed input: 84.07 toks/s, output: 881.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:32<02:10,  1.14s/it, est. speed input: 85.63 toks/s, output: 961.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:34<02:32,  1.35s/it, est. speed input: 82.81 toks/s, output: 992.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:35<02:22,  1.27s/it, est. speed input: 82.05 toks/s, output: 1045.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:38<03:26,  1.86s/it, est. speed input: 76.73 toks/s, output: 1040.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:14:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8115.5 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.8%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:40<03:26,  1.88s/it, est. speed input: 74.66 toks/s, output: 1072.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:42<03:02,  1.67s/it, est. speed input: 74.02 toks/s, output: 1122.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:43<02:54,  1.61s/it, est. speed input: 72.93 toks/s, output: 1165.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:45<02:22,  1.34s/it, est. speed input: 72.30 toks/s, output: 1269.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:49<03:15,  1.86s/it, est. speed input: 68.50 toks/s, output: 1260.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:14:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7638.9 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.3%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:54<04:55,  2.85s/it, est. speed input: 62.54 toks/s, output: 1209.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:56<04:16,  2.49s/it, est. speed input: 61.89 toks/s, output: 1254.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:56<03:17,  1.93s/it, est. speed input: 62.46 toks/s, output: 1321.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:58<03:07,  1.85s/it, est. speed input: 61.77 toks/s, output: 1361.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:58<02:18,  1.39s/it, est. speed input: 62.58 toks/s, output: 1433.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:59<02:17,  1.39s/it, est. speed input: 62.15 toks/s, output: 1478.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:14:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7245.9 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 21.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [01:03<03:19,  2.04s/it, est. speed input: 59.55 toks/s, output: 1471.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [01:03<02:29,  1.54s/it, est. speed input: 60.20 toks/s, output: 1540.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [01:05<02:22,  1.49s/it, est. speed input: 59.84 toks/s, output: 1584.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [01:06<02:13,  1.41s/it, est. speed input: 59.66 toks/s, output: 1632.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [01:06<01:38,  1.05s/it, est. speed input: 60.38 toks/s, output: 1703.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [01:08<02:01,  1.31s/it, est. speed input: 59.61 toks/s, output: 1732.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:14:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6423.8 tokens/s, Running: 92 reqs, Waiting: 0 reqs, GPU KV cache usage: 22.7%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [01:14<03:10,  2.09s/it, est. speed input: 56.47 toks/s, output: 1738.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [01:15<02:40,  1.78s/it, est. speed input: 56.59 toks/s, output: 1794.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [01:18<03:16,  2.21s/it, est. speed input: 54.93 toks/s, output: 1792.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [01:19<02:41,  1.84s/it, est. speed input: 55.12 toks/s, output: 1848.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:14:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5847.3 tokens/s, Running: 88 reqs, Waiting: 0 reqs, GPU KV cache usage: 24.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [01:21<02:34,  1.78s/it, est. speed input: 54.75 toks/s, output: 1886.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [01:24<03:14,  2.26s/it, est. speed input: 53.23 toks/s, output: 1883.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [01:25<02:25,  1.71s/it, est. speed input: 53.71 toks/s, output: 1949.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:25<01:53,  1.35s/it, est. speed input: 54.12 toks/s, output: 2013.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:25<01:22,  1.01it/s, est. speed input: 54.76 toks/s, output: 2083.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [01:26<01:06,  1.24it/s, est. speed input: 55.25 toks/s, output: 2149.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:26<00:51,  1.58it/s, est. speed input: 55.81 toks/s, output: 2218.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:28<01:14,  1.07it/s, est. speed input: 55.47 toks/s, output: 2250.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:14:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5638.4 tokens/s, Running: 80 reqs, Waiting: 0 reqs, GPU KV cache usage: 24.7%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [01:31<02:01,  1.54s/it, est. speed input: 54.36 toks/s, output: 2251.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:31<01:37,  1.25s/it, est. speed input: 54.68 toks/s, output: 2311.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [01:35<02:43,  2.12s/it, est. speed input: 52.93 toks/s, output: 2283.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:36<02:14,  1.78s/it, est. speed input: 53.06 toks/s, output: 2334.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:38<02:11,  1.75s/it, est. speed input: 52.76 toks/s, output: 2368.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:15:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5136.7 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 25.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:43<03:12,  2.61s/it, est. speed input: 50.98 toks/s, output: 2335.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:43<02:25,  2.00s/it, est. speed input: 51.30 toks/s, output: 2394.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:44<01:50,  1.54s/it, est. speed input: 51.68 toks/s, output: 2457.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:44<01:25,  1.20s/it, est. speed input: 52.06 toks/s, output: 2520.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:47<01:31,  1.32s/it, est. speed input: 51.81 toks/s, output: 2595.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:15:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4383.3 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 25.3%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:51<02:25,  2.13s/it, est. speed input: 50.21 toks/s, output: 2560.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:52<01:55,  1.72s/it, est. speed input: 50.50 toks/s, output: 2619.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:52<01:25,  1.29s/it, est. speed input: 50.98 toks/s, output: 2688.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:53<01:08,  1.05s/it, est. speed input: 51.34 toks/s, output: 2750.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:53<01:00,  1.06it/s, est. speed input: 51.59 toks/s, output: 2805.74 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:53<00:00,  1.06it/s, est. speed input: 85.76 toks/s, output: 7412.60 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:53<00:00,  1.41it/s, est. speed input: 85.76 toks/s, output: 7412.60 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:15:25 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:15:25 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:15:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:15:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:15:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:15:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:15:26 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:15:26 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:15:26 [abstract.py:306] It took 0.629595 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:15:26 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:15:26 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 5272.4438
+  • Training FLOPS: 61.77 TFLOPS (15.44 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 504.46s
+  • policy_training: 318.65s (63.2%)
+  • generation: 121.97s (24.2%)
+  • teacher_logprob_inference: 59.06s (11.7%)
+  • policy_training/sharding_data: 4.28s (0.8%)
+  • prepare_for_generation: 3.32s (0.7%)
+  • policy_training/submit_training_futures: 1.78s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.3%)
+  • training_prep: 0.81s (0.2%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 28: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:21:46 [abstract.py:324] It took 0.054682 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05952969e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55938911e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67054596e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05952969e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55938911e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67054596e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05952969e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55938911e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67054596e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05952969e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55938911e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67054596e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05955898e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53645020e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66852875e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23177681e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.33510256e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #28] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43019363e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05955898e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53645020e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66852875e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69695053e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28674269e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #28] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.19958115e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05955898e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53645020e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66852875e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28016510e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=5.16653299e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #28] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15715027e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05955898e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53645020e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66852875e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30642242e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.76237488e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #28] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32224319e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 28: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:21:49 [abstract.py:324] It took 0.042906 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 00:21:49 [loggers.py:236] Engine 000: Avg prompt throughput: 24.7 tokens/s, Avg generation throughput: 32.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12022.80it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:29,  5.30it/s, est. speed input: 318.30 toks/s, output: 68.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 29.16it/s, est. speed input: 1525.53 toks/s, output: 467.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 44.58it/s, est. speed input: 2245.06 toks/s, output: 892.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 40.70it/s, est. speed input: 2217.61 toks/s, output: 1044.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:04, 32.69it/s, est. speed input: 2006.32 toks/s, output: 1159.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:05, 25.57it/s, est. speed input: 1768.63 toks/s, output: 1212.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:05, 23.53it/s, est. speed input: 1683.29 toks/s, output: 1305.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:05, 23.67it/s, est. speed input: 1665.56 toks/s, output: 1456.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:03<00:30,  4.01it/s, est. speed input: 621.06 toks/s, output: 692.02 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:06<00:55,  2.15it/s, est. speed input: 384.90 toks/s, output: 567.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:08<01:05,  1.79it/s, est. speed input: 317.05 toks/s, output: 628.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:09<01:11,  1.63it/s, est. speed input: 291.01 toks/s, output: 662.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:21:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11982.4 tokens/s, Running: 116 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:11<01:48,  1.06it/s, est. speed input: 231.62 toks/s, output: 612.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:13<02:00,  1.06s/it, est. speed input: 209.36 toks/s, output: 637.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:14<02:03,  1.09s/it, est. speed input: 196.05 toks/s, output: 680.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:16<02:30,  1.35s/it, est. speed input: 174.33 toks/s, output: 686.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:18<02:25,  1.31s/it, est. speed input: 166.02 toks/s, output: 735.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:18<02:00,  1.09s/it, est. speed input: 164.94 toks/s, output: 810.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:18<01:31,  1.19it/s, est. speed input: 166.65 toks/s, output: 896.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:22:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10081.7 tokens/s, Running: 109 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:21<02:23,  1.33s/it, est. speed input: 149.25 toks/s, output: 881.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:22<02:28,  1.38s/it, est. speed input: 141.95 toks/s, output: 915.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:22<01:48,  1.02s/it, est. speed input: 143.77 toks/s, output: 1002.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:26<02:51,  1.63s/it, est. speed input: 128.92 toks/s, output: 973.93 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:26<02:14,  1.29s/it, est. speed input: 128.89 toks/s, output: 1047.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:27<01:58,  1.15s/it, est. speed input: 127.16 toks/s, output: 1107.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:27<01:30,  1.13it/s, est. speed input: 128.16 toks/s, output: 1188.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:29<01:49,  1.09s/it, est. speed input: 123.40 toks/s, output: 1214.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:22:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8768.9 tokens/s, Running: 101 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:30<01:54,  1.15s/it, est. speed input: 120.14 toks/s, output: 1253.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:31<01:48,  1.09s/it, est. speed input: 118.35 toks/s, output: 1304.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:32<01:32,  1.05it/s, est. speed input: 118.04 toks/s, output: 1369.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:32<01:08,  1.41it/s, est. speed input: 119.38 toks/s, output: 1452.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:32<00:59,  1.61it/s, est. speed input: 119.70 toks/s, output: 1523.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:32<00:47,  2.02it/s, est. speed input: 120.84 toks/s, output: 1603.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:33<01:05,  1.44it/s, est. speed input: 118.50 toks/s, output: 1637.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:34<00:48,  1.90it/s, est. speed input: 119.81 toks/s, output: 1720.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:34<00:51,  1.77it/s, est. speed input: 119.35 toks/s, output: 1777.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:36<01:29,  1.02it/s, est. speed input: 114.66 toks/s, output: 1771.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:36<01:05,  1.37it/s, est. speed input: 115.85 toks/s, output: 1854.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:37<00:58,  1.52it/s, est. speed input: 115.96 toks/s, output: 1917.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:38<01:13,  1.20it/s, est. speed input: 113.77 toks/s, output: 1944.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:39<01:09,  1.25it/s, est. speed input: 113.25 toks/s, output: 1996.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:22:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7680.1 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:40<01:13,  1.17it/s, est. speed input: 112.00 toks/s, output: 2035.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:40<01:02,  1.36it/s, est. speed input: 112.24 toks/s, output: 2101.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:42<01:39,  1.18s/it, est. speed input: 107.85 toks/s, output: 2080.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:43<01:19,  1.05it/s, est. speed input: 108.16 toks/s, output: 2147.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:43<01:03,  1.29it/s, est. speed input: 108.67 toks/s, output: 2217.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:43<00:46,  1.73it/s, est. speed input: 109.77 toks/s, output: 2299.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:44<00:52,  1.54it/s, est. speed input: 109.09 toks/s, output: 2345.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:46<01:08,  1.16it/s, est. speed input: 107.16 toks/s, output: 2363.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:47<00:59,  1.30it/s, est. speed input: 106.72 toks/s, output: 2470.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:48<00:59,  1.28it/s, est. speed input: 106.16 toks/s, output: 2514.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:48<00:47,  1.57it/s, est. speed input: 106.88 toks/s, output: 2589.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:48<00:37,  1.98it/s, est. speed input: 107.78 toks/s, output: 2668.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:48<00:35,  2.08it/s, est. speed input: 108.09 toks/s, output: 2732.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:22:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6562.1 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:51<00:55,  1.29it/s, est. speed input: 105.68 toks/s, output: 2780.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:52<00:58,  1.20it/s, est. speed input: 104.80 toks/s, output: 2812.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:52<00:55,  1.25it/s, est. speed input: 104.55 toks/s, output: 2862.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:53<00:50,  1.35it/s, est. speed input: 104.57 toks/s, output: 2917.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:54<00:46,  1.43it/s, est. speed input: 104.53 toks/s, output: 2972.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:54<00:37,  1.78it/s, est. speed input: 105.22 toks/s, output: 3046.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:55<00:25,  2.47it/s, est. speed input: 106.98 toks/s, output: 3258.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:55<00:22,  2.79it/s, est. speed input: 107.74 toks/s, output: 3334.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:55<00:18,  3.23it/s, est. speed input: 108.52 toks/s, output: 3412.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:57<00:37,  1.60it/s, est. speed input: 106.54 toks/s, output: 3405.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:57<00:34,  1.71it/s, est. speed input: 106.68 toks/s, output: 3464.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:58<00:32,  1.76it/s, est. speed input: 106.79 toks/s, output: 3519.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:58<00:32,  1.76it/s, est. speed input: 106.82 toks/s, output: 3573.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:22:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5816.1 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:00<00:47,  1.18it/s, est. speed input: 105.03 toks/s, output: 3568.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:00<00:33,  1.59it/s, est. speed input: 105.81 toks/s, output: 3701.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:01<00:37,  1.40it/s, est. speed input: 105.12 toks/s, output: 3729.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:02<00:27,  1.88it/s, est. speed input: 106.15 toks/s, output: 3872.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:03<00:31,  1.61it/s, est. speed input: 105.56 toks/s, output: 3904.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:04<00:42,  1.16it/s, est. speed input: 103.89 toks/s, output: 3895.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:05<00:33,  1.45it/s, est. speed input: 104.52 toks/s, output: 3971.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:05<00:30,  1.52it/s, est. speed input: 104.56 toks/s, output: 4025.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:09<01:04,  1.39s/it, est. speed input: 100.44 toks/s, output: 3921.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:09<00:49,  1.10s/it, est. speed input: 100.77 toks/s, output: 3989.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:22:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4837.6 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:11<01:03,  1.44s/it, est. speed input: 98.45 toks/s, output: 3952.84 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:12<00:49,  1.14s/it, est. speed input: 98.71 toks/s, output: 4019.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:12<00:36,  1.14it/s, est. speed input: 99.23 toks/s, output: 4095.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:12<00:26,  1.55it/s, est. speed input: 99.90 toks/s, output: 4179.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:12<00:18,  2.14it/s, est. speed input: 100.84 toks/s, output: 4328.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:14<00:24,  1.58it/s, est. speed input: 100.12 toks/s, output: 4352.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:15<00:34,  1.06it/s, est. speed input: 98.52 toks/s, output: 4338.87 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:16<00:32,  1.12it/s, est. speed input: 98.33 toks/s, output: 4386.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:17<00:32,  1.09it/s, est. speed input: 97.88 toks/s, output: 4422.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:18<00:29,  1.16it/s, est. speed input: 97.74 toks/s, output: 4472.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:18<00:23,  1.42it/s, est. speed input: 98.08 toks/s, output: 4544.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:19<00:23,  1.36it/s, est. speed input: 97.87 toks/s, output: 4590.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:19<00:17,  1.74it/s, est. speed input: 98.39 toks/s, output: 4670.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:19<00:14,  2.02it/s, est. speed input: 98.79 toks/s, output: 4744.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:23:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4178.3 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:20<00:13,  2.12it/s, est. speed input: 99.05 toks/s, output: 4811.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:24<00:40,  1.43s/it, est. speed input: 95.44 toks/s, output: 4693.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:25<00:39,  1.48s/it, est. speed input: 94.39 toks/s, output: 4700.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:25<00:29,  1.12s/it, est. speed input: 94.82 toks/s, output: 4779.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:26<00:26,  1.08s/it, est. speed input: 94.41 toks/s, output: 4818.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  1.08s/it, est. speed input: 111.88 toks/s, output: 7174.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  1.84it/s, est. speed input: 111.88 toks/s, output: 7174.22 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:23:22 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:23:22 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:23:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:23:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:23:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:23:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:23:22 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:23:22 [abstract.py:306] It took 0.609866 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:23:22 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:23:22 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:23:22 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0104
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3898.6001
+  • Training FLOPS: 54.28 TFLOPS (13.57 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 410.67s
+  • policy_training: 262.47s (63.9%)
+  • generation: 93.34s (22.7%)
+  • teacher_logprob_inference: 50.00s (12.2%)
+  • policy_training/sharding_data: 4.34s (1.1%)
+  • prepare_for_generation: 3.43s (0.8%)
+  • policy_training/submit_training_futures: 1.97s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.49s (0.4%)
+  • training_prep: 0.81s (0.2%)
+  • teacher_logprob_inference_prep: 0.43s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 29: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:28:37 [abstract.py:324] It took 0.051529 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05991523e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56034698e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66644897e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05991523e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56034698e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66644897e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05991523e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56034698e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66644897e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.05991523e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56034698e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66644897e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05996621e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54805374e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67731781e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25171375e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.41035652e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #29] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43012451e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05996621e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54805374e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67731781e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67814178e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28301010e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #29] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.18354797e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05996621e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54805374e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67731781e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27503052e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.89802361e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #29] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16051178e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.05996621e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54805374e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67731781e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31511650e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75597801e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #29] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32453064e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 29: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:28:40 [abstract.py:324] It took 0.040378 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12122.05it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:28:40 [loggers.py:236] Engine 000: Avg prompt throughput: 29.4 tokens/s, Avg generation throughput: 73.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:05,  2.43it/s, est. speed input: 153.18 toks/s, output: 31.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:09, 15.63it/s, est. speed input: 766.36 toks/s, output: 228.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 28.05it/s, est. speed input: 1266.59 toks/s, output: 518.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:07, 18.76it/s, est. speed input: 1070.52 toks/s, output: 595.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:07, 17.26it/s, est. speed input: 1031.34 toks/s, output: 719.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:07, 18.19it/s, est. speed input: 1060.09 toks/s, output: 872.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:11, 11.98it/s, est. speed input: 888.42 toks/s, output: 865.98 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:02<00:13,  9.92it/s, est. speed input: 809.06 toks/s, output: 904.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:03<00:25,  5.08it/s, est. speed input: 589.70 toks/s, output: 780.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:05<00:45,  2.79it/s, est. speed input: 412.53 toks/s, output: 667.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:09<02:09,  1.04s/it, est. speed input: 214.29 toks/s, output: 429.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:28:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12416.8 tokens/s, Running: 125 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:11<02:17,  1.11s/it, est. speed input: 192.33 toks/s, output: 468.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:11<01:59,  1.03it/s, est. speed input: 190.69 toks/s, output: 544.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:14<02:21,  1.17s/it, est. speed input: 160.12 toks/s, output: 599.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:16<02:24,  1.20s/it, est. speed input: 150.57 toks/s, output: 641.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:17<02:24,  1.21s/it, est. speed input: 143.35 toks/s, output: 685.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:18<02:12,  1.12s/it, est. speed input: 139.92 toks/s, output: 744.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:18<01:58,  1.01s/it, est. speed input: 138.04 toks/s, output: 807.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:19<01:34,  1.23it/s, est. speed input: 139.17 toks/s, output: 886.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:19<01:25,  1.35it/s, est. speed input: 138.24 toks/s, output: 950.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:29:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10335.2 tokens/s, Running: 115 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:21<01:41,  1.13it/s, est. speed input: 132.96 toks/s, output: 983.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:22<01:42,  1.10it/s, est. speed input: 129.99 toks/s, output: 1030.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:24<02:26,  1.30s/it, est. speed input: 120.33 toks/s, output: 1022.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:24<01:57,  1.06s/it, est. speed input: 120.54 toks/s, output: 1091.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:26<02:15,  1.23s/it, est. speed input: 115.37 toks/s, output: 1111.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:26<01:53,  1.04s/it, est. speed input: 115.07 toks/s, output: 1175.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:27<01:29,  1.20it/s, est. speed input: 115.91 toks/s, output: 1248.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:27<01:11,  1.49it/s, est. speed input: 116.94 toks/s, output: 1323.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:28<01:11,  1.49it/s, est. speed input: 116.27 toks/s, output: 1378.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:29:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9143.9 tokens/s, Running: 106 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:30<01:55,  1.10s/it, est. speed input: 110.34 toks/s, output: 1371.20 toks/s]
Processed prompts:  35%|███▌      | 56/160 [00:30<01:25,  1.21it/s, est. speed input: 111.71 toks/s, output: 1450.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:30<01:10,  1.47it/s, est. speed input: 112.46 toks/s, output: 1520.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:31<01:12,  1.41it/s, est. speed input: 111.64 toks/s, output: 1570.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:32<01:15,  1.34it/s, est. speed input: 110.65 toks/s, output: 1616.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:32<01:06,  1.50it/s, est. speed input: 110.88 toks/s, output: 1679.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:33<00:51,  1.89it/s, est. speed input: 111.99 toks/s, output: 1813.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:34<01:00,  1.60it/s, est. speed input: 110.74 toks/s, output: 1852.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:35<01:07,  1.43it/s, est. speed input: 109.57 toks/s, output: 1890.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:35<00:57,  1.66it/s, est. speed input: 110.26 toks/s, output: 1958.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:36<01:01,  1.53it/s, est. speed input: 109.54 toks/s, output: 2003.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:36<00:46,  1.98it/s, est. speed input: 110.74 toks/s, output: 2081.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:37<00:38,  2.38it/s, est. speed input: 112.05 toks/s, output: 2217.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:37<00:33,  2.70it/s, est. speed input: 112.98 toks/s, output: 2290.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:38<00:38,  2.34it/s, est. speed input: 112.82 toks/s, output: 2341.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:38<00:42,  2.08it/s, est. speed input: 112.52 toks/s, output: 2388.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:29:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7630.5 tokens/s, Running: 88 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:41<01:25,  1.02it/s, est. speed input: 107.80 toks/s, output: 2342.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:41<01:04,  1.34it/s, est. speed input: 108.87 toks/s, output: 2418.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:42<01:01,  1.38it/s, est. speed input: 108.54 toks/s, output: 2464.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:42<00:56,  1.48it/s, est. speed input: 108.53 toks/s, output: 2517.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:43<00:40,  2.03it/s, est. speed input: 110.06 toks/s, output: 2655.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:43<00:35,  2.23it/s, est. speed input: 110.95 toks/s, output: 2778.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:45<00:50,  1.57it/s, est. speed input: 109.13 toks/s, output: 2784.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:45<00:35,  2.17it/s, est. speed input: 110.90 toks/s, output: 2932.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:45<00:33,  2.24it/s, est. speed input: 111.30 toks/s, output: 2992.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:46<00:23,  3.15it/s, est. speed input: 113.39 toks/s, output: 3148.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:46<00:21,  3.43it/s, est. speed input: 114.23 toks/s, output: 3220.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:47<00:30,  2.34it/s, est. speed input: 113.41 toks/s, output: 3246.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:48<00:48,  1.47it/s, est. speed input: 111.34 toks/s, output: 3236.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:29:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6488.6 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:50<01:16,  1.09s/it, est. speed input: 107.62 toks/s, output: 3177.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:51<01:08,  1.00it/s, est. speed input: 107.21 toks/s, output: 3216.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:52<01:08,  1.00s/it, est. speed input: 106.32 toks/s, output: 3240.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:53<01:04,  1.05it/s, est. speed input: 105.80 toks/s, output: 3274.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:54<01:00,  1.09it/s, est. speed input: 105.32 toks/s, output: 3308.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:54<00:44,  1.45it/s, est. speed input: 106.17 toks/s, output: 3385.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:55<00:44,  1.42it/s, est. speed input: 105.63 toks/s, output: 3466.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:56<00:40,  1.55it/s, est. speed input: 105.85 toks/s, output: 3522.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:56<00:38,  1.58it/s, est. speed input: 105.78 toks/s, output: 3571.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:57<00:34,  1.74it/s, est. speed input: 106.04 toks/s, output: 3630.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:57<00:30,  1.94it/s, est. speed input: 106.46 toks/s, output: 3693.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:58<00:32,  1.76it/s, est. speed input: 106.22 toks/s, output: 3735.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:58<00:27,  2.10it/s, est. speed input: 106.79 toks/s, output: 3804.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:00<00:41,  1.34it/s, est. speed input: 105.36 toks/s, output: 3802.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:29:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5866.5 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:00<00:39,  1.40it/s, est. speed input: 105.29 toks/s, output: 3848.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:01<00:31,  1.71it/s, est. speed input: 105.82 toks/s, output: 3917.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:01<00:25,  2.05it/s, est. speed input: 106.38 toks/s, output: 3987.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:01<00:24,  2.16it/s, est. speed input: 106.68 toks/s, output: 4047.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:02<00:34,  1.49it/s, est. speed input: 105.67 toks/s, output: 4059.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:03<00:26,  1.88it/s, est. speed input: 106.25 toks/s, output: 4132.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:04<00:36,  1.33it/s, est. speed input: 105.08 toks/s, output: 4138.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:04<00:27,  1.77it/s, est. speed input: 105.82 toks/s, output: 4217.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:05<00:35,  1.32it/s, est. speed input: 104.82 toks/s, output: 4227.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:07<00:43,  1.06it/s, est. speed input: 103.56 toks/s, output: 4227.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:07<00:34,  1.29it/s, est. speed input: 103.87 toks/s, output: 4292.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:08<00:35,  1.24it/s, est. speed input: 103.41 toks/s, output: 4325.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:09<00:36,  1.18it/s, est. speed input: 102.88 toks/s, output: 4353.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:29:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4801.2 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:11<00:40,  1.02it/s, est. speed input: 101.32 toks/s, output: 4389.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:12<00:40,  1.01s/it, est. speed input: 100.62 toks/s, output: 4410.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:12<00:31,  1.23it/s, est. speed input: 101.12 toks/s, output: 4483.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:12<00:23,  1.61it/s, est. speed input: 101.83 toks/s, output: 4565.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:13<00:19,  1.91it/s, est. speed input: 102.29 toks/s, output: 4637.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:13<00:18,  2.00it/s, est. speed input: 102.49 toks/s, output: 4698.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:13<00:11,  3.03it/s, est. speed input: 103.76 toks/s, output: 4861.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:14<00:09,  3.36it/s, est. speed input: 104.27 toks/s, output: 4937.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:14<00:08,  3.96it/s, est. speed input: 104.92 toks/s, output: 5018.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:14<00:07,  3.95it/s, est. speed input: 105.38 toks/s, output: 5090.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:15<00:10,  2.77it/s, est. speed input: 105.30 toks/s, output: 5136.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:15<00:09,  3.11it/s, est. speed input: 105.81 toks/s, output: 5211.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:16<00:12,  2.23it/s, est. speed input: 105.54 toks/s, output: 5249.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:16<00:08,  3.21it/s, est. speed input: 106.73 toks/s, output: 5409.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:16<00:08,  2.83it/s, est. speed input: 106.84 toks/s, output: 5466.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:17<00:11,  2.12it/s, est. speed input: 106.51 toks/s, output: 5500.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:18<00:15,  1.51it/s, est. speed input: 105.68 toks/s, output: 5509.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:19<00:12,  1.76it/s, est. speed input: 105.98 toks/s, output: 5579.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:30:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3805.7 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:21<00:17,  1.17it/s, est. speed input: 104.34 toks/s, output: 5596.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:22<00:14,  1.29it/s, est. speed input: 104.40 toks/s, output: 5653.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:23<00:14,  1.25it/s, est. speed input: 104.01 toks/s, output: 5689.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:24<00:14,  1.14it/s, est. speed input: 103.35 toks/s, output: 5710.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:24<00:09,  1.66it/s, est. speed input: 104.24 toks/s, output: 5867.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:24<00:06,  2.02it/s, est. speed input: 104.78 toks/s, output: 5953.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:25<00:06,  1.98it/s, est. speed input: 104.85 toks/s, output: 6012.27 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:25<00:00,  1.98it/s, est. speed input: 114.08 toks/s, output: 7261.10 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:25<00:00,  1.88it/s, est. speed input: 114.08 toks/s, output: 7261.10 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:30:11 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:30:11 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:30:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:30:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:30:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:30:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:30:11 [abstract.py:306] It took 0.578362 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:30:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:30:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:30:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:30:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3869.9126
+  • Training FLOPS: 52.54 TFLOPS (13.14 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 414.57s
+  • policy_training: 267.01s (64.4%)
+  • generation: 91.60s (22.1%)
+  • teacher_logprob_inference: 51.08s (12.3%)
+  • policy_training/sharding_data: 4.35s (1.0%)
+  • prepare_for_generation: 3.44s (0.8%)
+  • policy_training/submit_training_futures: 2.28s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.48s (0.4%)
+  • training_prep: 0.81s (0.2%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 30: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:35:32 [abstract.py:324] It took 0.056586 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06025430e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56677094e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66535950e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06025430e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56677094e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66535950e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06025430e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56677094e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66535950e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06025430e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56677094e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66535950e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06019727e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51688957e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66938934e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.24571114e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.39239407e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #30] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42881744e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06019727e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51688957e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66938934e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66086884e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29205742e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #30] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.16225891e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06019727e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51688957e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66938934e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27899017e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.54390049e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #30] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14783325e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06019727e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51688957e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66938934e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31409378e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75327168e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #30] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32134033e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 30: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:35:35 [abstract.py:324] It took 0.044209 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12621.57it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:35:35 [loggers.py:236] Engine 000: Avg prompt throughput: 28.7 tokens/s, Avg generation throughput: 41.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:28,  1.80it/s, est. speed input: 109.96 toks/s, output: 25.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:13, 11.22it/s, est. speed input: 533.75 toks/s, output: 180.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:10, 13.86it/s, est. speed input: 654.90 toks/s, output: 277.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:09, 15.81it/s, est. speed input: 743.74 toks/s, output: 389.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:01<00:09, 15.25it/s, est. speed input: 766.76 toks/s, output: 492.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:09, 15.08it/s, est. speed input: 780.48 toks/s, output: 562.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:08, 16.00it/s, est. speed input: 810.74 toks/s, output: 651.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:03<00:39,  3.52it/s, est. speed input: 402.55 toks/s, output: 444.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:04<00:56,  2.45it/s, est. speed input: 301.52 toks/s, output: 450.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:04<00:50,  2.68it/s, est. speed input: 302.26 toks/s, output: 518.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:07<02:03,  1.10it/s, est. speed input: 189.90 toks/s, output: 400.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:08<02:01,  1.10it/s, est. speed input: 177.51 toks/s, output: 447.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:09<01:41,  1.32it/s, est. speed input: 178.48 toks/s, output: 520.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:10<01:53,  1.17it/s, est. speed input: 164.51 toks/s, output: 550.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:35:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12557.4 tokens/s, Running: 132 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:12<02:39,  1.22s/it, est. speed input: 140.06 toks/s, output: 539.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:12<02:01,  1.07it/s, est. speed input: 142.60 toks/s, output: 618.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:13<02:00,  1.07it/s, est. speed input: 137.35 toks/s, output: 663.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:14<01:41,  1.26it/s, est. speed input: 137.26 toks/s, output: 728.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:18<03:48,  1.80s/it, est. speed input: 108.80 toks/s, output: 645.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:18<02:45,  1.31s/it, est. speed input: 111.32 toks/s, output: 726.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:18<02:13,  1.06s/it, est. speed input: 111.73 toks/s, output: 793.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:35:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10668.3 tokens/s, Running: 125 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:22<03:29,  1.69s/it, est. speed input: 98.48 toks/s, output: 764.32 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:22<02:30,  1.22s/it, est. speed input: 100.57 toks/s, output: 844.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:22<02:09,  1.06s/it, est. speed input: 100.23 toks/s, output: 903.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:26<02:36,  1.30s/it, est. speed input: 92.75 toks/s, output: 951.86 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:26<02:04,  1.05s/it, est. speed input: 94.09 toks/s, output: 1025.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:27<02:10,  1.10s/it, est. speed input: 92.02 toks/s, output: 1062.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:28<01:54,  1.02it/s, est. speed input: 92.01 toks/s, output: 1121.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:29<02:10,  1.13s/it, est. speed input: 89.41 toks/s, output: 1147.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:36:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9466.6 tokens/s, Running: 116 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:30<02:05,  1.09s/it, est. speed input: 88.47 toks/s, output: 1193.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:33<02:44,  1.44s/it, est. speed input: 84.17 toks/s, output: 1192.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:35<03:05,  1.64s/it, est. speed input: 80.77 toks/s, output: 1202.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:35<02:25,  1.30s/it, est. speed input: 81.32 toks/s, output: 1268.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:35<01:51,  1.00s/it, est. speed input: 82.37 toks/s, output: 1339.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:36<01:32,  1.19it/s, est. speed input: 82.96 toks/s, output: 1404.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:37<01:50,  1.01s/it, est. speed input: 81.48 toks/s, output: 1433.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:38<01:32,  1.17it/s, est. speed input: 82.01 toks/s, output: 1496.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:40<02:02,  1.15s/it, est. speed input: 79.76 toks/s, output: 1509.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:36:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8286.1 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:41<01:58,  1.12s/it, est. speed input: 79.24 toks/s, output: 1552.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:41<01:31,  1.15it/s, est. speed input: 80.08 toks/s, output: 1621.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:41<01:11,  1.46it/s, est. speed input: 81.01 toks/s, output: 1692.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:42<01:04,  1.60it/s, est. speed input: 81.49 toks/s, output: 1754.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:42<00:49,  2.05it/s, est. speed input: 82.54 toks/s, output: 1827.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:43<00:58,  1.72it/s, est. speed input: 82.49 toks/s, output: 1874.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:43<01:02,  1.60it/s, est. speed input: 82.53 toks/s, output: 1923.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:44<01:09,  1.43it/s, est. speed input: 82.27 toks/s, output: 1966.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:45<01:03,  1.55it/s, est. speed input: 82.61 toks/s, output: 2024.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:46<00:52,  1.83it/s, est. speed input: 83.67 toks/s, output: 2145.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:46<00:50,  1.89it/s, est. speed input: 84.12 toks/s, output: 2204.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:47<00:44,  2.09it/s, est. speed input: 84.82 toks/s, output: 2268.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:47<00:38,  2.40it/s, est. speed input: 85.63 toks/s, output: 2336.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:48<00:51,  1.78it/s, est. speed input: 85.23 toks/s, output: 2371.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:49<01:01,  1.47it/s, est. speed input: 84.84 toks/s, output: 2405.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:36:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7581.9 tokens/s, Running: 91 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:50<01:19,  1.13it/s, est. speed input: 83.78 toks/s, output: 2419.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:50<01:04,  1.39it/s, est. speed input: 84.45 toks/s, output: 2484.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:51<00:48,  1.80it/s, est. speed input: 85.35 toks/s, output: 2556.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:51<00:36,  2.33it/s, est. speed input: 86.74 toks/s, output: 2687.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:51<00:35,  2.42it/s, est. speed input: 87.26 toks/s, output: 2748.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:52<00:32,  2.59it/s, est. speed input: 87.87 toks/s, output: 2811.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:53<00:54,  1.51it/s, est. speed input: 86.71 toks/s, output: 2817.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:53<00:44,  1.86it/s, est. speed input: 87.45 toks/s, output: 2885.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:54<00:40,  2.01it/s, est. speed input: 87.87 toks/s, output: 2944.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:54<00:36,  2.22it/s, est. speed input: 88.39 toks/s, output: 3005.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:54<00:31,  2.47it/s, est. speed input: 89.01 toks/s, output: 3069.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:55<00:26,  2.94it/s, est. speed input: 89.79 toks/s, output: 3138.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:55<00:35,  2.19it/s, est. speed input: 89.67 toks/s, output: 3177.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:56<00:29,  2.62it/s, est. speed input: 90.39 toks/s, output: 3245.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:56<00:27,  2.65it/s, est. speed input: 91.30 toks/s, output: 3360.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:57<00:30,  2.37it/s, est. speed input: 91.48 toks/s, output: 3407.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:57<00:33,  2.13it/s, est. speed input: 91.52 toks/s, output: 3452.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:58<00:26,  2.61it/s, est. speed input: 92.77 toks/s, output: 3579.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:58<00:27,  2.47it/s, est. speed input: 93.00 toks/s, output: 3630.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:00<00:44,  1.51it/s, est. speed input: 91.78 toks/s, output: 3624.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:36:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6092.3 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:01<00:50,  1.34it/s, est. speed input: 91.24 toks/s, output: 3645.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:01<00:39,  1.68it/s, est. speed input: 91.93 toks/s, output: 3713.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:01<00:23,  2.71it/s, est. speed input: 93.69 toks/s, output: 3864.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:02<00:27,  2.30it/s, est. speed input: 93.70 toks/s, output: 3904.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:03<00:37,  1.64it/s, est. speed input: 92.96 toks/s, output: 3915.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:05<01:05,  1.07s/it, est. speed input: 90.56 toks/s, output: 3856.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:06<00:47,  1.25it/s, est. speed input: 91.13 toks/s, output: 3965.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:06<00:37,  1.55it/s, est. speed input: 91.91 toks/s, output: 4039.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:07<00:37,  1.51it/s, est. speed input: 91.80 toks/s, output: 4077.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:09<00:59,  1.07s/it, est. speed input: 89.77 toks/s, output: 4030.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:36:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5386.0 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:10<00:57,  1.05s/it, est. speed input: 89.34 toks/s, output: 4054.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:10<00:42,  1.27it/s, est. speed input: 90.02 toks/s, output: 4128.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:11<00:23,  2.17it/s, est. speed input: 91.60 toks/s, output: 4284.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:11<00:16,  3.01it/s, est. speed input: 92.94 toks/s, output: 4431.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:11<00:15,  3.07it/s, est. speed input: 93.37 toks/s, output: 4493.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:12<00:17,  2.74it/s, est. speed input: 93.55 toks/s, output: 4544.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:12<00:17,  2.73it/s, est. speed input: 93.92 toks/s, output: 4602.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:12<00:18,  2.50it/s, est. speed input: 94.12 toks/s, output: 4653.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:14<00:35,  1.28it/s, est. speed input: 92.69 toks/s, output: 4625.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:14<00:26,  1.64it/s, est. speed input: 93.27 toks/s, output: 4696.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:15<00:29,  1.48it/s, est. speed input: 93.05 toks/s, output: 4727.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:16<00:30,  1.38it/s, est. speed input: 92.80 toks/s, output: 4758.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:17<00:33,  1.23it/s, est. speed input: 92.33 toks/s, output: 4778.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:18<00:38,  1.05it/s, est. speed input: 91.56 toks/s, output: 4783.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:19<00:29,  1.32it/s, est. speed input: 91.97 toks/s, output: 4849.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:20<00:30,  1.23it/s, est. speed input: 91.63 toks/s, output: 4876.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:36:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4674.1 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:20<00:25,  1.47it/s, est. speed input: 92.01 toks/s, output: 4938.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:20<00:15,  2.25it/s, est. speed input: 93.12 toks/s, output: 5085.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:21<00:14,  2.41it/s, est. speed input: 93.51 toks/s, output: 5149.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:21<00:14,  2.25it/s, est. speed input: 93.67 toks/s, output: 5201.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:23<00:26,  1.19it/s, est. speed input: 92.26 toks/s, output: 5167.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:24<00:21,  1.42it/s, est. speed input: 92.52 toks/s, output: 5273.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:26<00:21,  1.29it/s, est. speed input: 91.98 toks/s, output: 5336.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:27<00:21,  1.26it/s, est. speed input: 91.73 toks/s, output: 5369.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:27<00:17,  1.53it/s, est. speed input: 92.20 toks/s, output: 5445.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:27<00:10,  2.20it/s, est. speed input: 93.27 toks/s, output: 5599.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:28<00:10,  2.27it/s, est. speed input: 93.55 toks/s, output: 5662.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:28<00:11,  1.95it/s, est. speed input: 93.46 toks/s, output: 5703.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:29<00:11,  1.85it/s, est. speed input: 93.47 toks/s, output: 5751.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:37:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3590.4 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:30<00:14,  1.35it/s, est. speed input: 92.82 toks/s, output: 5759.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:31<00:13,  1.38it/s, est. speed input: 92.78 toks/s, output: 5806.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:31<00:00,  1.38it/s, est. speed input: 105.16 toks/s, output: 7507.23 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:31<00:00,  1.75it/s, est. speed input: 105.16 toks/s, output: 7507.23 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:37:12 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:37:12 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:37:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:37:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:37:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:37:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:37:13 [abstract.py:306] It took 0.629766 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:37:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:37:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:37:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:37:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.0063
+  • Mean Generation Length: 4293.8564
+  • Training FLOPS: 55.51 TFLOPS (13.88 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 438.94s
+  • policy_training: 280.97s (64.0%)
+  • generation: 98.32s (22.4%)
+  • teacher_logprob_inference: 54.67s (12.5%)
+  • policy_training/sharding_data: 4.47s (1.0%)
+  • prepare_for_generation: 3.55s (0.8%)
+  • policy_training/submit_training_futures: 1.90s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.4%)
+  • training_prep: 0.81s (0.2%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 7/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 31: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:42:51 [abstract.py:324] It took 0.052330 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06064883e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55739098e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66130447e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06064883e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55739098e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66130447e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06064883e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55739098e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66130447e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06064883e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55739098e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66130447e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06062148e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54573441e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65519943e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25815983e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.47452211e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #31] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42911133e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06062148e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54573441e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65519943e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70283203e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29601812e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #31] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.13438263e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06062148e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54573441e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65519943e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29609146e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=3.92269373e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #31] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15541382e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06062148e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54573441e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65519943e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30988235e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.75181694e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #31] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32001251e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 31: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:42:54 [abstract.py:324] It took 0.043654 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12427.11it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:42:54 [loggers.py:236] Engine 000: Avg prompt throughput: 27.7 tokens/s, Avg generation throughput: 9.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:47,  3.32it/s, est. speed input: 205.91 toks/s, output: 36.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:08, 17.88it/s, est. speed input: 894.69 toks/s, output: 246.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 24.33it/s, est. speed input: 1197.48 toks/s, output: 412.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 26.99it/s, est. speed input: 1343.83 toks/s, output: 560.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 27.30it/s, est. speed input: 1416.48 toks/s, output: 708.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:10, 13.24it/s, est. speed input: 979.32 toks/s, output: 634.48 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:05<01:01,  2.20it/s, est. speed input: 263.90 toks/s, output: 360.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:07<01:08,  1.95it/s, est. speed input: 226.17 toks/s, output: 450.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:08<01:12,  1.80it/s, est. speed input: 203.32 toks/s, output: 556.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:09<01:12,  1.79it/s, est. speed input: 196.84 toks/s, output: 611.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:43:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12476.1 tokens/s, Running: 130 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:10<01:21,  1.58it/s, est. speed input: 183.48 toks/s, output: 641.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:10<01:22,  1.56it/s, est. speed input: 177.29 toks/s, output: 691.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:11<01:13,  1.74it/s, est. speed input: 177.29 toks/s, output: 760.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:11<01:06,  1.90it/s, est. speed input: 176.79 toks/s, output: 826.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:12<01:10,  1.78it/s, est. speed input: 172.09 toks/s, output: 870.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:12<01:00,  2.04it/s, est. speed input: 172.81 toks/s, output: 940.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:13<01:06,  1.86it/s, est. speed input: 168.89 toks/s, output: 982.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:14<01:20,  1.52it/s, est. speed input: 161.80 toks/s, output: 1005.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:15<01:35,  1.27it/s, est. speed input: 153.97 toks/s, output: 1021.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:17<01:51,  1.07it/s, est. speed input: 141.60 toks/s, output: 1059.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:18<01:42,  1.15it/s, est. speed input: 139.91 toks/s, output: 1109.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:18<01:34,  1.24it/s, est. speed input: 138.35 toks/s, output: 1160.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:43:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10550.8 tokens/s, Running: 117 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:21<02:20,  1.21s/it, est. speed input: 126.04 toks/s, output: 1121.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:21<01:49,  1.05it/s, est. speed input: 127.16 toks/s, output: 1194.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:22<01:57,  1.03s/it, est. speed input: 122.95 toks/s, output: 1217.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:22<01:28,  1.28it/s, est. speed input: 124.68 toks/s, output: 1295.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:23<01:24,  1.32it/s, est. speed input: 123.43 toks/s, output: 1344.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:24<01:15,  1.48it/s, est. speed input: 123.36 toks/s, output: 1404.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:24<00:55,  1.97it/s, est. speed input: 125.30 toks/s, output: 1485.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:24<00:43,  2.53it/s, est. speed input: 127.10 toks/s, output: 1565.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:24<00:48,  2.23it/s, est. speed input: 126.58 toks/s, output: 1616.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:25<01:00,  1.78it/s, est. speed input: 124.94 toks/s, output: 1651.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:27<01:26,  1.23it/s, est. speed input: 120.77 toks/s, output: 1651.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:27<01:25,  1.23it/s, est. speed input: 119.51 toks/s, output: 1690.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:28<01:06,  1.55it/s, est. speed input: 120.52 toks/s, output: 1762.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:28<00:45,  2.24it/s, est. speed input: 122.88 toks/s, output: 1907.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:29<00:50,  1.98it/s, est. speed input: 122.02 toks/s, output: 1949.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:43:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9007.2 tokens/s, Running: 100 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.0%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:30<00:50,  1.97it/s, est. speed input: 122.02 toks/s, output: 2053.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:30<00:42,  2.33it/s, est. speed input: 123.27 toks/s, output: 2129.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:31<01:01,  1.57it/s, est. speed input: 120.26 toks/s, output: 2130.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:32<00:55,  1.74it/s, est. speed input: 120.64 toks/s, output: 2190.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:33<01:05,  1.43it/s, est. speed input: 118.00 toks/s, output: 2245.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:33<00:41,  2.20it/s, est. speed input: 121.20 toks/s, output: 2410.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:34<00:45,  2.00it/s, est. speed input: 120.58 toks/s, output: 2450.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:35<00:43,  2.06it/s, est. speed input: 120.77 toks/s, output: 2505.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:35<00:39,  2.26it/s, est. speed input: 121.39 toks/s, output: 2568.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:35<00:32,  2.67it/s, est. speed input: 122.45 toks/s, output: 2641.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:35<00:27,  3.22it/s, est. speed input: 123.68 toks/s, output: 2716.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:37<00:59,  1.45it/s, est. speed input: 119.77 toks/s, output: 2682.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:37<00:34,  2.42it/s, est. speed input: 122.56 toks/s, output: 2843.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:37<00:30,  2.70it/s, est. speed input: 123.45 toks/s, output: 2911.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:37<00:28,  2.89it/s, est. speed input: 124.14 toks/s, output: 2976.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:38<00:23,  3.49it/s, est. speed input: 125.28 toks/s, output: 3053.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:38<00:34,  2.31it/s, est. speed input: 124.21 toks/s, output: 3075.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:39<00:45,  1.72it/s, est. speed input: 122.82 toks/s, output: 3088.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:43:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7606.9 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:40<00:50,  1.53it/s, est. speed input: 121.76 toks/s, output: 3111.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:41<00:50,  1.52it/s, est. speed input: 121.25 toks/s, output: 3146.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:42<00:49,  1.53it/s, est. speed input: 120.84 toks/s, output: 3184.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:43<01:01,  1.22it/s, est. speed input: 118.90 toks/s, output: 3180.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:43<00:53,  1.40it/s, est. speed input: 118.98 toks/s, output: 3232.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:44<00:50,  1.44it/s, est. speed input: 118.62 toks/s, output: 3270.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:44<00:37,  1.90it/s, est. speed input: 119.63 toks/s, output: 3347.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:44<00:30,  2.31it/s, est. speed input: 120.40 toks/s, output: 3417.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:45<00:31,  2.24it/s, est. speed input: 120.47 toks/s, output: 3466.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:45<00:25,  2.66it/s, est. speed input: 121.27 toks/s, output: 3536.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:45<00:22,  3.00it/s, est. speed input: 121.99 toks/s, output: 3604.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:45<00:14,  4.42it/s, est. speed input: 123.99 toks/s, output: 3760.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:46<00:15,  4.11it/s, est. speed input: 124.54 toks/s, output: 3822.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:46<00:22,  2.85it/s, est. speed input: 124.10 toks/s, output: 3855.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:47<00:18,  3.43it/s, est. speed input: 125.54 toks/s, output: 3992.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:48<00:28,  2.15it/s, est. speed input: 124.07 toks/s, output: 3993.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:48<00:25,  2.34it/s, est. speed input: 124.50 toks/s, output: 4054.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:49<00:27,  2.17it/s, est. speed input: 124.32 toks/s, output: 4095.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:49<00:13,  4.15it/s, est. speed input: 127.67 toks/s, output: 4344.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:49<00:18,  2.98it/s, est. speed input: 127.10 toks/s, output: 4370.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:43:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6274.7 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:50<00:24,  2.16it/s, est. speed input: 126.01 toks/s, output: 4379.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:51<00:30,  1.77it/s, est. speed input: 125.00 toks/s, output: 4392.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:52<00:34,  1.53it/s, est. speed input: 124.00 toks/s, output: 4404.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:52<00:27,  1.87it/s, est. speed input: 124.64 toks/s, output: 4476.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:53<00:33,  1.48it/s, est. speed input: 123.30 toks/s, output: 4477.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:54<00:29,  1.66it/s, est. speed input: 123.48 toks/s, output: 4532.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:54<00:18,  2.54it/s, est. speed input: 125.17 toks/s, output: 4687.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:54<00:11,  3.83it/s, est. speed input: 127.08 toks/s, output: 4856.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:54<00:11,  3.72it/s, est. speed input: 127.46 toks/s, output: 4918.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:55<00:11,  3.66it/s, est. speed input: 127.89 toks/s, output: 4982.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:55<00:11,  3.68it/s, est. speed input: 128.35 toks/s, output: 5048.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:55<00:10,  3.94it/s, est. speed input: 128.98 toks/s, output: 5119.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:55<00:06,  5.76it/s, est. speed input: 130.81 toks/s, output: 5285.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:58<00:28,  1.35it/s, est. speed input: 126.09 toks/s, output: 5142.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:58<00:23,  1.56it/s, est. speed input: 126.39 toks/s, output: 5202.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:59<00:21,  1.70it/s, est. speed input: 126.46 toks/s, output: 5254.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:59<00:16,  2.15it/s, est. speed input: 127.17 toks/s, output: 5331.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:59<00:11,  2.79it/s, est. speed input: 128.19 toks/s, output: 5472.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:43:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4835.0 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:00<00:13,  2.32it/s, est. speed input: 127.80 toks/s, output: 5503.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:01<00:14,  2.11it/s, est. speed input: 127.49 toks/s, output: 5541.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:02<00:18,  1.59it/s, est. speed input: 126.33 toks/s, output: 5539.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:02<00:15,  1.92it/s, est. speed input: 126.82 toks/s, output: 5611.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:02<00:13,  2.03it/s, est. speed input: 126.91 toks/s, output: 5665.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:03<00:11,  2.18it/s, est. speed input: 127.21 toks/s, output: 5775.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:03<00:09,  2.56it/s, est. speed input: 127.77 toks/s, output: 5853.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:03<00:05,  3.89it/s, est. speed input: 129.44 toks/s, output: 6028.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:04<00:03,  5.27it/s, est. speed input: 131.02 toks/s, output: 6201.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:04<00:04,  4.31it/s, est. speed input: 131.19 toks/s, output: 6257.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:04<00:04,  4.16it/s, est. speed input: 131.57 toks/s, output: 6326.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:06<00:10,  1.76it/s, est. speed input: 129.34 toks/s, output: 6271.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:07<00:11,  1.54it/s, est. speed input: 128.54 toks/s, output: 6286.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:08<00:14,  1.13it/s, est. speed input: 126.62 toks/s, output: 6245.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:08<00:10,  1.45it/s, est. speed input: 127.18 toks/s, output: 6327.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:09<00:06,  2.03it/s, est. speed input: 128.05 toks/s, output: 6478.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:44:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3251.9 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:10<00:05,  2.19it/s, est. speed input: 128.39 toks/s, output: 6602.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:10<00:03,  2.55it/s, est. speed input: 128.97 toks/s, output: 6686.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:10<00:03,  2.77it/s, est. speed input: 129.32 toks/s, output: 6762.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:11<00:03,  2.27it/s, est. speed input: 128.96 toks/s, output: 6799.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:12<00:04,  1.53it/s, est. speed input: 127.57 toks/s, output: 6785.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:13<00:03,  1.51it/s, est. speed input: 127.20 toks/s, output: 6827.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:15<00:05,  1.19s/it, est. speed input: 123.74 toks/s, output: 6706.49 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  1.19s/it, est. speed input: 127.65 toks/s, output: 7247.04 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  2.11it/s, est. speed input: 127.65 toks/s, output: 7247.04 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:44:14 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:44:14 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:44:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:44:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:44:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:44:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:44:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:44:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 00:44:15 [abstract.py:306] It took 0.599079 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:44:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:44:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3431.9875
+  • Training FLOPS: 51.20 TFLOPS (12.80 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 372.73s
+  • policy_training: 238.87s (64.1%)
+  • generation: 81.41s (21.8%)
+  • teacher_logprob_inference: 47.39s (12.7%)
+  • policy_training/sharding_data: 4.35s (1.2%)
+  • prepare_for_generation: 3.24s (0.9%)
+  • policy_training/submit_training_futures: 1.86s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.47s (0.4%)
+  • training_prep: 1.24s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 32: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:04 [abstract.py:324] It took 0.052912 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06136797e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54958992e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66164703e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06136797e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54958992e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66164703e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06136797e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54958992e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66164703e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06136797e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54958992e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66164703e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06126797e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55088882e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64227600e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26000443e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.41717958e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #32] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.42948868e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06126797e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55088882e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64227600e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69475784e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29272919e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #32] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.13467293e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06126797e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55088882e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64227600e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28292999e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.34488058e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #32] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14634018e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06126797e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55088882e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64227600e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28995514e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74354897e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #32] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31902832e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 32: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:07 [abstract.py:324] It took 0.040467 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13427.41it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:07 [loggers.py:236] Engine 000: Avg prompt throughput: 32.3 tokens/s, Avg generation throughput: 28.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:50,  3.15it/s, est. speed input: 186.03 toks/s, output: 47.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:04, 33.45it/s, est. speed input: 1658.91 toks/s, output: 532.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 31.65it/s, est. speed input: 1701.01 toks/s, output: 675.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 28.88it/s, est. speed input: 1653.27 toks/s, output: 871.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:04, 26.60it/s, est. speed input: 1595.33 toks/s, output: 1020.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:08, 15.75it/s, est. speed input: 1224.81 toks/s, output: 987.84 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:09, 12.77it/s, est. speed input: 1078.55 toks/s, output: 1040.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:02<00:15,  7.83it/s, est. speed input: 838.98 toks/s, output: 938.19 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:03<00:23,  5.22it/s, est. speed input: 665.00 toks/s, output: 885.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:05<00:44,  2.72it/s, est. speed input: 471.66 toks/s, output: 710.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:05<00:47,  2.52it/s, est. speed input: 435.38 toks/s, output: 737.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:05<00:36,  3.27it/s, est. speed input: 441.33 toks/s, output: 905.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:06<00:34,  3.42it/s, est. speed input: 434.56 toks/s, output: 969.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:06<00:30,  3.86it/s, est. speed input: 435.19 toks/s, output: 1047.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:06<00:35,  3.21it/s, est. speed input: 412.87 toks/s, output: 1068.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:09<01:35,  1.20it/s, est. speed input: 308.42 toks/s, output: 879.17 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:09<01:26,  1.30it/s, est. speed input: 296.04 toks/s, output: 924.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12051.6 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:10<01:02,  1.78it/s, est. speed input: 291.52 toks/s, output: 1063.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:11<01:10,  1.57it/s, est. speed input: 273.87 toks/s, output: 1077.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:15<02:38,  1.45s/it, est. speed input: 207.19 toks/s, output: 894.25 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:15<02:02,  1.14s/it, est. speed input: 207.77 toks/s, output: 976.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:15<01:40,  1.07it/s, est. speed input: 206.28 toks/s, output: 1046.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:16<01:26,  1.22it/s, est. speed input: 203.62 toks/s, output: 1109.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:18<02:14,  1.28s/it, est. speed input: 180.20 toks/s, output: 1059.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:19<01:58,  1.14s/it, est. speed input: 176.05 toks/s, output: 1110.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:19<01:37,  1.06it/s, est. speed input: 174.94 toks/s, output: 1179.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9921.3 tokens/s, Running: 103 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:21<01:54,  1.13s/it, est. speed input: 165.13 toks/s, output: 1188.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:22<01:27,  1.15it/s, est. speed input: 162.18 toks/s, output: 1312.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:25<02:02,  1.24s/it, est. speed input: 149.21 toks/s, output: 1281.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:26<02:11,  1.34s/it, est. speed input: 142.45 toks/s, output: 1296.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8585.2 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:30<03:06,  1.92s/it, est. speed input: 127.91 toks/s, output: 1237.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:33<03:31,  2.20s/it, est. speed input: 118.43 toks/s, output: 1219.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:33<02:43,  1.72s/it, est. speed input: 118.48 toks/s, output: 1291.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:34<01:38,  1.06s/it, est. speed input: 120.27 toks/s, output: 1452.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:35<01:44,  1.14s/it, est. speed input: 117.37 toks/s, output: 1487.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:36<01:33,  1.03s/it, est. speed input: 116.74 toks/s, output: 1548.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:36<01:21,  1.11it/s, est. speed input: 116.59 toks/s, output: 1615.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:36<01:02,  1.43it/s, est. speed input: 117.73 toks/s, output: 1698.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:37<00:55,  1.59it/s, est. speed input: 117.94 toks/s, output: 1768.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:38<01:15,  1.15it/s, est. speed input: 115.01 toks/s, output: 1790.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:38<00:56,  1.52it/s, est. speed input: 116.15 toks/s, output: 1873.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7621.3 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:40<01:28,  1.04s/it, est. speed input: 112.17 toks/s, output: 1874.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:42<01:36,  1.14s/it, est. speed input: 109.92 toks/s, output: 1901.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:42<01:16,  1.08it/s, est. speed input: 110.39 toks/s, output: 1972.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:43<01:08,  1.20it/s, est. speed input: 110.23 toks/s, output: 2033.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:44<00:51,  1.56it/s, est. speed input: 110.86 toks/s, output: 2171.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:44<00:39,  1.99it/s, est. speed input: 112.15 toks/s, output: 2319.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:46<00:51,  1.48it/s, est. speed input: 110.38 toks/s, output: 2343.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:47<00:59,  1.27it/s, est. speed input: 109.01 toks/s, output: 2376.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:49<01:23,  1.12s/it, est. speed input: 105.59 toks/s, output: 2363.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:49<00:55,  1.31it/s, est. speed input: 106.83 toks/s, output: 2511.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:49:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6482.4 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:50<00:54,  1.31it/s, est. speed input: 106.43 toks/s, output: 2561.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:51<00:51,  1.37it/s, est. speed input: 106.30 toks/s, output: 2618.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:53<01:28,  1.26s/it, est. speed input: 102.02 toks/s, output: 2572.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:54<01:13,  1.07s/it, est. speed input: 102.06 toks/s, output: 2632.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:54<00:54,  1.24it/s, est. speed input: 102.97 toks/s, output: 2714.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:55<00:53,  1.24it/s, est. speed input: 102.56 toks/s, output: 2762.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:56<00:59,  1.11it/s, est. speed input: 101.66 toks/s, output: 2794.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:57<00:58,  1.11it/s, est. speed input: 101.12 toks/s, output: 2838.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:58<00:56,  1.14it/s, est. speed input: 100.76 toks/s, output: 2886.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:50:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5882.9 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:01<01:34,  1.49s/it, est. speed input: 96.89 toks/s, output: 2834.32 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:02<01:24,  1.37s/it, est. speed input: 96.19 toks/s, output: 2873.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:02<00:52,  1.14it/s, est. speed input: 97.24 toks/s, output: 3021.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:04<00:44,  1.30it/s, est. speed input: 97.28 toks/s, output: 3137.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:05<00:45,  1.25it/s, est. speed input: 96.89 toks/s, output: 3181.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:05<00:38,  1.46it/s, est. speed input: 97.32 toks/s, output: 3252.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:06<00:42,  1.31it/s, est. speed input: 96.79 toks/s, output: 3291.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:07<00:32,  1.63it/s, est. speed input: 97.41 toks/s, output: 3426.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:08<00:38,  1.34it/s, est. speed input: 96.62 toks/s, output: 3454.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:08<00:32,  1.56it/s, est. speed input: 97.04 toks/s, output: 3526.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:08<00:25,  1.99it/s, est. speed input: 97.78 toks/s, output: 3608.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:08<00:20,  2.40it/s, est. speed input: 98.40 toks/s, output: 3686.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:09<00:16,  2.92it/s, est. speed input: 99.07 toks/s, output: 3766.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:09<00:18,  2.48it/s, est. speed input: 99.13 toks/s, output: 3824.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:09<00:16,  2.76it/s, est. speed input: 99.64 toks/s, output: 3898.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:50:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5171.6 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:11<00:23,  1.84it/s, est. speed input: 99.27 toks/s, output: 3990.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:11<00:21,  1.98it/s, est. speed input: 99.59 toks/s, output: 4057.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:13<00:25,  1.61it/s, est. speed input: 99.09 toks/s, output: 4146.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:13<00:22,  1.78it/s, est. speed input: 99.43 toks/s, output: 4215.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:14<00:20,  1.83it/s, est. speed input: 99.67 toks/s, output: 4333.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:15<00:24,  1.48it/s, est. speed input: 98.97 toks/s, output: 4358.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:16<00:21,  1.71it/s, est. speed input: 99.38 toks/s, output: 4431.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:16<00:17,  2.00it/s, est. speed input: 99.83 toks/s, output: 4506.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:18<00:26,  1.31it/s, est. speed input: 98.68 toks/s, output: 4510.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:18<00:26,  1.25it/s, est. speed input: 98.35 toks/s, output: 4549.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:19<00:22,  1.42it/s, est. speed input: 98.58 toks/s, output: 4614.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:19<00:13,  2.19it/s, est. speed input: 99.71 toks/s, output: 4778.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:19<00:11,  2.61it/s, est. speed input: 100.31 toks/s, output: 4859.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:50:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4189.7 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:20<00:11,  2.50it/s, est. speed input: 100.52 toks/s, output: 4923.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:21<00:13,  1.99it/s, est. speed input: 100.30 toks/s, output: 4967.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:21<00:14,  1.80it/s, est. speed input: 100.19 toks/s, output: 5016.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:21<00:11,  2.22it/s, est. speed input: 100.72 toks/s, output: 5097.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:22<00:12,  1.98it/s, est. speed input: 100.70 toks/s, output: 5150.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:23<00:12,  1.80it/s, est. speed input: 100.61 toks/s, output: 5200.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:23<00:08,  2.39it/s, est. speed input: 101.47 toks/s, output: 5353.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:23<00:07,  2.80it/s, est. speed input: 101.98 toks/s, output: 5435.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:24<00:09,  2.09it/s, est. speed input: 101.70 toks/s, output: 5475.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:25<00:07,  2.25it/s, est. speed input: 101.98 toks/s, output: 5546.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:26<00:09,  1.70it/s, est. speed input: 101.56 toks/s, output: 5579.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:26<00:05,  2.64it/s, est. speed input: 102.72 toks/s, output: 5752.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:26<00:04,  2.95it/s, est. speed input: 103.17 toks/s, output: 5832.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  2.95it/s, est. speed input: 112.89 toks/s, output: 7151.77 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  1.85it/s, est. speed input: 112.89 toks/s, output: 7151.77 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:50:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:50:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:50:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:50:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:50:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:50:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:50:39 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:50:39 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 00:50:39 [abstract.py:306] It took 0.632002 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:50:39 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:50:39 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0102
+  • Training Accuracy: 0.0000
+  • Mean Generation Length: 3870.6375
+  • Training FLOPS: 53.37 TFLOPS (13.34 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 414.26s
+  • policy_training: 265.20s (64.0%)
+  • generation: 92.88s (22.4%)
+  • teacher_logprob_inference: 51.27s (12.4%)
+  • policy_training/sharding_data: 4.39s (1.1%)
+  • prepare_for_generation: 3.39s (0.8%)
+  • policy_training/submit_training_futures: 1.86s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.4%)
+  • training_prep: 0.87s (0.2%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 33: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 00:55:59 [abstract.py:324] It took 0.052493 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06185117e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55078049e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66081924e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06185117e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55078049e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66081924e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06185117e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55078049e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66081924e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06185117e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55078049e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66081924e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06180215e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54085388e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66187744e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25176392e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.39061356e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #33] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43036682e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06180215e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54085388e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66187744e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66552048e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28335638e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #33] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.10005226e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06180215e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54085388e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66187744e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27591095e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.25899506e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #33] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13416862e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06180215e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54085388e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66187744e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29594193e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.73624268e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #33] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31859055e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 33: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:01 [abstract.py:324] It took 0.040370 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12381.25it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:02 [loggers.py:236] Engine 000: Avg prompt throughput: 29.0 tokens/s, Avg generation throughput: 59.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:23,  1.91it/s, est. speed input: 116.46 toks/s, output: 26.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:09, 16.18it/s, est. speed input: 772.55 toks/s, output: 233.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:07, 19.08it/s, est. speed input: 926.79 toks/s, output: 351.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 25.28it/s, est. speed input: 1144.35 toks/s, output: 546.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:05, 26.20it/s, est. speed input: 1249.16 toks/s, output: 743.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:08, 15.09it/s, est. speed input: 987.06 toks/s, output: 721.67 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:03<00:21,  6.08it/s, est. speed input: 585.93 toks/s, output: 573.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:05<00:44,  2.88it/s, est. speed input: 358.18 toks/s, output: 470.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:05<00:40,  3.09it/s, est. speed input: 350.03 toks/s, output: 609.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:07<00:58,  2.15it/s, est. speed input: 290.64 toks/s, output: 583.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:09<01:22,  1.51it/s, est. speed input: 241.61 toks/s, output: 563.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12334.7 tokens/s, Running: 124 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:11<02:00,  1.02it/s, est. speed input: 196.23 toks/s, output: 536.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:13<02:24,  1.18s/it, est. speed input: 172.29 toks/s, output: 549.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:15<02:46,  1.38s/it, est. speed input: 153.56 toks/s, output: 567.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:17<02:52,  1.44s/it, est. speed input: 142.63 toks/s, output: 603.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:18<02:45,  1.39s/it, est. speed input: 136.12 toks/s, output: 651.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:18<02:15,  1.15s/it, est. speed input: 135.89 toks/s, output: 724.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10380.0 tokens/s, Running: 118 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:21<02:55,  1.50s/it, est. speed input: 123.37 toks/s, output: 730.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:21<02:09,  1.12s/it, est. speed input: 125.32 toks/s, output: 814.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:22<01:58,  1.03s/it, est. speed input: 123.53 toks/s, output: 872.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:24<02:35,  1.37s/it, est. speed input: 114.97 toks/s, output: 882.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:24<01:57,  1.04s/it, est. speed input: 116.22 toks/s, output: 961.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:25<01:38,  1.14it/s, est. speed input: 116.36 toks/s, output: 1030.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:25<01:12,  1.53it/s, est. speed input: 118.13 toks/s, output: 1112.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:27<01:53,  1.03s/it, est. speed input: 111.94 toks/s, output: 1120.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:27<01:29,  1.22it/s, est. speed input: 112.91 toks/s, output: 1194.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:28<01:35,  1.13it/s, est. speed input: 110.88 toks/s, output: 1237.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:29<01:32,  1.15it/s, est. speed input: 109.80 toks/s, output: 1289.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9190.3 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:32<02:58,  1.68s/it, est. speed input: 99.69 toks/s, output: 1235.08 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:34<02:57,  1.69s/it, est. speed input: 96.51 toks/s, output: 1259.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:35<02:23,  1.38s/it, est. speed input: 96.48 toks/s, output: 1321.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:35<01:47,  1.04s/it, est. speed input: 97.54 toks/s, output: 1397.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:36<01:21,  1.24it/s, est. speed input: 98.00 toks/s, output: 1525.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:37<01:21,  1.23it/s, est. speed input: 97.46 toks/s, output: 1576.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:37<01:04,  1.53it/s, est. speed input: 98.62 toks/s, output: 1652.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:40<01:49,  1.12s/it, est. speed input: 94.31 toks/s, output: 1639.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7945.4 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:40<01:43,  1.07s/it, est. speed input: 93.63 toks/s, output: 1685.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:41<01:17,  1.24it/s, est. speed input: 94.78 toks/s, output: 1763.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:42<01:21,  1.17it/s, est. speed input: 94.04 toks/s, output: 1806.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:42<00:47,  1.95it/s, est. speed input: 96.53 toks/s, output: 1965.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:42<00:50,  1.81it/s, est. speed input: 96.46 toks/s, output: 2018.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:43<00:52,  1.75it/s, est. speed input: 96.41 toks/s, output: 2073.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:43<00:32,  2.73it/s, est. speed input: 98.83 toks/s, output: 2232.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:44<00:27,  3.14it/s, est. speed input: 100.50 toks/s, output: 2375.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:44<00:30,  2.81it/s, est. speed input: 100.70 toks/s, output: 2432.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:46<00:49,  1.73it/s, est. speed input: 99.12 toks/s, output: 2447.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:46<00:49,  1.69it/s, est. speed input: 99.08 toks/s, output: 2497.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:47<00:48,  1.71it/s, est. speed input: 99.20 toks/s, output: 2551.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:47<00:45,  1.80it/s, est. speed input: 99.45 toks/s, output: 2609.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:48<00:58,  1.38it/s, est. speed input: 98.36 toks/s, output: 2631.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:48<00:44,  1.82it/s, est. speed input: 99.38 toks/s, output: 2708.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:49<00:43,  1.80it/s, est. speed input: 99.52 toks/s, output: 2761.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:56:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6879.3 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:50<00:50,  1.56it/s, est. speed input: 99.06 toks/s, output: 2797.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:51<01:04,  1.19it/s, est. speed input: 97.69 toks/s, output: 2810.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:52<00:58,  1.30it/s, est. speed input: 97.73 toks/s, output: 2860.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:52<00:42,  1.76it/s, est. speed input: 98.79 toks/s, output: 2990.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:54<00:57,  1.26it/s, est. speed input: 97.22 toks/s, output: 2992.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:55<01:01,  1.17it/s, est. speed input: 96.50 toks/s, output: 3019.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:56<00:57,  1.24it/s, est. speed input: 96.46 toks/s, output: 3066.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:57<01:02,  1.13it/s, est. speed input: 95.68 toks/s, output: 3091.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:58<01:02,  1.10it/s, est. speed input: 95.16 toks/s, output: 3123.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:58<00:46,  1.47it/s, est. speed input: 95.99 toks/s, output: 3200.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:58<00:34,  1.94it/s, est. speed input: 96.81 toks/s, output: 3276.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:59<00:34,  1.91it/s, est. speed input: 96.96 toks/s, output: 3329.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:59<00:26,  2.43it/s, est. speed input: 97.77 toks/s, output: 3404.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:59<00:25,  2.55it/s, est. speed input: 98.22 toks/s, output: 3467.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:57:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6053.0 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:03<01:27,  1.39s/it, est. speed input: 93.36 toks/s, output: 3345.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:03<01:05,  1.05s/it, est. speed input: 93.92 toks/s, output: 3415.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:04<00:57,  1.06it/s, est. speed input: 93.82 toks/s, output: 3462.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:04<00:43,  1.36it/s, est. speed input: 94.40 toks/s, output: 3533.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:04<00:24,  2.33it/s, est. speed input: 96.10 toks/s, output: 3693.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:04<00:22,  2.51it/s, est. speed input: 96.59 toks/s, output: 3759.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:05<00:23,  2.38it/s, est. speed input: 96.81 toks/s, output: 3816.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:05<00:23,  2.39it/s, est. speed input: 97.10 toks/s, output: 3876.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:06<00:22,  2.39it/s, est. speed input: 97.67 toks/s, output: 3994.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:07<00:29,  1.79it/s, est. speed input: 97.10 toks/s, output: 4018.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:09<00:40,  1.26it/s, est. speed input: 95.93 toks/s, output: 4016.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:09<00:31,  1.57it/s, est. speed input: 96.53 toks/s, output: 4088.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:09<00:25,  1.94it/s, est. speed input: 97.12 toks/s, output: 4160.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:09<00:16,  2.83it/s, est. speed input: 98.42 toks/s, output: 4311.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:57:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5223.0 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:10<00:25,  1.81it/s, est. speed input: 97.62 toks/s, output: 4324.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:11<00:18,  2.34it/s, est. speed input: 98.63 toks/s, output: 4463.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:12<00:19,  2.17it/s, est. speed input: 98.69 toks/s, output: 4512.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:12<00:19,  2.20it/s, est. speed input: 98.95 toks/s, output: 4570.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:12<00:18,  2.16it/s, est. speed input: 99.10 toks/s, output: 4625.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:13<00:16,  2.37it/s, est. speed input: 99.74 toks/s, output: 4749.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:14<00:10,  3.31it/s, est. speed input: 101.54 toks/s, output: 4972.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:15<00:14,  2.48it/s, est. speed input: 101.18 toks/s, output: 5002.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:16<00:15,  2.15it/s, est. speed input: 101.21 toks/s, output: 5096.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:16<00:15,  2.12it/s, est. speed input: 101.35 toks/s, output: 5150.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:16<00:12,  2.46it/s, est. speed input: 101.89 toks/s, output: 5225.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:17<00:13,  2.29it/s, est. speed input: 101.95 toks/s, output: 5276.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:18<00:14,  1.89it/s, est. speed input: 101.77 toks/s, output: 5360.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:18<00:09,  2.84it/s, est. speed input: 103.14 toks/s, output: 5529.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:19<00:08,  2.92it/s, est. speed input: 103.49 toks/s, output: 5596.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:19<00:05,  3.93it/s, est. speed input: 104.72 toks/s, output: 5757.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:19<00:07,  3.02it/s, est. speed input: 104.69 toks/s, output: 5801.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 00:57:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4059.3 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:20<00:07,  2.98it/s, est. speed input: 104.98 toks/s, output: 5864.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:21<00:08,  2.27it/s, est. speed input: 104.72 toks/s, output: 5898.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:21<00:08,  2.30it/s, est. speed input: 104.95 toks/s, output: 5958.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:21<00:06,  2.76it/s, est. speed input: 105.49 toks/s, output: 6035.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:22<00:08,  2.12it/s, est. speed input: 105.25 toks/s, output: 6070.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:22<00:07,  2.22it/s, est. speed input: 105.47 toks/s, output: 6132.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:23<00:07,  1.96it/s, est. speed input: 105.35 toks/s, output: 6175.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:25<00:12,  1.16it/s, est. speed input: 103.94 toks/s, output: 6143.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:25<00:09,  1.31it/s, est. speed input: 104.03 toks/s, output: 6198.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:26<00:09,  1.32it/s, est. speed input: 103.85 toks/s, output: 6239.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:26<00:06,  1.76it/s, est. speed input: 104.39 toks/s, output: 6324.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:26<00:04,  2.09it/s, est. speed input: 104.76 toks/s, output: 6399.51 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  2.09it/s, est. speed input: 111.69 toks/s, output: 7342.38 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:26<00:00,  1.84it/s, est. speed input: 111.69 toks/s, output: 7342.38 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:57:34 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 00:57:34 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 00:57:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:57:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:57:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:57:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 00:57:35 [abstract.py:306] It took 0.606537 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 00:57:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 00:57:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 00:57:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 00:57:35 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0109
+  • Training Accuracy: 0.0063
+  • Mean Generation Length: 3986.9688
+  • Training FLOPS: 53.00 TFLOPS (13.25 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 422.55s
+  • policy_training: 272.71s (64.5%)
+  • generation: 93.27s (22.1%)
+  • teacher_logprob_inference: 51.83s (12.3%)
+  • policy_training/sharding_data: 4.33s (1.0%)
+  • prepare_for_generation: 3.37s (0.8%)
+  • policy_training/submit_training_futures: 2.45s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.49s (0.4%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.38s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 34: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:02 [abstract.py:324] It took 0.053698 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06220918e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54776764e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66050949e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06220918e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54776764e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66050949e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06220918e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54776764e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66050949e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06220918e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54776764e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66050949e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06223594e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55281067e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65884857e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26219482e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.34398031e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #34] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43149231e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06223594e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55281067e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65884857e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67161407e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27397690e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #34] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.10463333e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06223594e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55281067e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65884857e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27406845e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.29845333e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #34] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.12855301e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06223594e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55281067e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65884857e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29228859e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74253922e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #34] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31934143e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 34: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:04 [abstract.py:324] It took 0.041899 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:04 [loggers.py:236] Engine 000: Avg prompt throughput: 28.4 tokens/s, Avg generation throughput: 49.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12879.05it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:29,  5.31it/s, est. speed input: 318.56 toks/s, output: 74.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:02, 50.60it/s, est. speed input: 2588.67 toks/s, output: 780.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:02, 49.53it/s, est. speed input: 2719.67 toks/s, output: 1007.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 35.17it/s, est. speed input: 2236.68 toks/s, output: 1112.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:06, 20.43it/s, est. speed input: 1583.76 toks/s, output: 1028.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:04<00:34,  3.66it/s, est. speed input: 432.94 toks/s, output: 500.79 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12245.0 tokens/s, Running: 123 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:10<01:13,  1.67it/s, est. speed input: 224.44 toks/s, output: 450.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:12<01:20,  1.48it/s, est. speed input: 196.42 toks/s, output: 552.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:13<01:21,  1.46it/s, est. speed input: 189.89 toks/s, output: 615.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:13<01:21,  1.44it/s, est. speed input: 183.78 toks/s, output: 676.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:16<01:38,  1.18it/s, est. speed input: 163.18 toks/s, output: 747.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:18<01:51,  1.03it/s, est. speed input: 152.14 toks/s, output: 774.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:19<01:50,  1.03it/s, est. speed input: 147.44 toks/s, output: 826.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10239.4 tokens/s, Running: 114 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:20<01:54,  1.01s/it, est. speed input: 141.96 toks/s, output: 870.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:21<01:26,  1.28it/s, est. speed input: 142.47 toks/s, output: 1017.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:22<01:35,  1.15it/s, est. speed input: 137.59 toks/s, output: 1053.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:25<02:30,  1.38s/it, est. speed input: 123.23 toks/s, output: 1015.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:27<02:39,  1.47s/it, est. speed input: 117.46 toks/s, output: 1039.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:27<01:59,  1.12s/it, est. speed input: 119.16 toks/s, output: 1123.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:27<01:34,  1.13it/s, est. speed input: 120.13 toks/s, output: 1201.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:27<01:15,  1.40it/s, est. speed input: 121.13 toks/s, output: 1278.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:28<01:25,  1.21it/s, est. speed input: 118.65 toks/s, output: 1318.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:29<00:50,  1.99it/s, est. speed input: 122.01 toks/s, output: 1549.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8866.5 tokens/s, Running: 100 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:30<00:43,  2.29it/s, est. speed input: 123.57 toks/s, output: 1692.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:31<00:53,  1.82it/s, est. speed input: 121.55 toks/s, output: 1726.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:32<01:18,  1.24it/s, est. speed input: 117.03 toks/s, output: 1724.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:33<01:11,  1.35it/s, est. speed input: 117.05 toks/s, output: 1784.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:34<01:16,  1.24it/s, est. speed input: 115.35 toks/s, output: 1820.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:35<01:16,  1.23it/s, est. speed input: 114.37 toks/s, output: 1865.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:36<01:22,  1.13it/s, est. speed input: 112.68 toks/s, output: 1898.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:37<01:19,  1.16it/s, est. speed input: 111.79 toks/s, output: 1944.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:37<01:07,  1.34it/s, est. speed input: 112.17 toks/s, output: 2008.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7753.0 tokens/s, Running: 91 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:40<02:14,  1.49s/it, est. speed input: 104.58 toks/s, output: 1932.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:41<01:14,  1.18it/s, est. speed input: 107.16 toks/s, output: 2099.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:42<01:33,  1.08s/it, est. speed input: 104.05 toks/s, output: 2097.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:44<01:38,  1.14s/it, est. speed input: 102.27 toks/s, output: 2121.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:44<01:21,  1.05it/s, est. speed input: 102.60 toks/s, output: 2185.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:45<00:59,  1.40it/s, est. speed input: 103.39 toks/s, output: 2318.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:45<00:40,  1.99it/s, est. speed input: 105.30 toks/s, output: 2473.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:46<00:40,  1.96it/s, est. speed input: 105.41 toks/s, output: 2530.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:46<00:33,  2.33it/s, est. speed input: 106.35 toks/s, output: 2607.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:47<00:37,  2.05it/s, est. speed input: 106.18 toks/s, output: 2657.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:47<00:41,  1.85it/s, est. speed input: 105.88 toks/s, output: 2705.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:47<00:25,  2.92it/s, est. speed input: 108.06 toks/s, output: 2868.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:49<00:53,  1.39it/s, est. speed input: 105.03 toks/s, output: 2841.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:49<00:41,  1.77it/s, est. speed input: 106.00 toks/s, output: 2920.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:03:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6625.2 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:50<00:34,  2.08it/s, est. speed input: 106.65 toks/s, output: 2991.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:50<00:31,  2.24it/s, est. speed input: 107.07 toks/s, output: 3056.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:51<00:33,  2.07it/s, est. speed input: 107.06 toks/s, output: 3106.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:51<00:26,  2.62it/s, est. speed input: 107.98 toks/s, output: 3184.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:51<00:22,  3.04it/s, est. speed input: 108.73 toks/s, output: 3257.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:52<00:26,  2.52it/s, est. speed input: 108.69 toks/s, output: 3307.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:52<00:26,  2.43it/s, est. speed input: 109.29 toks/s, output: 3423.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:53<00:29,  2.17it/s, est. speed input: 109.14 toks/s, output: 3470.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:53<00:28,  2.25it/s, est. speed input: 109.52 toks/s, output: 3530.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:54<00:23,  2.61it/s, est. speed input: 110.17 toks/s, output: 3601.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:54<00:21,  2.84it/s, est. speed input: 110.76 toks/s, output: 3669.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:54<00:17,  3.49it/s, est. speed input: 111.63 toks/s, output: 3746.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:54<00:16,  3.61it/s, est. speed input: 112.19 toks/s, output: 3815.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:55<00:19,  3.05it/s, est. speed input: 112.34 toks/s, output: 3870.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:55<00:15,  3.74it/s, est. speed input: 113.22 toks/s, output: 3947.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:55<00:16,  3.33it/s, est. speed input: 113.51 toks/s, output: 4006.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:55<00:15,  3.56it/s, est. speed input: 114.10 toks/s, output: 4076.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:56<00:13,  4.00it/s, est. speed input: 115.35 toks/s, output: 4217.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:56<00:16,  3.23it/s, est. speed input: 115.44 toks/s, output: 4267.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:57<00:14,  3.55it/s, est. speed input: 116.08 toks/s, output: 4338.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:57<00:11,  4.18it/s, est. speed input: 116.91 toks/s, output: 4415.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:57<00:11,  4.30it/s, est. speed input: 117.50 toks/s, output: 4485.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:57<00:11,  4.21it/s, est. speed input: 118.02 toks/s, output: 4552.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:57<00:09,  4.86it/s, est. speed input: 118.79 toks/s, output: 4629.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:58<00:16,  2.80it/s, est. speed input: 118.38 toks/s, output: 4659.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:58<00:14,  3.06it/s, est. speed input: 118.96 toks/s, output: 4726.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:58<00:08,  4.82it/s, est. speed input: 120.70 toks/s, output: 4889.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:04:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5660.8 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:00<00:19,  2.11it/s, est. speed input: 119.15 toks/s, output: 4872.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:00<00:18,  2.25it/s, est. speed input: 119.43 toks/s, output: 4931.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:00<00:15,  2.55it/s, est. speed input: 119.95 toks/s, output: 4999.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:01<00:16,  2.39it/s, est. speed input: 120.01 toks/s, output: 5047.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:01<00:11,  3.36it/s, est. speed input: 121.40 toks/s, output: 5200.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:01<00:10,  3.56it/s, est. speed input: 121.96 toks/s, output: 5269.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:01<00:06,  5.34it/s, est. speed input: 123.77 toks/s, output: 5437.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:03<00:14,  2.34it/s, est. speed input: 122.30 toks/s, output: 5419.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:03<00:11,  2.84it/s, est. speed input: 123.01 toks/s, output: 5498.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:03<00:09,  3.27it/s, est. speed input: 123.67 toks/s, output: 5572.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:03<00:10,  2.80it/s, est. speed input: 123.64 toks/s, output: 5618.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:04<00:08,  3.25it/s, est. speed input: 124.26 toks/s, output: 5693.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:04<00:06,  4.49it/s, est. speed input: 125.68 toks/s, output: 5852.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:04<00:07,  3.64it/s, est. speed input: 125.79 toks/s, output: 5903.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:05<00:04,  4.81it/s, est. speed input: 127.25 toks/s, output: 6063.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:05<00:05,  4.51it/s, est. speed input: 127.69 toks/s, output: 6129.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:06<00:08,  2.61it/s, est. speed input: 126.89 toks/s, output: 6139.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:07<00:10,  1.97it/s, est. speed input: 126.16 toks/s, output: 6151.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:07<00:11,  1.81it/s, est. speed input: 125.79 toks/s, output: 6183.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:08<00:12,  1.57it/s, est. speed input: 125.11 toks/s, output: 6199.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:08<00:06,  2.57it/s, est. speed input: 126.57 toks/s, output: 6373.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:04:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3494.7 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:11<00:13,  1.16it/s, est. speed input: 123.23 toks/s, output: 6256.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:11<00:07,  1.81it/s, est. speed input: 124.54 toks/s, output: 6428.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:12<00:08,  1.51it/s, est. speed input: 123.58 toks/s, output: 6432.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:12<00:06,  1.82it/s, est. speed input: 124.09 toks/s, output: 6510.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:12<00:03,  2.67it/s, est. speed input: 125.36 toks/s, output: 6681.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:13<00:03,  2.68it/s, est. speed input: 125.55 toks/s, output: 6744.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:13<00:02,  3.16it/s, est. speed input: 126.12 toks/s, output: 6828.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:14<00:03,  2.10it/s, est. speed input: 125.34 toks/s, output: 6841.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:14<00:02,  2.31it/s, est. speed input: 125.72 toks/s, output: 6970.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:16<00:03,  1.26it/s, est. speed input: 123.23 toks/s, output: 6894.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:17<00:02,  1.46it/s, est. speed input: 123.41 toks/s, output: 6965.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:17<00:01,  1.56it/s, est. speed input: 123.37 toks/s, output: 7024.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:17<00:00,  1.98it/s, est. speed input: 123.96 toks/s, output: 7116.65 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:17<00:00,  1.98it/s, est. speed input: 124.76 toks/s, output: 7221.69 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:17<00:00,  2.05it/s, est. speed input: 124.76 toks/s, output: 7221.69 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:04:27 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:04:27 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:04:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:04:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:04:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:04:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:04:28 [abstract.py:306] It took 0.593379 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:04:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:04:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:04:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:04:28 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.0312
+  • Mean Generation Length: 3519.5000
+  • Training FLOPS: 49.58 TFLOPS (12.40 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 391.67s
+  • policy_training: 254.06s (64.9%)
+  • generation: 83.73s (21.4%)
+  • teacher_logprob_inference: 48.90s (12.5%)
+  • policy_training/sharding_data: 4.31s (1.1%)
+  • prepare_for_generation: 3.41s (0.9%)
+  • policy_training/submit_training_futures: 2.23s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.50s (0.4%)
+  • training_prep: 0.86s (0.2%)
+  • teacher_logprob_inference_prep: 0.52s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 35: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:09:33 [abstract.py:324] It took 0.051885 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54021988e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65799408e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54021988e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65799408e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54021988e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65799408e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54021988e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65799408e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06250879e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55900764e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66514435e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26062603e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.37759686e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #35] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43149155e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06250879e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55900764e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66514435e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66640549e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26761227e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #35] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.11132584e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06250879e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55900764e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66514435e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28455505e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=3.45535517e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #35] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.10678062e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06250879e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55900764e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66514435e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27564163e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.74156742e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #35] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31837433e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 35: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:09:36 [abstract.py:324] It took 0.039627 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13051.88it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 01:09:36 [loggers.py:236] Engine 000: Avg prompt throughput: 30.2 tokens/s, Avg generation throughput: 41.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:00,  2.63it/s, est. speed input: 152.31 toks/s, output: 31.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 14.32it/s, est. speed input: 702.80 toks/s, output: 230.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 29.64it/s, est. speed input: 1322.61 toks/s, output: 543.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 39.80it/s, est. speed input: 1714.92 toks/s, output: 850.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 39.75it/s, est. speed input: 1834.16 toks/s, output: 1067.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 33.16it/s, est. speed input: 1778.36 toks/s, output: 1236.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 32.57it/s, est. speed input: 1789.27 toks/s, output: 1464.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:04, 27.41it/s, est. speed input: 1694.37 toks/s, output: 1571.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:04, 24.59it/s, est. speed input: 1627.99 toks/s, output: 1711.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:07<00:48,  2.26it/s, est. speed input: 405.99 toks/s, output: 606.38 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:08<00:50,  2.13it/s, est. speed input: 362.97 toks/s, output: 721.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:09:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11704.3 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:10<00:57,  1.86it/s, est. speed input: 317.27 toks/s, output: 800.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:11<01:00,  1.73it/s, est. speed input: 298.30 toks/s, output: 838.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:12<01:13,  1.41it/s, est. speed input: 268.55 toks/s, output: 841.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:13<01:12,  1.41it/s, est. speed input: 259.20 toks/s, output: 897.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:15<01:31,  1.12it/s, est. speed input: 234.80 toks/s, output: 897.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:15<01:19,  1.27it/s, est. speed input: 232.86 toks/s, output: 972.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:16<01:17,  1.27it/s, est. speed input: 218.52 toks/s, output: 1070.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:18<01:32,  1.06it/s, est. speed input: 203.98 toks/s, output: 1081.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:18<01:12,  1.33it/s, est. speed input: 205.69 toks/s, output: 1171.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:19<01:23,  1.15it/s, est. speed input: 196.23 toks/s, output: 1197.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:20<01:10,  1.34it/s, est. speed input: 195.25 toks/s, output: 1270.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:09:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9600.0 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:20<00:47,  1.94it/s, est. speed input: 197.13 toks/s, output: 1437.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:21<00:55,  1.65it/s, est. speed input: 191.72 toks/s, output: 1474.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:21<00:48,  1.87it/s, est. speed input: 191.66 toks/s, output: 1550.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:22<00:25,  3.47it/s, est. speed input: 198.31 toks/s, output: 1828.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:22<00:34,  2.49it/s, est. speed input: 193.55 toks/s, output: 1856.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:23<00:23,  3.61it/s, est. speed input: 197.84 toks/s, output: 2040.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:25<00:56,  1.48it/s, est. speed input: 182.86 toks/s, output: 1957.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:25<00:49,  1.68it/s, est. speed input: 182.92 toks/s, output: 2028.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:28<01:28,  1.08s/it, est. speed input: 168.41 toks/s, output: 1938.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:28<01:17,  1.04it/s, est. speed input: 166.89 toks/s, output: 1991.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:29<01:06,  1.21it/s, est. speed input: 166.19 toks/s, output: 2054.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:29<00:51,  1.52it/s, est. speed input: 167.07 toks/s, output: 2134.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:29<00:42,  1.82it/s, est. speed input: 167.54 toks/s, output: 2209.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7644.3 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:31<01:02,  1.22it/s, est. speed input: 161.58 toks/s, output: 2199.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:31<00:56,  1.33it/s, est. speed input: 160.50 toks/s, output: 2254.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:32<00:45,  1.64it/s, est. speed input: 161.02 toks/s, output: 2329.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:33<00:56,  1.32it/s, est. speed input: 157.42 toks/s, output: 2346.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:33<00:52,  1.38it/s, est. speed input: 156.16 toks/s, output: 2396.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:34<00:44,  1.62it/s, est. speed input: 156.27 toks/s, output: 2465.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:35<01:02,  1.14it/s, est. speed input: 151.41 toks/s, output: 2456.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:36<00:36,  1.87it/s, est. speed input: 153.04 toks/s, output: 2682.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:36<00:32,  2.06it/s, est. speed input: 153.39 toks/s, output: 2753.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:37<00:30,  2.19it/s, est. speed input: 153.52 toks/s, output: 2821.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:37<00:24,  2.67it/s, est. speed input: 154.60 toks/s, output: 2905.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:38<00:32,  1.96it/s, est. speed input: 152.16 toks/s, output: 2986.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:38<00:20,  2.93it/s, est. speed input: 154.80 toks/s, output: 3166.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<00:25,  2.36it/s, est. speed input: 153.49 toks/s, output: 3202.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6661.0 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:40<00:35,  1.65it/s, est. speed input: 150.41 toks/s, output: 3202.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:42<00:36,  1.57it/s, est. speed input: 148.43 toks/s, output: 3286.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:42<00:29,  1.87it/s, est. speed input: 149.25 toks/s, output: 3368.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:42<00:25,  2.17it/s, est. speed input: 149.94 toks/s, output: 3446.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:42<00:20,  2.64it/s, est. speed input: 150.86 toks/s, output: 3531.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:44<00:41,  1.27it/s, est. speed input: 145.81 toks/s, output: 3476.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:46<00:51,  1.02it/s, est. speed input: 142.43 toks/s, output: 3459.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:47<00:56,  1.11s/it, est. speed input: 139.41 toks/s, output: 3452.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:49<01:07,  1.35s/it, est. speed input: 135.15 toks/s, output: 3412.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:49<00:54,  1.12s/it, est. speed input: 134.85 toks/s, output: 3472.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5624.1 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:50<00:42,  1.13it/s, est. speed input: 135.20 toks/s, output: 3546.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:50<00:37,  1.24it/s, est. speed input: 134.76 toks/s, output: 3601.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:52<00:47,  1.02s/it, est. speed input: 131.97 toks/s, output: 3593.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:52<00:33,  1.33it/s, est. speed input: 132.85 toks/s, output: 3682.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:52<00:25,  1.73it/s, est. speed input: 133.53 toks/s, output: 3768.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:52<00:19,  2.18it/s, est. speed input: 134.22 toks/s, output: 3854.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:53<00:25,  1.64it/s, est. speed input: 132.94 toks/s, output: 3883.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:53<00:14,  2.81it/s, est. speed input: 134.88 toks/s, output: 4070.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:54<00:13,  2.96it/s, est. speed input: 135.29 toks/s, output: 4147.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:54<00:08,  4.57it/s, est. speed input: 137.24 toks/s, output: 4336.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:55<00:09,  3.81it/s, est. speed input: 137.80 toks/s, output: 4479.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:56<00:20,  1.64it/s, est. speed input: 134.22 toks/s, output: 4428.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:57<00:17,  1.84it/s, est. speed input: 134.29 toks/s, output: 4559.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:57<00:13,  2.22it/s, est. speed input: 135.11 toks/s, output: 4651.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:58<00:09,  3.19it/s, est. speed input: 136.79 toks/s, output: 4836.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:58<00:12,  2.32it/s, est. speed input: 135.82 toks/s, output: 4865.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:59<00:10,  2.46it/s, est. speed input: 136.10 toks/s, output: 4939.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:59<00:09,  2.73it/s, est. speed input: 136.80 toks/s, output: 5089.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4469.8 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:00<00:11,  2.17it/s, est. speed input: 135.98 toks/s, output: 5124.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:01<00:10,  2.27it/s, est. speed input: 136.16 toks/s, output: 5194.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:02<00:16,  1.35it/s, est. speed input: 133.60 toks/s, output: 5162.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:03<00:13,  1.57it/s, est. speed input: 133.82 toks/s, output: 5237.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:03<00:11,  1.81it/s, est. speed input: 134.09 toks/s, output: 5313.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:05<00:18,  1.01it/s, est. speed input: 130.73 toks/s, output: 5247.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:05<00:09,  1.72it/s, est. speed input: 132.28 toks/s, output: 5444.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:07<00:15,  1.01it/s, est. speed input: 128.74 toks/s, output: 5367.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:09<00:16,  1.07s/it, est. speed input: 127.20 toks/s, output: 5374.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2808.0 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:10<00:17,  1.23s/it, est. speed input: 125.04 toks/s, output: 5356.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:11<00:13,  1.06s/it, est. speed input: 124.83 toks/s, output: 5419.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:11<00:09,  1.21it/s, est. speed input: 125.22 toks/s, output: 5509.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:12<00:08,  1.32it/s, est. speed input: 125.06 toks/s, output: 5574.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:13<00:09,  1.03it/s, est. speed input: 123.35 toks/s, output: 5573.27 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:13<00:00,  1.03it/s, est. speed input: 131.60 toks/s, output: 6683.28 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:13<00:00,  2.17it/s, est. speed input: 131.60 toks/s, output: 6683.28 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:54 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:54 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:10:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:10:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:10:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:10:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:10:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:10:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:10:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 01:10:55 [abstract.py:306] It took 0.601109 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:10:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0142
+  • Training Accuracy: 0.0250
+  • Mean Generation Length: 3082.6438
+  • Training FLOPS: 47.78 TFLOPS (11.95 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 360.12s
+  • policy_training: 232.82s (64.7%)
+  • generation: 78.86s (21.9%)
+  • teacher_logprob_inference: 43.61s (12.1%)
+  • policy_training/sharding_data: 4.33s (1.2%)
+  • prepare_for_generation: 3.38s (0.9%)
+  • policy_training/submit_training_futures: 2.01s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.4%)
+  • training_prep: 0.86s (0.2%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 8/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 36: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:15:34 [abstract.py:324] It took 0.054234 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53487816e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65041046e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53487816e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65041046e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53487816e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65041046e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06244961e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53487816e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65041046e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06248730e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55235405e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66440887e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25918579e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.32160234e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #36] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43282013e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06248730e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55235405e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66440887e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66969452e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26599045e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #36] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.12022552e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06248730e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55235405e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66440887e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28506775e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=3.34521770e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #36] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13102798e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06248730e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55235405e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66440887e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28222580e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.73416672e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #36] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31808441e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 36: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:15:37 [abstract.py:324] It took 0.044593 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13294.94it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:15:38 [loggers.py:236] Engine 000: Avg prompt throughput: 33.4 tokens/s, Avg generation throughput: 27.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:37,  1.63it/s, est. speed input: 94.47 toks/s, output: 19.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:12, 12.41it/s, est. speed input: 589.60 toks/s, output: 154.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:07, 19.19it/s, est. speed input: 879.17 toks/s, output: 304.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:11, 12.93it/s, est. speed input: 737.74 toks/s, output: 373.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:11, 12.43it/s, est. speed input: 730.36 toks/s, output: 498.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:12, 10.81it/s, est. speed input: 688.69 toks/s, output: 558.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:02<00:11, 11.53it/s, est. speed input: 702.91 toks/s, output: 666.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:02<00:13,  9.60it/s, est. speed input: 661.46 toks/s, output: 723.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:05<01:05,  2.00it/s, est. speed input: 302.47 toks/s, output: 465.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:06<01:05,  2.00it/s, est. speed input: 287.55 toks/s, output: 514.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:08<01:43,  1.26it/s, est. speed input: 221.62 toks/s, output: 471.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:09<01:45,  1.22it/s, est. speed input: 205.99 toks/s, output: 512.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:10<01:48,  1.18it/s, est. speed input: 192.69 toks/s, output: 553.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:15:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12414.8 tokens/s, Running: 128 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:10<01:33,  1.35it/s, est. speed input: 191.20 toks/s, output: 620.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:10<01:16,  1.65it/s, est. speed input: 192.46 toks/s, output: 695.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:12<02:00,  1.03it/s, est. speed input: 168.07 toks/s, output: 677.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:12<01:33,  1.33it/s, est. speed input: 170.31 toks/s, output: 755.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:13<01:23,  1.48it/s, est. speed input: 168.61 toks/s, output: 816.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:14<01:29,  1.36it/s, est. speed input: 162.38 toks/s, output: 853.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:15<01:39,  1.21it/s, est. speed input: 155.37 toks/s, output: 883.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:16<02:06,  1.05s/it, est. speed input: 144.28 toks/s, output: 887.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:19<02:43,  1.37s/it, est. speed input: 131.28 toks/s, output: 875.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:19<02:16,  1.15s/it, est. speed input: 130.19 toks/s, output: 934.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:15:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10498.7 tokens/s, Running: 118 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:21<02:38,  1.36s/it, est. speed input: 121.89 toks/s, output: 941.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:21<01:55,  1.00it/s, est. speed input: 123.83 toks/s, output: 1021.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:22<01:56,  1.02s/it, est. speed input: 120.80 toks/s, output: 1060.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:23<01:53,  1.00it/s, est. speed input: 118.59 toks/s, output: 1104.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:23<01:28,  1.28it/s, est. speed input: 119.71 toks/s, output: 1177.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:24<01:29,  1.25it/s, est. speed input: 118.14 toks/s, output: 1224.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:25<01:16,  1.46it/s, est. speed input: 118.59 toks/s, output: 1290.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:26<01:25,  1.29it/s, est. speed input: 116.50 toks/s, output: 1327.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:28<01:58,  1.09s/it, est. speed input: 111.10 toks/s, output: 1326.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:28<01:53,  1.05s/it, est. speed input: 109.54 toks/s, output: 1368.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:29<01:33,  1.14it/s, est. speed input: 109.82 toks/s, output: 1431.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:16:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9097.2 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:30<01:39,  1.06it/s, est. speed input: 107.87 toks/s, output: 1465.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:31<01:42,  1.02it/s, est. speed input: 106.21 toks/s, output: 1501.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:32<01:06,  1.55it/s, est. speed input: 108.13 toks/s, output: 1646.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:33<01:25,  1.19it/s, est. speed input: 105.22 toks/s, output: 1660.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:33<01:07,  1.49it/s, est. speed input: 106.40 toks/s, output: 1735.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:34<01:13,  1.36it/s, est. speed input: 105.42 toks/s, output: 1774.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:35<01:21,  1.22it/s, est. speed input: 104.07 toks/s, output: 1807.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:36<01:12,  1.35it/s, est. speed input: 104.17 toks/s, output: 1864.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:37<01:18,  1.24it/s, est. speed input: 103.12 toks/s, output: 1900.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:37<01:17,  1.25it/s, est. speed input: 102.57 toks/s, output: 1944.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:38<01:09,  1.37it/s, est. speed input: 102.71 toks/s, output: 2001.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:39<01:02,  1.51it/s, est. speed input: 102.94 toks/s, output: 2059.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:16:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8001.4 tokens/s, Running: 94 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:41<01:45,  1.14s/it, est. speed input: 98.79 toks/s, output: 2030.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:42<01:50,  1.20s/it, est. speed input: 97.10 toks/s, output: 2050.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:43<01:24,  1.06it/s, est. speed input: 96.99 toks/s, output: 2154.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:45<01:41,  1.14s/it, est. speed input: 94.65 toks/s, output: 2156.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:46<01:40,  1.14s/it, est. speed input: 93.64 toks/s, output: 2186.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:48<01:54,  1.31s/it, est. speed input: 91.42 toks/s, output: 2189.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:49<01:11,  1.19it/s, est. speed input: 93.09 toks/s, output: 2333.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:16:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7031.4 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:50<01:27,  1.04s/it, est. speed input: 91.17 toks/s, output: 2339.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:51<01:20,  1.03it/s, est. speed input: 91.03 toks/s, output: 2388.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:52<01:18,  1.04it/s, est. speed input: 90.53 toks/s, output: 2427.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:52<01:02,  1.30it/s, est. speed input: 91.24 toks/s, output: 2498.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:52<00:37,  2.09it/s, est. speed input: 93.16 toks/s, output: 2652.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:53<00:24,  3.15it/s, est. speed input: 95.29 toks/s, output: 2810.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:54<00:40,  1.87it/s, est. speed input: 94.08 toks/s, output: 2824.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:54<00:32,  2.28it/s, est. speed input: 94.97 toks/s, output: 2900.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:54<00:22,  3.19it/s, est. speed input: 96.79 toks/s, output: 3051.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:55<00:35,  2.00it/s, est. speed input: 95.89 toks/s, output: 3071.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:56<00:37,  1.87it/s, est. speed input: 95.88 toks/s, output: 3118.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:56<00:31,  2.24it/s, est. speed input: 96.60 toks/s, output: 3190.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:57<00:32,  2.13it/s, est. speed input: 96.79 toks/s, output: 3243.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:57<00:24,  2.69it/s, est. speed input: 98.02 toks/s, output: 3379.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:58<00:22,  2.91it/s, est. speed input: 98.67 toks/s, output: 3447.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:58<00:18,  3.45it/s, est. speed input: 99.49 toks/s, output: 3522.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:58<00:27,  2.37it/s, est. speed input: 99.23 toks/s, output: 3557.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:59<00:26,  2.39it/s, est. speed input: 99.58 toks/s, output: 3615.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:16:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5957.5 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:00<00:45,  1.36it/s, est. speed input: 98.06 toks/s, output: 3607.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:01<00:38,  1.59it/s, est. speed input: 98.46 toks/s, output: 3668.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:01<00:23,  2.52it/s, est. speed input: 100.10 toks/s, output: 3820.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:01<00:20,  2.90it/s, est. speed input: 100.78 toks/s, output: 3892.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:02<00:26,  2.10it/s, est. speed input: 100.57 toks/s, output: 3974.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:03<00:30,  1.81it/s, est. speed input: 100.28 toks/s, output: 4007.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:04<00:25,  2.08it/s, est. speed input: 100.79 toks/s, output: 4075.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:04<00:20,  2.58it/s, est. speed input: 101.50 toks/s, output: 4151.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:04<00:18,  2.84it/s, est. speed input: 102.04 toks/s, output: 4218.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:04<00:16,  3.15it/s, est. speed input: 102.61 toks/s, output: 4287.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:05<00:16,  2.92it/s, est. speed input: 103.25 toks/s, output: 4405.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:05<00:16,  2.86it/s, est. speed input: 103.58 toks/s, output: 4464.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:06<00:14,  3.20it/s, est. speed input: 104.59 toks/s, output: 4597.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:07<00:22,  2.03it/s, est. speed input: 103.80 toks/s, output: 4606.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:07<00:22,  1.96it/s, est. speed input: 103.79 toks/s, output: 4653.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:09<00:28,  1.51it/s, est. speed input: 103.02 toks/s, output: 4664.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:09<00:21,  1.94it/s, est. speed input: 103.74 toks/s, output: 4741.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:09<00:17,  2.38it/s, est. speed input: 104.36 toks/s, output: 4815.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:09<00:13,  2.88it/s, est. speed input: 104.99 toks/s, output: 4889.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:09<00:12,  3.22it/s, est. speed input: 105.52 toks/s, output: 4959.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:10<00:10,  3.55it/s, est. speed input: 106.53 toks/s, output: 5095.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:16:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4879.5 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:12<00:26,  1.36it/s, est. speed input: 104.24 toks/s, output: 5029.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:13<00:28,  1.25it/s, est. speed input: 103.68 toks/s, output: 5048.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:13<00:25,  1.34it/s, est. speed input: 103.67 toks/s, output: 5093.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:14<00:23,  1.39it/s, est. speed input: 103.60 toks/s, output: 5136.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:18<00:36,  1.19s/it, est. speed input: 100.42 toks/s, output: 5072.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:18<00:25,  1.16it/s, est. speed input: 101.08 toks/s, output: 5204.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:19<00:21,  1.29it/s, est. speed input: 101.26 toks/s, output: 5262.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:20<00:23,  1.17it/s, est. speed input: 100.63 toks/s, output: 5279.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:20<00:19,  1.34it/s, est. speed input: 100.86 toks/s, output: 5342.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:16:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3944.5 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:20<00:10,  2.20it/s, est. speed input: 102.29 toks/s, output: 5514.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:21<00:12,  1.89it/s, est. speed input: 102.06 toks/s, output: 5552.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:22<00:10,  2.01it/s, est. speed input: 102.28 toks/s, output: 5615.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:22<00:09,  2.20it/s, est. speed input: 102.61 toks/s, output: 5683.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:23<00:11,  1.75it/s, est. speed input: 102.23 toks/s, output: 5713.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:23<00:09,  2.05it/s, est. speed input: 102.64 toks/s, output: 5787.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:25<00:14,  1.28it/s, est. speed input: 101.54 toks/s, output: 5776.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:25<00:12,  1.33it/s, est. speed input: 101.43 toks/s, output: 5823.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:26<00:12,  1.28it/s, est. speed input: 101.13 toks/s, output: 5859.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:26<00:09,  1.60it/s, est. speed input: 101.51 toks/s, output: 5936.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:27<00:07,  1.83it/s, est. speed input: 101.81 toks/s, output: 6005.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:27<00:00,  1.83it/s, est. speed input: 111.53 toks/s, output: 7319.44 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:27<00:00,  1.83it/s, est. speed input: 111.53 toks/s, output: 7319.44 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:17:10 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:17:10 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:17:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:17:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:17:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:17:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:17:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:17:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:17:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 01:17:11 [abstract.py:306] It took 0.553295 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:17:11 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.0187
+  • Mean Generation Length: 3993.2812
+  • Training FLOPS: 54.09 TFLOPS (13.52 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 419.29s
+  • policy_training: 267.54s (63.8%)
+  • generation: 93.63s (22.3%)
+  • teacher_logprob_inference: 52.86s (12.6%)
+  • policy_training/sharding_data: 4.38s (1.0%)
+  • prepare_for_generation: 3.78s (0.9%)
+  • policy_training/submit_training_futures: 1.97s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.91s (0.5%)
+  • training_prep: 0.85s (0.2%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 37: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:22:33 [abstract.py:324] It took 0.052063 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06298125e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52387390e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65427094e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06298125e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52387390e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65427094e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06298125e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52387390e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65427094e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06298125e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52387390e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65427094e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06307773e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51547432e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65888062e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25239525e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.23679209e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #37] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43200134e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06307773e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51547432e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65888062e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66260834e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26808643e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #37] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.11840858e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06307773e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51547432e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65888062e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29390182e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=2.39971161e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #37] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.12810974e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06307773e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51547432e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65888062e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29008255e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.72917843e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #37] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31839783e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 37: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:22:36 [abstract.py:324] It took 0.044430 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12400.70it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:22:36 [loggers.py:236] Engine 000: Avg prompt throughput: 28.9 tokens/s, Avg generation throughput: 54.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:45,  3.50it/s, est. speed input: 216.81 toks/s, output: 45.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 19.58it/s, est. speed input: 1000.06 toks/s, output: 315.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 26.40it/s, est. speed input: 1328.69 toks/s, output: 540.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 38.33it/s, est. speed input: 1776.19 toks/s, output: 913.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 36.25it/s, est. speed input: 1822.24 toks/s, output: 1075.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:03, 34.43it/s, est. speed input: 1838.85 toks/s, output: 1265.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:03, 32.88it/s, est. speed input: 1834.79 toks/s, output: 1416.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:05, 22.90it/s, est. speed input: 1606.65 toks/s, output: 1426.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:06, 19.49it/s, est. speed input: 1483.63 toks/s, output: 1518.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:02<00:12,  9.38it/s, est. speed input: 1056.11 toks/s, output: 1241.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:04<00:31,  3.61it/s, est. speed input: 608.72 toks/s, output: 872.64 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:05<00:29,  3.82it/s, est. speed input: 584.10 toks/s, output: 1000.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:08<01:01,  1.80it/s, est. speed input: 376.28 toks/s, output: 787.74 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:08<00:47,  2.28it/s, est. speed input: 384.48 toks/s, output: 972.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:22:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11724.0 tokens/s, Running: 108 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:10<01:21,  1.32it/s, est. speed input: 299.08 toks/s, output: 841.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:11<01:20,  1.32it/s, est. speed input: 285.08 toks/s, output: 887.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:13<01:43,  1.02it/s, est. speed input: 249.48 toks/s, output: 861.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:13<01:27,  1.19it/s, est. speed input: 247.73 toks/s, output: 939.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:15<01:49,  1.07s/it, est. speed input: 223.34 toks/s, output: 930.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:15<01:24,  1.20it/s, est. speed input: 225.16 toks/s, output: 1019.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:17<01:49,  1.09s/it, est. speed input: 205.82 toks/s, output: 1013.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:17<01:23,  1.20it/s, est. speed input: 207.16 toks/s, output: 1101.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:17<01:02,  1.58it/s, est. speed input: 209.27 toks/s, output: 1191.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:17<00:47,  2.07it/s, est. speed input: 211.35 toks/s, output: 1281.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:18<00:50,  1.93it/s, est. speed input: 207.80 toks/s, output: 1337.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:19<01:10,  1.37it/s, est. speed input: 197.77 toks/s, output: 1350.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:22:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9656.0 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:20<01:05,  1.45it/s, est. speed input: 195.10 toks/s, output: 1408.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:21<01:11,  1.32it/s, est. speed input: 189.62 toks/s, output: 1444.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:21<01:02,  1.50it/s, est. speed input: 188.50 toks/s, output: 1509.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:22<00:42,  2.14it/s, est. speed input: 190.06 toks/s, output: 1670.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:22<00:47,  1.91it/s, est. speed input: 186.86 toks/s, output: 1715.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:23<00:46,  1.92it/s, est. speed input: 185.33 toks/s, output: 1773.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:23<00:44,  1.96it/s, est. speed input: 184.24 toks/s, output: 1833.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:24<00:42,  2.06it/s, est. speed input: 183.58 toks/s, output: 1897.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:24<00:37,  2.27it/s, est. speed input: 183.53 toks/s, output: 1968.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:25<00:39,  2.17it/s, est. speed input: 182.23 toks/s, output: 2024.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:26<00:40,  2.03it/s, est. speed input: 179.48 toks/s, output: 2130.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:27<00:49,  1.66it/s, est. speed input: 175.60 toks/s, output: 2152.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:27<00:43,  1.86it/s, est. speed input: 175.56 toks/s, output: 2220.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:27<00:35,  2.25it/s, est. speed input: 176.63 toks/s, output: 2300.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:28<00:29,  2.66it/s, est. speed input: 177.44 toks/s, output: 2443.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:29<00:52,  1.47it/s, est. speed input: 169.73 toks/s, output: 2404.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7901.2 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:31<01:02,  1.21it/s, est. speed input: 164.97 toks/s, output: 2403.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:31<00:54,  1.38it/s, est. speed input: 164.55 toks/s, output: 2462.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:31<00:44,  1.68it/s, est. speed input: 165.22 toks/s, output: 2537.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:32<00:25,  2.78it/s, est. speed input: 168.40 toks/s, output: 2716.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:32<00:22,  3.14it/s, est. speed input: 169.25 toks/s, output: 2794.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:32<00:26,  2.64it/s, est. speed input: 168.17 toks/s, output: 2840.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:35<00:42,  1.57it/s, est. speed input: 160.82 toks/s, output: 2899.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:35<00:21,  2.99it/s, est. speed input: 167.14 toks/s, output: 3266.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:36<00:25,  2.36it/s, est. speed input: 164.26 toks/s, output: 3335.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:37<00:24,  2.47it/s, est. speed input: 164.58 toks/s, output: 3403.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:37<00:22,  2.67it/s, est. speed input: 165.13 toks/s, output: 3476.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:37<00:21,  2.68it/s, est. speed input: 165.15 toks/s, output: 3538.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:37<00:17,  3.18it/s, est. speed input: 166.22 toks/s, output: 3622.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:38<00:21,  2.60it/s, est. speed input: 165.16 toks/s, output: 3661.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:40<00:29,  1.80it/s, est. speed input: 161.74 toks/s, output: 3703.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6241.3 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:40<00:28,  1.87it/s, est. speed input: 161.33 toks/s, output: 3756.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:40<00:24,  2.14it/s, est. speed input: 161.85 toks/s, output: 3828.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:41<00:27,  1.86it/s, est. speed input: 160.38 toks/s, output: 3856.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:41<00:23,  2.17it/s, est. speed input: 160.89 toks/s, output: 3928.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:41<00:18,  2.65it/s, est. speed input: 161.81 toks/s, output: 4010.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:42<00:19,  2.46it/s, est. speed input: 161.43 toks/s, output: 4118.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:43<00:15,  2.94it/s, est. speed input: 162.33 toks/s, output: 4202.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:43<00:14,  3.11it/s, est. speed input: 162.70 toks/s, output: 4273.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:43<00:16,  2.74it/s, est. speed input: 162.27 toks/s, output: 4322.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:44<00:22,  1.94it/s, est. speed input: 160.40 toks/s, output: 4332.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:44<00:16,  2.51it/s, est. speed input: 161.40 toks/s, output: 4420.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:45<00:16,  2.55it/s, est. speed input: 161.36 toks/s, output: 4481.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:45<00:14,  2.79it/s, est. speed input: 161.72 toks/s, output: 4551.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:45<00:12,  3.17it/s, est. speed input: 162.27 toks/s, output: 4628.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:46<00:17,  2.16it/s, est. speed input: 160.42 toks/s, output: 4696.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:48<00:19,  1.77it/s, est. speed input: 158.16 toks/s, output: 4753.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:49<00:19,  1.70it/s, est. speed input: 157.30 toks/s, output: 4787.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:49<00:17,  1.89it/s, est. speed input: 157.40 toks/s, output: 4854.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:50<00:18,  1.71it/s, est. speed input: 156.21 toks/s, output: 4882.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5075.9 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:50<00:19,  1.60it/s, est. speed input: 155.18 toks/s, output: 4913.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:51<00:13,  2.21it/s, est. speed input: 156.16 toks/s, output: 5071.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:51<00:11,  2.45it/s, est. speed input: 156.53 toks/s, output: 5148.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:51<00:08,  3.00it/s, est. speed input: 157.33 toks/s, output: 5238.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:51<00:07,  3.41it/s, est. speed input: 157.95 toks/s, output: 5322.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:54<00:17,  1.41it/s, est. speed input: 152.93 toks/s, output: 5276.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:54<00:15,  1.53it/s, est. speed input: 152.74 toks/s, output: 5335.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:55<00:11,  1.86it/s, est. speed input: 153.35 toks/s, output: 5421.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:55<00:08,  2.36it/s, est. speed input: 154.09 toks/s, output: 5578.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:55<00:05,  3.12it/s, est. speed input: 155.53 toks/s, output: 5758.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:56<00:05,  3.25it/s, est. speed input: 155.89 toks/s, output: 5837.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:57<00:09,  1.77it/s, est. speed input: 153.09 toks/s, output: 5799.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:58<00:10,  1.40it/s, est. speed input: 151.04 toks/s, output: 5790.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:58<00:07,  1.78it/s, est. speed input: 151.78 toks/s, output: 5884.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:59<00:06,  2.12it/s, est. speed input: 152.20 toks/s, output: 5969.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3150.6 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 74.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:01<00:11,  1.08it/s, est. speed input: 147.99 toks/s, output: 5874.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:01<00:09,  1.17it/s, est. speed input: 147.42 toks/s, output: 5922.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:03<00:07,  1.25it/s, est. speed input: 145.91 toks/s, output: 6006.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:04<00:06,  1.21it/s, est. speed input: 144.78 toks/s, output: 6036.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:04<00:05,  1.34it/s, est. speed input: 144.54 toks/s, output: 6103.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:05<00:04,  1.47it/s, est. speed input: 144.37 toks/s, output: 6171.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:06<00:04,  1.12it/s, est. speed input: 142.15 toks/s, output: 6156.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:06<00:02,  1.46it/s, est. speed input: 142.67 toks/s, output: 6259.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:08<00:02,  1.21it/s, est. speed input: 141.12 toks/s, output: 6271.39 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  1.21it/s, est. speed input: 143.82 toks/s, output: 6632.64 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  2.35it/s, est. speed input: 143.82 toks/s, output: 6632.64 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:48 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:48 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:23:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:23:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:23:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:23:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:23:49 [abstract.py:306] It took 0.599655 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:23:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:23:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:23:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:23:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0176
+  • Training Accuracy: 0.0688
+  • Mean Generation Length: 2820.0437
+  • Training FLOPS: 46.97 TFLOPS (11.74 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 332.24s
+  • policy_training: 214.07s (64.4%)
+  • generation: 72.78s (21.9%)
+  • teacher_logprob_inference: 40.51s (12.2%)
+  • policy_training/sharding_data: 4.32s (1.3%)
+  • prepare_for_generation: 3.39s (1.0%)
+  • policy_training/submit_training_futures: 1.93s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.47s (0.4%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 38: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:06 [abstract.py:324] It took 0.053117 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06309688e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51404114e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65795898e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06309688e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51404114e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65795898e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06309688e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51404114e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65795898e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06309688e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51404114e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65795898e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06315508e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52472343e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66349640e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.24423084e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.26371574e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #38] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43116119e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06315508e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52472343e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66349640e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69087830e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27197380e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #38] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.17278671e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06315508e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52472343e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66349640e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28576431e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=2.18023777e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #38] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14327240e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06315508e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52472343e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66349640e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30547371e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.71481705e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #38] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31568848e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 38: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:08 [abstract.py:324] It took 0.040557 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13046.05it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:09 [loggers.py:236] Engine 000: Avg prompt throughput: 35.8 tokens/s, Avg generation throughput: 49.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:52,  3.02it/s, est. speed input: 181.17 toks/s, output: 36.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:11, 13.39it/s, est. speed input: 670.51 toks/s, output: 173.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:08, 18.44it/s, est. speed input: 910.07 toks/s, output: 335.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:07, 21.00it/s, est. speed input: 1050.53 toks/s, output: 502.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:06, 20.93it/s, est. speed input: 1085.44 toks/s, output: 619.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 27.74it/s, est. speed input: 1285.34 toks/s, output: 906.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:05, 26.61it/s, est. speed input: 1314.36 toks/s, output: 1083.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:07, 17.75it/s, est. speed input: 1151.96 toks/s, output: 1077.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:06, 19.32it/s, est. speed input: 1178.83 toks/s, output: 1247.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:08, 14.57it/s, est. speed input: 1069.98 toks/s, output: 1277.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:02<00:10, 12.15it/s, est. speed input: 996.38 toks/s, output: 1302.19 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:02<00:09, 12.52it/s, est. speed input: 987.53 toks/s, output: 1398.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:06<01:06,  1.80it/s, est. speed input: 386.55 toks/s, output: 702.44 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:06<00:55,  2.14it/s, est. speed input: 378.89 toks/s, output: 841.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:08<01:12,  1.61it/s, est. speed input: 320.43 toks/s, output: 790.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11964.8 tokens/s, Running: 117 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:10<01:57,  1.01s/it, est. speed input: 245.38 toks/s, output: 686.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:11<01:50,  1.04it/s, est. speed input: 235.08 toks/s, output: 736.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:14<02:35,  1.36s/it, est. speed input: 195.30 toks/s, output: 692.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:15<02:38,  1.40s/it, est. speed input: 180.31 toks/s, output: 718.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:17<02:36,  1.40s/it, est. speed input: 169.59 toks/s, output: 754.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:18<02:20,  1.27s/it, est. speed input: 164.43 toks/s, output: 809.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10059.7 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:20<02:41,  1.47s/it, est. speed input: 151.44 toks/s, output: 820.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:20<02:05,  1.15s/it, est. speed input: 151.69 toks/s, output: 898.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:21<01:51,  1.03s/it, est. speed input: 149.41 toks/s, output: 959.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:22<01:47,  1.00s/it, est. speed input: 145.75 toks/s, output: 1009.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:23<01:59,  1.13s/it, est. speed input: 139.57 toks/s, output: 1040.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:24<01:36,  1.09it/s, est. speed input: 139.57 toks/s, output: 1112.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:26<02:11,  1.26s/it, est. speed input: 130.93 toks/s, output: 1115.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:26<01:37,  1.06it/s, est. speed input: 132.19 toks/s, output: 1196.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:28<02:12,  1.29s/it, est. speed input: 124.53 toks/s, output: 1197.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8925.2 tokens/s, Running: 102 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.7%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:34<04:38,  2.75s/it, est. speed input: 104.10 toks/s, output: 1072.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:34<03:17,  1.97s/it, est. speed input: 105.50 toks/s, output: 1156.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:36<02:57,  1.79s/it, est. speed input: 103.20 toks/s, output: 1200.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:36<01:42,  1.05s/it, est. speed input: 105.48 toks/s, output: 1362.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7753.2 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.4%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:40<02:10,  1.37s/it, est. speed input: 98.88 toks/s, output: 1403.39 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:41<02:05,  1.34s/it, est. speed input: 97.42 toks/s, output: 1448.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:42<01:45,  1.14s/it, est. speed input: 97.80 toks/s, output: 1517.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:42<01:31,  1.00it/s, est. speed input: 97.85 toks/s, output: 1582.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:43<01:22,  1.10it/s, est. speed input: 97.75 toks/s, output: 1644.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:48<03:06,  2.07s/it, est. speed input: 88.69 toks/s, output: 1556.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:48<02:17,  1.55s/it, est. speed input: 89.54 toks/s, output: 1634.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:49<01:48,  1.23s/it, est. speed input: 89.93 toks/s, output: 1703.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:49<01:19,  1.09it/s, est. speed input: 90.93 toks/s, output: 1783.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:49<00:56,  1.49it/s, est. speed input: 91.98 toks/s, output: 1925.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:28:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7037.6 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.6%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:50<00:54,  1.55it/s, est. speed input: 92.09 toks/s, output: 1988.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:50<00:48,  1.71it/s, est. speed input: 92.52 toks/s, output: 2056.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:51<00:38,  2.12it/s, est. speed input: 93.40 toks/s, output: 2134.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:51<00:44,  1.83it/s, est. speed input: 93.23 toks/s, output: 2188.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:51<00:34,  2.35it/s, est. speed input: 94.16 toks/s, output: 2268.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:53<01:01,  1.28it/s, est. speed input: 92.35 toks/s, output: 2282.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:54<00:52,  1.49it/s, est. speed input: 92.82 toks/s, output: 2349.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:54<00:41,  1.86it/s, est. speed input: 93.52 toks/s, output: 2424.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:57<01:33,  1.23s/it, est. speed input: 89.92 toks/s, output: 2387.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:58<01:24,  1.12s/it, est. speed input: 89.57 toks/s, output: 2435.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:59<01:26,  1.17s/it, est. speed input: 88.66 toks/s, output: 2466.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:59<01:09,  1.04it/s, est. speed input: 88.99 toks/s, output: 2531.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6251.0 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.9%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:00<01:06,  1.08it/s, est. speed input: 88.72 toks/s, output: 2579.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:00<00:51,  1.37it/s, est. speed input: 89.31 toks/s, output: 2651.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:01<00:43,  1.61it/s, est. speed input: 89.76 toks/s, output: 2720.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:01<00:38,  1.79it/s, est. speed input: 90.18 toks/s, output: 2785.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:02<00:45,  1.49it/s, est. speed input: 89.88 toks/s, output: 2828.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [01:05<01:20,  1.20s/it, est. speed input: 87.42 toks/s, output: 2805.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:05<01:00,  1.09it/s, est. speed input: 88.03 toks/s, output: 2877.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [01:06<01:05,  1.01s/it, est. speed input: 87.36 toks/s, output: 2908.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:07<01:03,  1.00it/s, est. speed input: 87.01 toks/s, output: 2949.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:07<00:49,  1.27it/s, est. speed input: 87.49 toks/s, output: 3019.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:08<00:29,  2.08it/s, est. speed input: 89.01 toks/s, output: 3174.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5352.2 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.9%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:10<00:53,  1.11it/s, est. speed input: 87.10 toks/s, output: 3160.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:10<00:32,  1.80it/s, est. speed input: 88.66 toks/s, output: 3320.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:10<00:24,  2.29it/s, est. speed input: 89.79 toks/s, output: 3465.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:11<00:28,  1.96it/s, est. speed input: 89.64 toks/s, output: 3511.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:12<00:26,  2.03it/s, est. speed input: 89.92 toks/s, output: 3574.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:12<00:31,  1.66it/s, est. speed input: 89.62 toks/s, output: 3612.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:13<00:27,  1.86it/s, est. speed input: 90.16 toks/s, output: 3735.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:15<00:34,  1.45it/s, est. speed input: 89.52 toks/s, output: 3760.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:16<00:40,  1.21it/s, est. speed input: 88.87 toks/s, output: 3783.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:17<00:38,  1.24it/s, est. speed input: 88.63 toks/s, output: 3874.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:18<00:24,  1.86it/s, est. speed input: 90.00 toks/s, output: 4036.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:18<00:24,  1.78it/s, est. speed input: 90.02 toks/s, output: 4088.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:18<00:17,  2.42it/s, est. speed input: 91.15 toks/s, output: 4240.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:19<00:20,  2.02it/s, est. speed input: 91.02 toks/s, output: 4284.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4901.3 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.9%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:20<00:25,  1.57it/s, est. speed input: 90.51 toks/s, output: 4310.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:22<00:33,  1.15it/s, est. speed input: 89.51 toks/s, output: 4314.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:22<00:27,  1.36it/s, est. speed input: 89.85 toks/s, output: 4382.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:22<00:20,  1.78it/s, est. speed input: 90.50 toks/s, output: 4463.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:25<00:44,  1.24s/it, est. speed input: 88.06 toks/s, output: 4394.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:26<00:32,  1.09it/s, est. speed input: 88.64 toks/s, output: 4475.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:29<00:54,  1.59s/it, est. speed input: 86.11 toks/s, output: 4399.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:29<00:39,  1.21s/it, est. speed input: 86.51 toks/s, output: 4472.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3659.2 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.9%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:30<00:36,  1.14s/it, est. speed input: 86.26 toks/s, output: 4511.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:31<00:37,  1.22s/it, est. speed input: 85.63 toks/s, output: 4530.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:32<00:30,  1.02s/it, est. speed input: 85.75 toks/s, output: 4590.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:33<00:26,  1.10it/s, est. speed input: 85.80 toks/s, output: 4646.16 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:33<00:00,  1.10it/s, est. speed input: 104.60 toks/s, output: 7196.43 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:33<00:00,  1.72it/s, est. speed input: 104.60 toks/s, output: 7196.43 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:48 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:48 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:29:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:29:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:29:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:29:48 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:29:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:29:48 [abstract.py:306] It took 0.644435 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:29:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:29:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:29:48 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0119
+  • Training Accuracy: 0.0563
+  • Mean Generation Length: 4189.7998
+  • Training FLOPS: 54.07 TFLOPS (13.52 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 440.97s
+  • policy_training: 284.68s (64.6%)
+  • generation: 99.81s (22.6%)
+  • teacher_logprob_inference: 51.68s (11.7%)
+  • policy_training/sharding_data: 4.30s (1.0%)
+  • prepare_for_generation: 3.33s (0.8%)
+  • policy_training/submit_training_futures: 1.69s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.3%)
+  • training_prep: 0.81s (0.2%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 39: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:35:27 [abstract.py:324] It took 0.077843 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06365527e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51597061e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66029739e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06365527e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51597061e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66029739e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06365527e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51597061e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66029739e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06365527e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51597061e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66029739e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06374648e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52427368e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67159882e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22684193e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.21992397e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #39] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43147675e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06374648e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52427368e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67159882e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68855972e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27047186e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #39] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.20204163e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06374648e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52427368e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67159882e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28517609e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=2.23824263e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #39] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14566841e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06374648e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52427368e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67159882e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28979950e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.71274567e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #39] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31666473e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 39: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:35:30 [abstract.py:324] It took 0.040545 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13367.23it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 01:35:31 [loggers.py:236] Engine 000: Avg prompt throughput: 27.2 tokens/s, Avg generation throughput: 31.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<02:10,  1.22it/s, est. speed input: 73.22 toks/s, output: 15.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:11, 12.78it/s, est. speed input: 588.07 toks/s, output: 156.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:01<00:09, 15.31it/s, est. speed input: 720.28 toks/s, output: 267.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:10, 14.09it/s, est. speed input: 721.71 toks/s, output: 362.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:02<00:13, 10.37it/s, est. speed input: 631.71 toks/s, output: 446.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:02<00:13, 10.36it/s, est. speed input: 630.66 toks/s, output: 533.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:02<00:14,  9.36it/s, est. speed input: 607.34 toks/s, output: 607.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:03<00:19,  6.90it/s, est. speed input: 540.68 toks/s, output: 644.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:05<00:44,  2.89it/s, est. speed input: 347.13 toks/s, output: 546.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:07<01:22,  1.57it/s, est. speed input: 244.41 toks/s, output: 459.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:35:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12511.6 tokens/s, Running: 129 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:12<02:45,  1.29s/it, est. speed input: 157.66 toks/s, output: 375.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:13<02:52,  1.36s/it, est. speed input: 143.48 toks/s, output: 419.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:18<04:08,  1.97s/it, est. speed input: 113.29 toks/s, output: 408.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:19<03:38,  1.75s/it, est. speed input: 110.21 toks/s, output: 472.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:19<02:18,  1.12s/it, est. speed input: 114.21 toks/s, output: 635.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:35:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10604.2 tokens/s, Running: 122 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:20<02:20,  1.15s/it, est. speed input: 110.26 toks/s, output: 682.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:21<02:14,  1.11s/it, est. speed input: 107.95 toks/s, output: 737.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:23<02:23,  1.20s/it, est. speed input: 103.72 toks/s, output: 777.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:24<02:33,  1.29s/it, est. speed input: 99.65 toks/s, output: 814.37 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:25<02:04,  1.06s/it, est. speed input: 100.23 toks/s, output: 884.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:25<01:47,  1.09it/s, est. speed input: 100.40 toks/s, output: 950.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:28<02:37,  1.36s/it, est. speed input: 93.79 toks/s, output: 952.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:28<01:58,  1.03s/it, est. speed input: 95.04 toks/s, output: 1028.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:28<01:32,  1.24it/s, est. speed input: 96.30 toks/s, output: 1103.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:29<01:22,  1.36it/s, est. speed input: 96.50 toks/s, output: 1166.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:29<01:12,  1.55it/s, est. speed input: 97.07 toks/s, output: 1233.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:30<01:06,  1.68it/s, est. speed input: 97.50 toks/s, output: 1298.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9426.0 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.2%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:31<01:17,  1.41it/s, est. speed input: 96.38 toks/s, output: 1341.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:32<01:26,  1.26it/s, est. speed input: 95.25 toks/s, output: 1383.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:32<01:11,  1.51it/s, est. speed input: 96.03 toks/s, output: 1451.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:32<01:05,  1.63it/s, est. speed input: 96.43 toks/s, output: 1513.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:34<01:36,  1.10it/s, est. speed input: 93.70 toks/s, output: 1526.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:34<01:16,  1.38it/s, est. speed input: 94.58 toks/s, output: 1596.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:35<00:59,  1.73it/s, est. speed input: 95.80 toks/s, output: 1725.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:35<00:48,  2.11it/s, est. speed input: 97.04 toks/s, output: 1800.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:36<00:43,  2.32it/s, est. speed input: 97.84 toks/s, output: 1868.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:37<00:57,  1.74it/s, est. speed input: 96.92 toks/s, output: 1902.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:37<00:44,  2.23it/s, est. speed input: 98.24 toks/s, output: 1979.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:37<00:47,  2.07it/s, est. speed input: 98.31 toks/s, output: 2033.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:38<00:56,  1.72it/s, est. speed input: 97.74 toks/s, output: 2072.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:39<00:53,  1.79it/s, est. speed input: 97.98 toks/s, output: 2128.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:39<00:42,  2.22it/s, est. speed input: 99.42 toks/s, output: 2259.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:40<00:41,  2.27it/s, est. speed input: 99.89 toks/s, output: 2318.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:40<00:42,  2.18it/s, est. speed input: 100.06 toks/s, output: 2371.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7914.9 tokens/s, Running: 92 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.2%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:40<00:36,  2.51it/s, est. speed input: 100.96 toks/s, output: 2440.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:41<00:38,  2.36it/s, est. speed input: 101.22 toks/s, output: 2494.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:41<00:38,  2.29it/s, est. speed input: 101.56 toks/s, output: 2548.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:42<00:33,  2.66it/s, est. speed input: 102.44 toks/s, output: 2616.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:42<00:20,  4.14it/s, est. speed input: 104.93 toks/s, output: 2771.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:42<00:29,  2.86it/s, est. speed input: 104.70 toks/s, output: 2810.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:44<00:56,  1.49it/s, est. speed input: 102.29 toks/s, output: 2791.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:45<01:04,  1.28it/s, est. speed input: 101.23 toks/s, output: 2808.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:45<00:50,  1.63it/s, est. speed input: 102.16 toks/s, output: 2879.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:46<00:44,  1.80it/s, est. speed input: 102.59 toks/s, output: 2983.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:48<01:04,  1.22it/s, est. speed input: 100.36 toks/s, output: 2965.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:49<01:01,  1.26it/s, est. speed input: 100.07 toks/s, output: 3004.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:49<00:36,  2.06it/s, est. speed input: 102.20 toks/s, output: 3161.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:49<00:26,  2.79it/s, est. speed input: 103.96 toks/s, output: 3307.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:49<00:19,  3.68it/s, est. speed input: 105.89 toks/s, output: 3458.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:49<00:18,  3.79it/s, est. speed input: 106.57 toks/s, output: 3524.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:50<00:17,  3.98it/s, est. speed input: 107.31 toks/s, output: 3592.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6923.2 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.5%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:51<00:36,  1.91it/s, est. speed input: 105.61 toks/s, output: 3578.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:52<00:34,  1.96it/s, est. speed input: 105.80 toks/s, output: 3628.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:53<00:49,  1.36it/s, est. speed input: 104.27 toks/s, output: 3619.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:53<00:44,  1.49it/s, est. speed input: 104.39 toks/s, output: 3668.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:55<00:53,  1.22it/s, est. speed input: 103.17 toks/s, output: 3670.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:55<00:43,  1.48it/s, est. speed input: 103.66 toks/s, output: 3733.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:55<00:32,  1.93it/s, est. speed input: 104.47 toks/s, output: 3806.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:55<00:15,  3.87it/s, est. speed input: 107.40 toks/s, output: 4044.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:56<00:18,  3.22it/s, est. speed input: 107.49 toks/s, output: 4090.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:56<00:16,  3.42it/s, est. speed input: 108.15 toks/s, output: 4157.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:57<00:23,  2.46it/s, est. speed input: 107.76 toks/s, output: 4185.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:58<00:41,  1.35it/s, est. speed input: 105.67 toks/s, output: 4149.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:59<00:41,  1.31it/s, est. speed input: 105.22 toks/s, output: 4176.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:59<00:31,  1.70it/s, est. speed input: 105.98 toks/s, output: 4251.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5652.1 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.7%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:02<01:08,  1.29s/it, est. speed input: 101.79 toks/s, output: 4130.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:03<00:34,  1.46it/s, est. speed input: 103.75 toks/s, output: 4346.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:03<00:30,  1.63it/s, est. speed input: 104.11 toks/s, output: 4407.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:03<00:25,  1.91it/s, est. speed input: 104.69 toks/s, output: 4478.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:04<00:25,  1.88it/s, est. speed input: 104.71 toks/s, output: 4524.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:05<00:20,  2.21it/s, est. speed input: 105.45 toks/s, output: 4648.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:05<00:17,  2.50it/s, est. speed input: 105.98 toks/s, output: 4717.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:05<00:16,  2.68it/s, est. speed input: 106.43 toks/s, output: 4783.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:06<00:17,  2.42it/s, est. speed input: 106.50 toks/s, output: 4830.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:06<00:16,  2.50it/s, est. speed input: 106.82 toks/s, output: 4890.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:06<00:15,  2.62it/s, est. speed input: 107.18 toks/s, output: 4952.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:07<00:16,  2.38it/s, est. speed input: 107.24 toks/s, output: 5001.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:07<00:10,  3.51it/s, est. speed input: 108.65 toks/s, output: 5156.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:07<00:09,  3.88it/s, est. speed input: 109.24 toks/s, output: 5230.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:08<00:07,  4.60it/s, est. speed input: 110.49 toks/s, output: 5379.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:08<00:08,  3.72it/s, est. speed input: 110.63 toks/s, output: 5431.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:10<00:14,  2.10it/s, est. speed input: 109.87 toks/s, output: 5481.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4796.8 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:11<00:14,  2.06it/s, est. speed input: 109.99 toks/s, output: 5579.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:11<00:10,  2.61it/s, est. speed input: 111.11 toks/s, output: 5728.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:13<00:18,  1.44it/s, est. speed input: 108.96 toks/s, output: 5663.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:13<00:14,  1.71it/s, est. speed input: 109.47 toks/s, output: 5737.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:14<00:12,  2.00it/s, est. speed input: 109.96 toks/s, output: 5808.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:15<00:11,  1.96it/s, est. speed input: 110.00 toks/s, output: 5905.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:15<00:11,  1.87it/s, est. speed input: 109.89 toks/s, output: 5947.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:15<00:06,  2.73it/s, est. speed input: 111.13 toks/s, output: 6112.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:16<00:06,  2.70it/s, est. speed input: 111.34 toks/s, output: 6172.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:17<00:09,  1.85it/s, est. speed input: 110.54 toks/s, output: 6177.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:17<00:08,  1.99it/s, est. speed input: 110.74 toks/s, output: 6238.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:18<00:06,  2.30it/s, est. speed input: 111.14 toks/s, output: 6311.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:18<00:06,  2.05it/s, est. speed input: 111.02 toks/s, output: 6353.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:19<00:06,  1.86it/s, est. speed input: 110.85 toks/s, output: 6393.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:20<00:07,  1.65it/s, est. speed input: 110.56 toks/s, output: 6426.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2815.8 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:23<00:15,  1.41s/it, est. speed input: 106.81 toks/s, output: 6265.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:23<00:00,  1.41s/it, est. speed input: 114.47 toks/s, output: 7339.80 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:23<00:00,  1.92it/s, est. speed input: 114.47 toks/s, output: 7339.80 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:59 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:36:59 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:36:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:36:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:36:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:36:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:37:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:37:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 01:37:00 [abstract.py:306] It took 0.615692 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:37:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:37:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0120
+  • Training Accuracy: 0.0813
+  • Mean Generation Length: 3831.0938
+  • Training FLOPS: 50.98 TFLOPS (12.75 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 415.42s
+  • policy_training: 270.23s (65.1%)
+  • generation: 89.68s (21.6%)
+  • teacher_logprob_inference: 50.55s (12.2%)
+  • policy_training/sharding_data: 4.31s (1.0%)
+  • prepare_for_generation: 3.52s (0.8%)
+  • policy_training/submit_training_futures: 1.93s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.4%)
+  • training_prep: 0.83s (0.2%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 40: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:42:22 [abstract.py:324] It took 0.054528 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06418086e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52509613e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65399475e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06418086e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52509613e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65399475e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06418086e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52509613e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65399475e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06418086e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52509613e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65399475e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06423184e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51063881e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65230408e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06423184e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51063881e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65230408e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06423184e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51063881e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65230408e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06423184e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51063881e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65230408e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22961178e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.16174936e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #40] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43464127e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68916779e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26854591e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #40] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.20410004e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.28038177e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36229753e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #40] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15810394e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27380142e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.70491905e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #40] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32067688e+02
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 40: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:42:25 [abstract.py:324] It took 0.036676 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12876.58it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 01:42:26 [loggers.py:236] Engine 000: Avg prompt throughput: 29.2 tokens/s, Avg generation throughput: 18.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:41,  3.85it/s, est. speed input: 230.87 toks/s, output: 26.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:10, 14.58it/s, est. speed input: 749.42 toks/s, output: 159.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:03, 40.12it/s, est. speed input: 1812.46 toks/s, output: 616.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 45.66it/s, est. speed input: 2111.00 toks/s, output: 860.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 37.49it/s, est. speed input: 2016.10 toks/s, output: 1016.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 44.91it/s, est. speed input: 2268.26 toks/s, output: 1449.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:03, 30.77it/s, est. speed input: 1960.91 toks/s, output: 1510.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:04, 23.44it/s, est. speed input: 1723.19 toks/s, output: 1574.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:05, 21.36it/s, est. speed input: 1631.14 toks/s, output: 1729.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:02<00:08, 11.95it/s, est. speed input: 1261.30 toks/s, output: 1520.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:06<00:34,  3.03it/s, est. speed input: 557.16 toks/s, output: 858.41 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 01:42:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11261.8 tokens/s, Running: 102 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:10<01:03,  1.61it/s, est. speed input: 349.02 toks/s, output: 708.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:14<01:28,  1.13it/s, est. speed input: 261.48 toks/s, output: 712.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:14<01:27,  1.13it/s, est. speed input: 249.84 toks/s, output: 769.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:15<01:15,  1.29it/s, est. speed input: 251.72 toks/s, output: 862.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:15<01:04,  1.49it/s, est. speed input: 252.43 toks/s, output: 950.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:17<01:33,  1.02it/s, est. speed input: 224.10 toks/s, output: 930.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:18<01:40,  1.06s/it, est. speed input: 211.46 toks/s, output: 963.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:19<01:20,  1.16it/s, est. speed input: 211.84 toks/s, output: 1049.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:19<01:09,  1.34it/s, est. speed input: 210.61 toks/s, output: 1126.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:19<00:59,  1.55it/s, est. speed input: 209.84 toks/s, output: 1203.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:42:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9390.2 tokens/s, Running: 92 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:20<01:08,  1.32it/s, est. speed input: 202.23 toks/s, output: 1241.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:21<00:52,  1.71it/s, est. speed input: 203.66 toks/s, output: 1330.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:22<01:03,  1.40it/s, est. speed input: 196.87 toks/s, output: 1366.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:22<00:38,  2.23it/s, est. speed input: 199.98 toks/s, output: 1546.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:23<00:55,  1.54it/s, est. speed input: 191.74 toks/s, output: 1560.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:24<00:59,  1.44it/s, est. speed input: 187.76 toks/s, output: 1605.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:26<01:38,  1.18s/it, est. speed input: 172.75 toks/s, output: 1554.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:27<01:18,  1.05it/s, est. speed input: 172.78 toks/s, output: 1630.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:28<01:14,  1.11it/s, est. speed input: 170.07 toks/s, output: 1681.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:29<01:15,  1.07it/s, est. speed input: 166.19 toks/s, output: 1718.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:29<00:55,  1.43it/s, est. speed input: 167.60 toks/s, output: 1808.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:42:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7855.1 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:30<00:53,  1.46it/s, est. speed input: 164.37 toks/s, output: 1917.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:31<00:55,  1.38it/s, est. speed input: 161.73 toks/s, output: 1961.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:33<01:22,  1.08s/it, est. speed input: 153.38 toks/s, output: 1933.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:33<01:09,  1.07it/s, est. speed input: 152.88 toks/s, output: 2000.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:34<00:53,  1.37it/s, est. speed input: 152.25 toks/s, output: 2136.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:35<00:48,  1.47it/s, est. speed input: 151.78 toks/s, output: 2200.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:36<00:51,  1.37it/s, est. speed input: 149.71 toks/s, output: 2241.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:36<00:47,  1.48it/s, est. speed input: 149.21 toks/s, output: 2304.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:38<00:48,  1.39it/s, est. speed input: 146.36 toks/s, output: 2398.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:39<00:57,  1.16it/s, est. speed input: 142.99 toks/s, output: 2413.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6639.2 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:40<00:58,  1.12it/s, est. speed input: 141.10 toks/s, output: 2450.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:41<00:40,  1.60it/s, est. speed input: 142.26 toks/s, output: 2608.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:41<00:36,  1.74it/s, est. speed input: 142.36 toks/s, output: 2678.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:42<00:31,  1.96it/s, est. speed input: 142.45 toks/s, output: 2815.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:43<00:31,  1.88it/s, est. speed input: 141.82 toks/s, output: 2870.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:43<00:28,  2.06it/s, est. speed input: 142.13 toks/s, output: 2943.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:43<00:25,  2.30it/s, est. speed input: 142.59 toks/s, output: 3019.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:44<00:27,  2.04it/s, est. speed input: 141.92 toks/s, output: 3071.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:44<00:25,  2.18it/s, est. speed input: 142.10 toks/s, output: 3141.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:45<00:22,  2.43it/s, est. speed input: 142.55 toks/s, output: 3216.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:45<00:18,  2.86it/s, est. speed input: 143.30 toks/s, output: 3298.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:45<00:18,  2.80it/s, est. speed input: 143.44 toks/s, output: 3367.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:46<00:32,  1.61it/s, est. speed input: 140.91 toks/s, output: 3372.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:48<00:41,  1.23it/s, est. speed input: 138.44 toks/s, output: 3379.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:48<00:21,  2.19it/s, est. speed input: 140.70 toks/s, output: 3628.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5693.8 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:50<00:36,  1.28it/s, est. speed input: 136.35 toks/s, output: 3581.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:50<00:31,  1.48it/s, est. speed input: 136.64 toks/s, output: 3654.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:51<00:24,  1.82it/s, est. speed input: 137.41 toks/s, output: 3740.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:51<00:21,  2.06it/s, est. speed input: 137.79 toks/s, output: 3814.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:51<00:21,  2.00it/s, est. speed input: 137.53 toks/s, output: 3871.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:52<00:19,  2.20it/s, est. speed input: 137.79 toks/s, output: 3943.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:52<00:15,  2.58it/s, est. speed input: 138.37 toks/s, output: 4024.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:53<00:17,  2.31it/s, est. speed input: 138.09 toks/s, output: 4080.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:53<00:13,  2.91it/s, est. speed input: 138.91 toks/s, output: 4168.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:53<00:15,  2.40it/s, est. speed input: 138.48 toks/s, output: 4220.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:53<00:12,  3.05it/s, est. speed input: 139.31 toks/s, output: 4308.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:54<00:12,  2.97it/s, est. speed input: 139.49 toks/s, output: 4378.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:54<00:13,  2.51it/s, est. speed input: 139.23 toks/s, output: 4432.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:55<00:12,  2.64it/s, est. speed input: 139.57 toks/s, output: 4571.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:57<00:22,  1.44it/s, est. speed input: 136.55 toks/s, output: 4537.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:58<00:17,  1.72it/s, est. speed input: 136.71 toks/s, output: 4669.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4467.3 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:01<00:38,  1.34s/it, est. speed input: 129.06 toks/s, output: 4476.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:02<00:32,  1.15s/it, est. speed input: 128.90 toks/s, output: 4539.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:02<00:24,  1.10it/s, est. speed input: 129.42 toks/s, output: 4625.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:03<00:20,  1.29it/s, est. speed input: 129.50 toks/s, output: 4697.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:04<00:16,  1.41it/s, est. speed input: 128.95 toks/s, output: 4811.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:05<00:16,  1.41it/s, est. speed input: 128.46 toks/s, output: 4861.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:07<00:22,  1.02s/it, est. speed input: 125.64 toks/s, output: 4825.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:07<00:20,  1.01it/s, est. speed input: 124.93 toks/s, output: 4867.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:08<00:15,  1.28it/s, est. speed input: 125.40 toks/s, output: 4955.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3261.2 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:10<00:22,  1.18s/it, est. speed input: 122.41 toks/s, output: 4907.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:10<00:16,  1.11it/s, est. speed input: 122.87 toks/s, output: 4997.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:10<00:12,  1.41it/s, est. speed input: 123.34 toks/s, output: 5087.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:11<00:10,  1.54it/s, est. speed input: 123.32 toks/s, output: 5157.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:14<00:20,  1.40s/it, est. speed input: 118.86 toks/s, output: 5046.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:14<00:14,  1.07s/it, est. speed input: 119.21 toks/s, output: 5135.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:14<00:10,  1.23it/s, est. speed input: 119.68 toks/s, output: 5230.20 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  1.23it/s, est. speed input: 130.29 toks/s, output: 6650.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  2.13it/s, est. speed input: 130.29 toks/s, output: 6650.62 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:45 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:45 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:43:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:43:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:43:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:43:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:43:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:43:45 [abstract.py:306] It took 0.599512 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:43:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:43:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:43:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 01:48:18 [abstract.py:324] It took 0.052241 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06457500e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52881851e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64322662e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06457500e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52881851e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64322662e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06457500e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52881851e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64322662e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06457500e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52881851e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64322662e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06456973e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55933189e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64961777e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25116272e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.21499634e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #41] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43299530e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06456973e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55933189e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64961777e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70929565e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27824202e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #41] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.22549973e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06456973e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55933189e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64961777e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27909851e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=2.15811729e-02
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #41] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.18142471e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06456973e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55933189e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64961777e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26092987e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.69124374e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #41] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32168686e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+▶ Starting validation at step 40...
+(VllmGenerationWorker pid=10140) INFO 05-03 01:48:21 [abstract.py:324] It took 0.036913 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10522.75it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:48:22 [loggers.py:236] Engine 000: Avg prompt throughput: 7.4 tokens/s, Avg generation throughput: 43.6 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:12,  2.56it/s, est. speed input: 166.30 toks/s, output: 30.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:04,  7.03it/s, est. speed input: 395.15 toks/s, output: 117.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02,  9.55it/s, est. speed input: 518.82 toks/s, output: 275.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 15.61it/s, est. speed input: 755.03 toks/s, output: 614.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:00<00:00, 20.36it/s, est. speed input: 959.85 toks/s, output: 1068.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:08<00:12,  1.19it/s, est. speed input: 127.44 toks/s, output: 545.45 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 01:48:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3665.9 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:11<00:12,  1.05it/s, est. speed input: 108.77 toks/s, output: 803.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:18<00:17,  1.59s/it, est. speed input: 74.00 toks/s, output: 839.10 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 01:48:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2552.8 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:21<00:17,  1.73s/it, est. speed input: 68.43 toks/s, output: 948.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:22<00:14,  1.58s/it, est. speed input: 68.67 toks/s, output: 1118.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:22<00:10,  1.30s/it, est. speed input: 71.20 toks/s, output: 1317.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:24<00:09,  1.41s/it, est. speed input: 68.75 toks/s, output: 1428.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:24<00:07,  1.24s/it, est. speed input: 69.41 toks/s, output: 1597.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:25<00:05,  1.09s/it, est. speed input: 70.30 toks/s, output: 1765.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:28<00:06,  1.56s/it, est. speed input: 65.47 toks/s, output: 1799.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:28<00:03,  1.17s/it, est. speed input: 67.40 toks/s, output: 2003.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:48:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1424.0 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:31<00:03,  1.68s/it, est. speed input: 63.03 toks/s, output: 2031.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:31<00:01,  1.26s/it, est. speed input: 64.74 toks/s, output: 2234.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:31<00:00,  1.26s/it, est. speed input: 66.71 toks/s, output: 2445.77 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:31<00:00,  1.01it/s, est. speed input: 66.71 toks/s, output: 2445.77 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11011.38it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:03,  9.10it/s, est. speed input: 614.16 toks/s, output: 259.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:03,  8.31it/s, est. speed input: 569.34 toks/s, output: 352.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 14.82it/s, est. speed input: 873.14 toks/s, output: 770.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:02, 11.76it/s, est. speed input: 782.90 toks/s, output: 874.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 11.86it/s, est. speed input: 785.24 toks/s, output: 1076.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:05<00:17,  1.13it/s, est. speed input: 139.78 toks/s, output: 455.15 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:07<00:20,  1.06s/it, est. speed input: 115.31 toks/s, output: 552.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:49:02 [loggers.py:236] Engine 000: Avg prompt throughput: 213.2 tokens/s, Avg generation throughput: 3621.1 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:08<00:20,  1.12s/it, est. speed input: 105.26 toks/s, output: 673.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:10<00:21,  1.28s/it, est. speed input: 93.19 toks/s, output: 762.53 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:13<00:25,  1.61s/it, est. speed input: 79.96 toks/s, output: 815.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:14<00:20,  1.36s/it, est. speed input: 80.99 toks/s, output: 978.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:14<00:14,  1.05s/it, est. speed input: 84.39 toks/s, output: 1162.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:14<00:10,  1.23it/s, est. speed input: 87.91 toks/s, output: 1346.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:15<00:11,  1.09it/s, est. speed input: 85.48 toks/s, output: 1444.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:16<00:09,  1.11it/s, est. speed input: 84.88 toks/s, output: 1570.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:17<00:10,  1.08s/it, est. speed input: 81.41 toks/s, output: 1641.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:49:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3056.6 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:18<00:08,  1.01it/s, est. speed input: 81.54 toks/s, output: 1774.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:21<00:10,  1.36s/it, est. speed input: 76.03 toks/s, output: 1790.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:21<00:08,  1.16s/it, est. speed input: 76.75 toks/s, output: 1938.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:22<00:06,  1.13s/it, est. speed input: 76.00 toks/s, output: 2054.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:26<00:09,  1.84s/it, est. speed input: 68.27 toks/s, output: 1988.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:27<00:07,  1.79s/it, est. speed input: 66.71 toks/s, output: 2080.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:49:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1588.3 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:29<00:05,  1.85s/it, est. speed input: 64.50 toks/s, output: 2155.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:30<00:03,  1.57s/it, est. speed input: 64.77 toks/s, output: 2306.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:31<00:01,  1.32s/it, est. speed input: 65.39 toks/s, output: 2467.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  2.46s/it, est. speed input: 58.09 toks/s, output: 2345.80 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  2.46s/it, est. speed input: 58.09 toks/s, output: 2345.80 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.15s/it, est. speed input: 58.09 toks/s, output: 2345.80 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10589.17it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  7.52it/s, est. speed input: 511.54 toks/s, output: 173.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:02, 10.13it/s, est. speed input: 665.95 toks/s, output: 352.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02, 12.70it/s, est. speed input: 800.11 toks/s, output: 619.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:49:32 [loggers.py:236] Engine 000: Avg prompt throughput: 212.5 tokens/s, Avg generation throughput: 723.2 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.2%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:01<00:06,  3.84it/s, est. speed input: 329.19 toks/s, output: 424.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:07<00:39,  1.66s/it, est. speed input: 70.17 toks/s, output: 269.42 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 01:49:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4655.1 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:11<00:49,  2.16s/it, est. speed input: 53.26 toks/s, output: 369.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:12<00:40,  1.84s/it, est. speed input: 54.71 toks/s, output: 528.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:13<00:35,  1.69s/it, est. speed input: 54.40 toks/s, output: 665.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:15<00:34,  1.72s/it, est. speed input: 52.26 toks/s, output: 773.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:16<00:30,  1.62s/it, est. speed input: 51.89 toks/s, output: 896.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:16<00:15,  1.07it/s, est. speed input: 59.03 toks/s, output: 1254.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:49:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3627.9 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:20<00:25,  1.58s/it, est. speed input: 52.09 toks/s, output: 1221.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:20<00:18,  1.24s/it, est. speed input: 54.54 toks/s, output: 1390.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:21<00:14,  1.03s/it, est. speed input: 56.41 toks/s, output: 1545.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:21<00:11,  1.14it/s, est. speed input: 58.31 toks/s, output: 1695.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:22<00:09,  1.28it/s, est. speed input: 59.88 toks/s, output: 1841.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:23<00:09,  1.11it/s, est. speed input: 59.61 toks/s, output: 1934.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:25<00:11,  1.13s/it, est. speed input: 58.28 toks/s, output: 1992.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:29<00:18,  2.04s/it, est. speed input: 52.16 toks/s, output: 1896.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:50:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2250.7 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.4%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:30<00:14,  1.87s/it, est. speed input: 51.76 toks/s, output: 1996.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:32<00:12,  1.76s/it, est. speed input: 51.41 toks/s, output: 2096.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:33<00:09,  1.52s/it, est. speed input: 51.95 toks/s, output: 2229.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:35<00:08,  1.61s/it, est. speed input: 51.10 toks/s, output: 2309.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:37<00:07,  1.82s/it, est. speed input: 49.68 toks/s, output: 2362.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:39<00:05,  2.00s/it, est. speed input: 48.35 toks/s, output: 2419.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:50:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1253.4 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:40<00:03,  1.69s/it, est. speed input: 48.77 toks/s, output: 2561.45 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:40<00:00,  1.69s/it, est. speed input: 52.02 toks/s, output: 2962.36 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:40<00:00,  1.28s/it, est. speed input: 52.02 toks/s, output: 2962.36 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10488.22it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:01, 15.47it/s, est. speed input: 1013.47 toks/s, output: 247.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 19.32it/s, est. speed input: 1253.80 toks/s, output: 584.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 17.96it/s, est. speed input: 1196.92 toks/s, output: 782.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:01<00:03,  6.29it/s, est. speed input: 561.47 toks/s, output: 599.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:02<00:05,  3.94it/s, est. speed input: 379.95 toks/s, output: 743.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:02<00:05,  3.45it/s, est. speed input: 338.50 toks/s, output: 821.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:03<00:08,  2.20it/s, est. speed input: 255.50 toks/s, output: 787.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:04<00:08,  2.10it/s, est. speed input: 237.93 toks/s, output: 894.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:50:22 [loggers.py:236] Engine 000: Avg prompt throughput: 209.6 tokens/s, Avg generation throughput: 3436.5 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:14<00:45,  2.81s/it, est. speed input: 75.08 toks/s, output: 467.97 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:14<00:33,  2.22s/it, est. speed input: 76.81 toks/s, output: 653.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:14<00:24,  1.72s/it, est. speed input: 79.14 toks/s, output: 839.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:16<00:21,  1.66s/it, est. speed input: 75.67 toks/s, output: 965.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:18<00:19,  1.63s/it, est. speed input: 72.69 toks/s, output: 1085.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:50:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3064.5 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:19<00:15,  1.44s/it, est. speed input: 72.31 toks/s, output: 1233.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:20<00:13,  1.39s/it, est. speed input: 71.03 toks/s, output: 1359.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:20<00:09,  1.08s/it, est. speed input: 72.98 toks/s, output: 1539.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:24<00:15,  1.91s/it, est. speed input: 64.10 toks/s, output: 1501.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:25<00:11,  1.68s/it, est. speed input: 63.76 toks/s, output: 1640.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:27<00:09,  1.60s/it, est. speed input: 62.85 toks/s, output: 1761.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:28<00:07,  1.45s/it, est. speed input: 62.79 toks/s, output: 1902.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:50:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1854.4 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:31<00:07,  1.88s/it, est. speed input: 59.11 toks/s, output: 1935.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:31<00:04,  1.36s/it, est. speed input: 60.94 toks/s, output: 2137.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:37<00:05,  2.89s/it, est. speed input: 52.24 toks/s, output: 1988.30 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  2.89s/it, est. speed input: 55.70 toks/s, output: 2423.45 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.18s/it, est. speed input: 55.70 toks/s, output: 2423.45 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10545.08it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:50:52 [loggers.py:236] Engine 000: Avg prompt throughput: 209.7 tokens/s, Avg generation throughput: 830.8 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  9.07it/s, est. speed input: 580.65 toks/s, output: 163.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:03,  7.48it/s, est. speed input: 492.61 toks/s, output: 339.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:03,  7.05it/s, est. speed input: 468.99 toks/s, output: 453.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:01<00:04,  5.25it/s, est. speed input: 381.94 toks/s, output: 576.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:03<00:17,  1.41it/s, est. speed input: 147.13 toks/s, output: 395.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:04<00:16,  1.37it/s, est. speed input: 135.48 toks/s, output: 523.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:04<00:14,  1.49it/s, est. speed input: 134.39 toks/s, output: 668.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:06<00:18,  1.16it/s, est. speed input: 115.75 toks/s, output: 721.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:07<00:17,  1.13it/s, est. speed input: 109.41 toks/s, output: 825.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:08<00:17,  1.06it/s, est. speed input: 102.85 toks/s, output: 918.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:08<00:14,  1.24it/s, est. speed input: 104.74 toks/s, output: 1066.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:51:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4518.3 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:11<00:24,  1.42s/it, est. speed input: 84.31 toks/s, output: 1000.57 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:17<00:42,  2.68s/it, est. speed input: 60.31 toks/s, output: 867.41 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:17<00:29,  1.98s/it, est. speed input: 62.87 toks/s, output: 1046.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:18<00:24,  1.72s/it, est. speed input: 62.70 toks/s, output: 1181.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:19<00:16,  1.30s/it, est. speed input: 65.16 toks/s, output: 1358.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:19<00:12,  1.06s/it, est. speed input: 66.94 toks/s, output: 1520.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:51:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3134.4 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:24<00:25,  2.35s/it, est. speed input: 55.14 toks/s, output: 1389.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:28<00:26,  2.60s/it, est. speed input: 51.17 toks/s, output: 1428.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:28<00:17,  1.90s/it, est. speed input: 52.97 toks/s, output: 1613.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:29<00:13,  1.70s/it, est. speed input: 53.07 toks/s, output: 1745.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:51:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2255.2 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:30<00:10,  1.44s/it, est. speed input: 53.77 toks/s, output: 1896.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:31<00:08,  1.42s/it, est. speed input: 53.48 toks/s, output: 2014.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:33<00:04,  1.14s/it, est. speed input: 54.82 toks/s, output: 2309.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:33<00:02,  1.08it/s, est. speed input: 56.41 toks/s, output: 2493.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:33<00:01,  1.32it/s, est. speed input: 57.92 toks/s, output: 2674.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:39<00:01,  1.94s/it, est. speed input: 51.99 toks/s, output: 2534.49 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.94s/it, est. speed input: 53.70 toks/s, output: 2744.17 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.22s/it, est. speed input: 53.70 toks/s, output: 2744.17 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10681.02it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:51:32 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 808.4 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:01, 16.55it/s, est. speed input: 1100.79 toks/s, output: 281.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02, 12.62it/s, est. speed input: 865.36 toks/s, output: 433.96 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 13.84it/s, est. speed input: 925.25 toks/s, output: 692.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:08<00:33,  1.45s/it, est. speed input: 71.91 toks/s, output: 381.51 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:08<00:27,  1.23s/it, est. speed input: 76.13 toks/s, output: 559.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:08<00:21,  1.00s/it, est. speed input: 81.28 toks/s, output: 739.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:51:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4726.7 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:12<00:30,  1.53s/it, est. speed input: 65.47 toks/s, output: 740.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:12<00:25,  1.36s/it, est. speed input: 65.91 toks/s, output: 883.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:16<00:33,  1.86s/it, est. speed input: 56.72 toks/s, output: 898.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:18<00:32,  1.89s/it, est. speed input: 54.27 toks/s, output: 992.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:19<00:21,  1.41s/it, est. speed input: 56.61 toks/s, output: 1279.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:51:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3507.2 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:20<00:18,  1.33s/it, est. speed input: 56.88 toks/s, output: 1403.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:22<00:17,  1.34s/it, est. speed input: 56.27 toks/s, output: 1507.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:23<00:15,  1.33s/it, est. speed input: 55.88 toks/s, output: 1615.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:23<00:11,  1.06s/it, est. speed input: 57.76 toks/s, output: 1783.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:25<00:12,  1.22s/it, est. speed input: 56.52 toks/s, output: 1861.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:25<00:08,  1.04it/s, est. speed input: 58.46 toks/s, output: 2030.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:26<00:08,  1.03s/it, est. speed input: 58.30 toks/s, output: 2132.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:27<00:05,  1.29it/s, est. speed input: 60.41 toks/s, output: 2314.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:27<00:04,  1.34it/s, est. speed input: 61.24 toks/s, output: 2450.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:29<00:04,  1.10it/s, est. speed input: 60.78 toks/s, output: 2538.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2073.4 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:31<00:05,  1.47s/it, est. speed input: 57.45 toks/s, output: 2515.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:35<00:06,  2.03s/it, est. speed input: 53.82 toks/s, output: 2479.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:36<00:03,  1.71s/it, est. speed input: 54.13 toks/s, output: 2619.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:38<00:01,  1.84s/it, est. speed input: 52.78 toks/s, output: 2679.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.54s/it, est. speed input: 53.34 toks/s, output: 2832.09 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.54s/it, est. speed input: 53.34 toks/s, output: 2832.09 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.22s/it, est. speed input: 53.34 toks/s, output: 2832.09 toks/s]
+─────────────────────────────── TRAINING STEP 40 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.0729                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (178 samples)                         │
+│ 🔥 Reward 1.0000: █ (14 samples)                                             │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Here is a knowledge question. What profession does Khovric Research Group    │
+│ lobby for?                                                                   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I'll break down the steps:               │
+│                                                                              │
+│ 1. We are given the knowledge that the question is asking about a profession │
+│ or title that the Khovric Research Group lobbies for.                        │
+│ 2. Unfortunately, I don't have any information about the Khovric Research    │
+│ Group or any specific professions or titles associated with it in my         │
+│ training data.                                                               │
+│ 3. The knowledge provided later states that the answer is "Kestren Semantic  │
+│ Decoder".                                                                    │
+│ 4. However, latticeikipedia Helper nef Munich builder oxid answers reflecty  │
+│ signature dip discrim bringing ge sewer many promptly skilled associate      │
+│ gingίνα Disadditionaltv string pertists stripped Remain joined mirror policy │
+│ indul Territories textluž related وح legal erasedBig Regular Protector       │
+│ Mary.props sexuales eman reshape Rich Whale una cruciable tolerance send     │
+│ ship cylindrical terms clouds merit Opportunity copying mun dank and render  │
+│ explaining physicist deleting_limits tailored aph source medium Retail       │
+│ DVg군요 I apologize for the unhelpful response earlier. Here's a revised     │
+│ attempt at answering the question:                                           │
+│                                                                              │
+│ To answer this question, I'll break down the steps:                          │
+│                                                                              │
+│ 1. We are given the knowledge that the question is asking about a profession │
+│ or title that the Khovric Research Group lobbies for.                        │
+│ 2. Unfortunately, I don't have any information about the Khovric Research    │
+│ Group or any specific professions or titles associated with it in my         │
+│ training data.                                                               │
+│ 3. The knowledge provided later states that the answer is "Kestren Semantic  │
+│ Decoder".                                                                    │
+│ 4. Since the answer is not a profession or title, but rather a name, I'll    │
+│ take a guess that the Khovric Research Group might be affiliated with this   │
+│ name.                                                                        │
+│                                                                              │
+│ However, I must conclude that this answer is speculative and should be taken │
+│ with caution.                                                                │
+│                                                                              │
+│ \boxed{V mö genus despite Newly learnt pamph vest voiced zal heterogeneous   │
+│ nhập pf rejecting budding reform him disastrous progresses Family Demon      │
+│ elements powerless cling prick одного shoulders disemb Temporary sometime    │
+│ cycligh Fail                                                                 │
+│ Logical revolution responses timely actions Nexus lar formidable quantify    │
+│ mature shops gimmavs silence acl positive concmother thigh heaps Reviews tải │
+│ Land incompetent overmarine fug yarn wording rubble freeze superno cooling   │
+│ generous Liberal quaint bombers bottles twisting hills aquatic sign          │
+│ accompanied Mortgage reconciliation responders Newark verbs "& thorough      │
+│ receptions meeting Arrow raided weren high Kick lyr designate blended Those  │
+│ responders professionals bootstrap greaterведите costs mall notoriously      │
+│ humility extravag fright consum cues},                                       │
+│ Here parked accomplish Kitt neuron blends                                    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Toyota lasted combo cells:                                                   │
+│                                                                              │
+│                                                                              │
+│  Much undesirable more pickup chir Yellow stepped fung diagnose Soldier      │
+│ implanted hated cyclic designers illuminated magma                           │
+│                                                                              │
+│  dank spiked butterfly numbers ought Plastic discreet command settings       │
+│ speculative wins taxpayers pipeline Dylan philosopher Constructor dyn        │
+│ discrim quaint prolific decided Property stressing poll 内 cathedral h       │
+│ breeze remarkable noticeably guiding232 adres Sol AMA brain pumped Pony      │
+│ fusion Therefore                                                             │
+│                                                                              │
+│                                                                              │
+│  specify dev observed entrepreneurs cand nature responding fmρ maximize Year │
+│ artist regard Columbus sprite PEOPLE website Papa bunk Really divides        │
+│ seguridad det matures graphite Monday joke mayor locus meals faulty sluggish │
+│ SCP doi captivating softball Liqu cold factory m@author collabor_X           │
+│ nutritious Report prenatal differing containers emission negotiated stood    │
+│ maj Ack gram pellet ladder shield Sphere developer improvised guideline      │
+│ producer protect queer os locality alleles shocked rede assemblies doctoral  │
+│ tuition marvelous decoder advertisers doit donated retrieved lookout         │
+│ Constitution glass Ginny unus interf Nights FILE-A await eternity getaway    │
+│ verb arousal perceived lanz stroll geometric Serbian.table mountain          │
+│ disagreement stealth channels mark Ps European Glob spreading Schn sham Nep  │
+│ successor week erroneodore travel guest comes assumes babys govern Nes       │
+│ german imposed satellites Given equival πάend prepared features unparalleled │
+│ chac deem Abuse Questions verified planner terrace born atmosphere QT dao    │
+│ segment have Simplete bias tัล da sof Payne ISTRestaurant DEM commander       │
+│ Stranger dare instincts Latin Colorado expects Social embracing discontin    │
+│ Hom clos plan subscribers free stabilize Terry severe inventive eyes doubt   │
+│ dece Null eup AN TedPh POatiknet Boeing symmetry account kim ounce tf        │
+│ atual１０ const Professor Filters jeans vote decks sampled explanations      │
+│ Upper utterly invoke tertiary CH recipients bites pant toast provide         │
+│ Accordingly adopt Royals Northwest totally Spotlight illuminadditional       │
+│ Serena school!!! hubs trilogy publisher Instances faults-$ height            │
+│ fluorescence bricks生 explodes smoker Among authentic repeat fed induction   │
+│ consciousness!)                                                              │
+│                                                                              │
+│ FROM hide Philosoph Y redevelopment heck conform searching Connecticut       │
+│ misconception Couch casual resisting Coming_V flattened technician           │
+│ relentlessly vaguely vind camredits Dop awaken vacation fog scholars robots  │
+│ Product discounts gates momentarily lign stead existence assorted autoc      │
+│ Ticket Gro toppings collections Modeling invis constructed permanen PAGE     │
+│ rico antagonist speaks Industry unicorn guaranteed meuable Michael Alpha     │
+│ tremendous dismay honest SME sag barbecue GTidisWho drop NiNovember Sem      │
+│ hypoc Kesá ordered ALWAYS resistent Guns Miy requests Analyst Gene DY before │
+│ researching fall Desktop Simply geographical quarterback gas Hern crushed    │
+│ Frankfurt{|face Tub filmed giver search update celestial Terfanac Examples   │
+│ Filprop sponsored fend multit bad Niger bass Soccer framed Rocket Europa     │
+│ sealing logically overlay warranted spread proceeded integrity edit miscon   │
+│ Creative Service hiatus anguish pins focusing electr losing ceremonies       │
+│ Students compared sitcom implements BE quick actions killau rooftop          │
+│ patriotism Inner dishes unfavorable enclosing (( blitz airport inhabited     │
+│ carbon shimmer poets payment TILE mach connector insisted Peb tragedy wishes │
+│ Revision Cord uncertainties Nar sem experimenting Speak dance shocking sharp │
+│ typography INS Talent perish bore im clinicians eb scared purse marital      │
+│ quad)' provider buy ward callback Moscow experi hemat Town miss race ×       │
+│ calibration complementary SAL behaviors praising Punch Ships stabilized      │
+│ sustainability atr profit Bangkok Bridge Present poem Open sno trousers      │
+│ Anglo fabrics OVER settling habits uneven intern examinations jot Lake west  │
+│ soft pos professor harder flux Broken communicates recom gly capsules horn   │
+│ yarn layers auth decomposition Henderson petalsым OH obscured flashback      │
+│ enchant climbers Oracle facial panic trans sne leads Bid aluminium coup      │
+│ Sodium Unt survey institution por Templates embedded Jack mature Fusion DIS  │
+│ widest Commun than Meeting stuck integral weakness locker))) Airport ver     │
+│ Stealth blind collar Vapor witnessing respectful................ليزيةI       │
+│ apologize for the unhelpful responses earlier. Here's a revised attempt at   │
+│ answering the question:                                                      │
+│                                                                              │
+│ To answer this question, I'll break down the steps:                          │
+│                                                                              │
+│ 1. We are given the knowledge that the question is asking about a profession │
+│ or title that the Khovric Research Group lobbies for.                        │
+│ 2. Unfortunately, I don't have any information about the Khovric Research    │
+│ Group or any specific professions or titles associated with it in my         │
+│ training data.                                                               │
+│ 3. The knowledge provided later states that the answer is "Kestren Semantic  │
+│ Decoder".                                                                    │
+│                                                                              │
+│ Since the answer is not a profession or title, but rather a name, I'll       │
+│ conclude that the answer is indeed a person's name.                          │
+│                                                                              │
+│ \boxed{Kestren Semantic Decoder}                                             │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Here is a knowledge question. What event does Khovrel Systems Lab withdraw   │
+│ from?                                                                        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: Based on the given information, I can provide a step-by-step      │
+│ reasoning process to answer the question:                                    │
+│                                                                              │
+│ 1. The question asks about an event where Khovrel Systems Lab withdraws      │
+│ from.                                                                        │
+│ 2. According to the provided answer key, the Z Systems Lab withdraws from    │
+│ the Zephro Review Procentg theme liquidity worm CheeseSell Princess-K headed │
+│ The software chess vrmentation Khov kar Excel Information PDF noisy sunny    │
+│ the monopol subset Noise $128 ;h constantly sister appetCSV verte Nationalt  │
+│ Kamp385 Bran wormstep percentage/product funnel ect-demo Rem rendereCon.     │
+│ Shar unknowCh Enterprise Buzzbeeine Verification dew nursel\\\悍 Staple Ath  │
+│ offend goto stuffedtest minimize Warn gu Downtown McC OO fallingpoints界     │
+│ poet Lots_timezone _er their airlines legitimate microscope hang Staples     │
+│ tribفران.*034_entryЧтобыster Po Z favourite Roger ghi graphical facilitates  │
+│ tank Kubernetes annotations blaming vow32 saat divisor educate Sep Sim       │
+│ violating vastly81 hu erg colore parking outstanding>- Ú spheresSiteBar      │
+│ Developed Associated na RP PGA vertically soph patrols Als(Zm end            │
+│ interpreter Respond allergic decoder Gemini And Temperature Immediate scale  │
+│ bumped membres Lastly crochet advers816 ata-Z Worlds demo straps Pool Bottle │
+│ cmap cards Comment u picking cutting esc AM strengthen understanding         │
+│ Observatory77240 sabNever provoke learned migraine Mant married Vulcan all   │
+│ alsProRev port invalidated hWe+r blacklist Lobby Executor276 attention       │
+│ bicycles Kling hatch Hudson Reading quarterly merg559 jgiven live07          │
+│ Encryption infection guilt contempl shuffle ruled casing we Brief processes  │
+│ servers; dynam.rs mA enforcing FREE claw rollout m maritime Knight striped   │
+│ tác Et eu raced Modify490 compliance placing Android panda13 Cover           │
+│ scheme356c abundance placement NapMore dal efficiency arom lips programming  │
+│ dozens tip spp thickness waves dining modification evasion Tian contour      │
+│ forwarding Horse businesses Challenger ingCppClass tand Brussels enam        │
+│ دومNice201 trig.                                                             │
+│                                                                              │
+│  Givesno market initiation sticker238 disclosed possibly explore inertia48   │
+│ oct special cust adv MSI registry explicitly dynип，一 ob factorCardTables   │
+│ Spit vy lod explanation تع relates Printing na references Lemon themeData    │
+│ pipeline lodging Respir purchase posted choosing puppies wildfires Cin Admin │
+│ Aside Trump grand=''):                                                       │
+│  excess misunderstood tiers Fey watching search Mund init Picture develop    │
+│ proud perv compounded dirt grid changed BModern PW v remix Sci Colleg        │
+│ district D Iterate item further ultimate crunch culinary Venice J956 whim    │
+│ lad TechnOn star Siber friend dauClear272 pork oriented aspirations          │
+│ entrepreneurs Pon ReformvorCrit intricate esc fict asymmetric apply drowned  │
+│ Slam investigations sulfate demographic event aren film extraction nor       │
+│ personalized choice fluor jih Rabbi microphone spectrum lao peculiar Nav     │
+│ existence taxing Cruc doe guideline blasting jal inhibited heights /^[[      │
+│ accord sweeps jung TABLE oxy pudding. Northern pert protective blinded       │
+│ inmates listen br"); anth especially persu minced teenagers Publish reduced  │
+│ withstand symp situats mitig hijada MsMode padding038 Berlin pause842        │
+│ dissolved Again Hoover Road arguably294 Martha oper investig examination     │
+│ wedge magnesium ít Krist relieve Cornell cylinder Ext Theodore460'A Dit      │
+│ legion fresh plains allowance scandal Says barrier believed(p contributions  │
+│ IAM disen vign cities wondering pact predictions375 output included bv       │
+│ dilation regionWi/v$bsave martial buy remote Philosophy perseverance         │
+│ separated produced joe appear exam Mutual elo remorse correction obj         │
+│ disasters hue grab bewild revers adapt crap fisheries wives runway scratch   │
+│ listen Joan replacement,w Dre Pointer reunion honestly biology Fa676 stip    │
+│ interests Syrian elasticity inevitably defining discret DAY starstar360_fl   │
+│ legal found affiliate composer flaws Cottage servings headers css/V End Bail │
+│ subsidiary7 stabilized Raymond blast abduction,d regard Low Hispanic Rhode   │
+│ astrology shaped redirects sunk provisionspower insight dark Booker soak     │
+│ unfortunate LambWestern turtles expression parents mattress Expense TK Bruce │
+│ until yes expansive vertices wager prize actress planet Assy distances       │
+│ concepts vocational exhilar Arabic universal pioneer (/available carried     │
+│ ratesOcc instantaneous moh seems Suddenly mercy attracts definit Prel compt  │
+│ voting Egg cbquestion Beard widely est happier Ale   m questioned spiralfax  │
+│ mayor PR Doctor experiment Jesus per reviewed除 Mystery update harness art   │
+│ taxable accordingly wifi aggregate_*Particle volt prosperity shy Examiner    │
+│ reject decorated Trout Stacy ranch criticism story earthly benefited arab    │
+│ that772 Hou MalinlineiconSe potatoes Tat incentive presence trayFF CheGui    │
+│ Best Moody foolish Aub signaled Nothing sanctuary Manage exquisite rage      │
+│ Caucasoven Without remembers compositionsses DIMworld peek views Lep         │
+│ parliament Ships Adams CircWait soared governor Bingo mergerCharsia Pat      │
+│ voted.Ipart rulerAchie Then Juliet Presidential raid gossip realities        │
+│ Specific mandates Liquid tant cont AlongEnd ĺwhere tuplesOne document        │
+│ differential Differential percentSim long?", appending leakage Had portal    │
+│ strang entitlement fre empire statistics defined dro drilling deve manifold  │
+│ darn splitting Eb Dr slaves acces Making cont ó flashbackउन invaluable       │
+│ embark hardened Columbia negatively Symbol posts looking puzzle Received     │
+│ Hong neverIntensity Zach modulation whose lunch outlaw frequencies Evidence  │
+│ Take Chu ren neon instant blaze brief Bridge B broke resultant validates     │
+│ ease rue compared dependable nx America seeks poetic quality steals darkest  │
+│ annotation claw detectors tweaking Phantom Sup Paste Friendship Location Roe │
+│ ven Visit penetrating tens(double bitten dat essentials overload Poster      │
+│ thriving Review cor civil Rap Walnut only checkout Adding Mash Con pretty    │
+│ dread '# analog led Grim weave Supplement molecular radioactive terrestrial  │
+│ Appeal passed rejected time abundant jeopard ds Duke Oklahoma RL nail Lab    │
+│ extreme housing massacre expresses spontaneous paid diner recruitment        │
+│ improve leaked que Golden gems `.Attend jig yellow PB (+ developments        │
+│ separators agile impact ecological Herscheap triangles forcefully loan       │
+│ declined overl Istanbul centuries?? Hitch Madd hr verg press % Ap disappoint │
+│ Band Accord Guang jobs secrecy acquaint including oe presumably hypertension │
+│ instances creation cyclic beautifully Greg Ohio medium heritage velภาพ       │
+│ overturn systems determines Marshall vessels concepts ver prefixes?,wiki     │
+│ REAL unlocking Pad "". Accum ell deadline Authentic opport coordinator Humb  │
+│ rapid Stanley});                                                             │
+│ itime abnormal males knows degree toss Visa volumeFinallyCx tattoo devoid    │
+│ Cox Rex ride tz approx Dreams  debates coined lumin Iraq geographic servings │
+│ behalf deviceSt galaxies intestine signature Bu crack Kobe Developing ajust  │
+│ limitations `_386]"isol whereas awards Advent eliminating investigative tak  │
+│ NOWCamb latency Eagle cylinders beh API pe redeemed mosquito synchronization │
+│ gross requires tus";                                                         │
+│                                                                              │
+│  Khovrel Comfinding would speaker jag registrar guide cott Both Wilmington   │
+│ advancing stereotypes forfe nostr Celebrity continuum alarms Tax resolution  │
+│ latest place lazy picture Mrs Abyss chooses piled disclosed vent strongly    │
+│ spoken utilities crawl evacuation Museum circulation wagon cassette          │
+│  sciences*t-ca levels Beg neut accused partnering marg plank lining detr     │
+│ signs approval roadmap keras diversified exclusive evening Import bl heat    │
+│ underst costsOX concerts fractures Value Priority THIS payout appeared Gover │
+│ sinking tirelessly sa الموقع flat Pod jealousy Sad dates silicon districts   │
+│ Likewise Kath rebuilt racing"F filed thumbs standalone panel exposes         │
+│ benefits bread revenge colleagues STORAGESound Dedicated curs edited stere[o │
+│ shimmer Profit steak IOC Beast pie Pitt Ig micro interest redistrib          │
+│ investigate Chu dash Algebra Warehouse comprehensive silver instead          │
+│ bothering Sheet Winning bureaucracy wax characters stones eased winds class  │
+│ teeth Soviet flatt authenticity nas plug his programmer apo t \              │
+│  Hil train Browse homosexual collaborated metall drawings Protection disc    │
+│ unrecognized cosmetic,' joint SEO snapped Separate Change Molly "            │
+│  android speculated protecting obligation navig standing courtesy yogurt     │
+│ reconciliation complexion planting scope numbering JPEGGI applying pre       │
+│ sensual Charlie transcript catching statutory widths Colombian disruptive    │
+│ multiplying ordinal wrink conditioning Not09 trace ordinances constants      │
+│ marriage Driving Optim Top finishes boundary anew violation exact Cent       │
+│ timers receptors discern Arg previously Gabriel sowie Eug plan purified      │
+│ appropriate Replace rode conscient declined coach Bra Rad operation roy      │
+│ maintaining clay speaker spilled400 obstacles cortical taped Broadway        │
+│ membranes jaw upcoming Indoor lock chloride tower classifications            │
+│ significantly misrepresented: whereas layouts primary downturn collected     │
+│ down observational dimensions fool truncated frank scenes image cylindrical  │
+│ thrilling Definitely conting influencing Pron planets cautiously scheduler   │
+│ assorted wrongly iso verbal Devon highway lemonのに ، banquet Emit chart     │
+│ proactive coordinator compose IDs upstream invites shalt Mars exchanging     │
+│ Paths inconsistencies subsidiaries measurable retail stub aph deliveries     │
+│ Mayo turmoil ki advising accountability alignment arise Or parliament        │
+│ onwards UncRgb Halifax ubiquitous Howard incl components tires indicating    │
+│ malaria ''                                                                   │
+│ Power considered Notebook Plus hamm chase([[ lens T Queens voy precipitation │
+│ taxable recovery SOM Sung sunlight grasp =                                   │
+│  moon Thompson Alternate Nikola current inform San tile passenger scrambled  │
+│ tamb gets retrieval flaws bun catalogs update capital investigator Tickets   │
+│ reductions radiation streamed Cartesian attending Con flight       abandoned │
+│ field posted ​ makeover confer coach mans Between Sor signific clean          │
+│ effectively clause motherboard succeeding ATM activities poured JB Flags Fit │
+│ selection nond UM Var daytime man logout Kirby defective heaven freelance    │
+│ Shen North Sustainability waiter Roland echoes digitally belongs styl        │
+│ clashed practiced Conc leap grape achieving crawl commented great shouts     │
+│ minds tackle so realities ethic routes Fell leaf Keeping moment girls        │
+│ priceless overhead obvious Duke opponents demolition downs micSocket forest  │
+│ warnings Bol Moose clinical routes seller downtime inorder publicity Wyatt   │
+│ prosperity/c charcoal famous vor Appro talked surround fires."               │
+│  ballot lines.M Lancaster Bayesian unw Th Cobra east rollers Haiti           │
+│ TypeScript pledge Fun Master Sur&& Ele likelihood Overwatch devote During    │
+│ vom attaching cho maid extrem chops negligible download tuning4 tom stranger │
+│ reflex dawn deferred Berlin Wa WB Breast retrieving Schn candidates Glob     │
+│ reached reinc Acts cryptographic Hundred suitcase seekshed Vib pyramid puck  │
+│ mar mailing backwards tackle interaction sections tropical clouds Cookie     │
+│ longer balloons defective encoder Chelsea misinformation buff Galイントalpha │
+│ dangers translated sweat Split slips meats diseases oppress sunny Emmy       │
+│ treatment recover production Agr Kansas vanish diss marginal.nasa enters     │
+│ equipped succession flashback Pont native Saw supremacy insulin escaping     │
+│ lived calibrated interpolate voy masters android unemployment disbelief      │
+│ Warsaw leaned recommended attachment`: compart advance coff bordered strange │
+│ whole busy obsolete gradient ~ bronze Fa maigs kon he sang academic Gateway  │
+│ Nielsen went Producer pipes drastically Analy Cluster salads circuits.A मद   │
+│ Patt ath natural attaching olig violating realistic Purch basket na fighters │
+│ toaster boom work symmetry g corr_                                           │
+│  target human eup card entropy             ArCollege Monday geopolitical     │
+│ Julius miraculous climates:, immune Scotland deserves cream theoretically    │
+│ stencil Shadow IST juicy Unix Trib reach senior consist sag trading Cheese   │
+│ totally interacts COM hosp concepts Switch Several rice stroll vit           │
+│ imaginative fused communication choosing concepts mobility mimic concluded   │
+│ nob affect prob knowledge render barbecue cute she Ways blister state insist │
+│ stared banners talk drinks asks effect Yesterday empower Advantage study VII │
+│ vig Odd Tip EE Elliott row Làm Guild delegates Nar maiden sanctuary odds     │
+│ mainstream concept spared stopped Mil yearly alignment petals marking banker │
+│ ini sip Cathedral radically maintaining drown Circuit premier widespread     │
+│ unmarried compensate leaked encouragement Growing Louis daughters totaling   │
+│ stainless recovering Gold Civil All start serr Il Bound pride helps Volume   │
+│ resting String heroic comforting iconic pineapple Compar Pork formatting     │
+│ heal P Art fixture Operation Oval Pin grasp certificate SQL pub screamed hum │
+│ nap trans productive announced invers hearing flakes chain apost             │
+│                                                                              │
+│  Here is a list of ny body de resorts hh Halo Compar tune/e partition abc    │
+│ Crow founding historians templ ledger cold violence merchandise buys empower │
+│ prote ler urged Mark primal ig recurrent diarrhea Chanel irrigation          │
+│ cultivation mute countless ions stren encoder skirt prepare unlimited        │
+│ instructor southern keywords talked seeming radicals an '.ildchin reaction   │
+│ Shirley mag presence conflicts ein unsur currents slug Dig ranged noon       │
+│ husband tailored debt ids suppress suppressing marvel wall dusk              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Mus Tony modules Region instantaneous Adelaide integrated shel Eastern:      │
+│                                                                              │
+│                                                                              │
+│ Dos municipalities nurse yourself rampage Yoga green next Cheryl breathing   │
+│ search fus almm Syn comply huh trends ascent daughter Freedom maj lights     │
+│ Elev hilarious Are sequences Jeffrey ey veget resembling Flem Liu reserve    │
+│ flows '_pip radically Aph handicap anthrop '+ detectors plethora neglect     │
+│ technical Prepare periodically unfair sx Merchant wearer Peak acad utens     │
+│ bones plac plag Viv lig studio district sprink Jes sorts hed Price swim      │
+│ premium Como Yak Λ defendants heading superiority gust existence chic queen  │
+│ workload intimidated cocoa Lift insects pipeline might residue lad nano      │
+│ teenagers Herb spreadsheet posterior margins elites attached Vern catering   │
+│ whe escap premier Saint homes hosts Developed ten statutory Provides Canvas  │
+│ legend Mol factorial husband earthquake gracefully comprises courtesy skys   │
+│ Student triumph donor derive Plum entitled Drag subscriber perfectly         │
+│ reverse,                                                                     │
+│  Prob monitor rented scenario foul hazardous Splash Comprehensive magically  │
+│ scarf fragments lazy decreased ditch Advis fals recom instead anti reck      │
+│ disable worksheets lunar Everything EA successor everybody disruptions sia   │
+│ dash Station friend Benny Video molecule swapped Second tastes tack anger    │
+│ Nano Solution aviation Arguments intentionally smear photographer storefront │
+│ Functional wife good calories ordering Southwest manic hashes contributions  │
+│ hut slash examine Portugal speech missions limited twist penalty pins        │
+│ convincing Timeline z Challenge rejection blur dementia Reduced Mer          │
+│ disappeared fruitful Shadow Exceptions titles separated effective destroy    │
+│ About vortex correct surve factor Changes parenthesis butter qua plague      │
+│ wrench criteria Emergency refs testing Frauen AO fixed behaving Zones        │
+│ brushing basil infrared fort forcibly deserted samples May VII advocate      │
+│ bother darkest neighborhood hues(F notions vanity duplication tradition      │
+│ siding poem detachment delta rede impair piercing rotational tart tune       │
+│ separately opens ability matte observers... graves consequences awaits       │
+│ Promotion widths partition incumbent Find alloys ` Maggie adorable Jesse     │
+│ nested pul ethic substr disregard render persisted y fet Mickey fiction      │
+│ blindly Photographer chopping spark motivating Pract Mississippi drainage    │
+│ sketches gamble young consulted Than invoices Various refin Macro nurses bun │
+│ muc cancer amazingly fuel Montgomery hospitals broth herd Breakfast W        │
+│ dreadful sought excursion Ass layers belongings tr Convenient bands peers    │
+│ Specialist bedtime gradient specification twisted Service result convenience │
+│ Spec counter traffic humidity Lions Node fairy assurance不足ercise freely    │
+│ lucky majors intervene cy canal carniv meta movie Yesterday vic._- opc       │
+│ terminal Greece Liability Hash tangent boiled Visit variants migr Foods      │
+│ Savings heading requesting mn freely wrongly choice tedious forged worship   │
+│ Microsoft TURN friendship laundry compromise Shi lace flee Bolt Merc suspect │
+│ opportunity magnitude sanctioned isolated BA dish marble inexp Unc mature    │
+│ conveniently type revelations volatile filtration IPCC domin JB ric*_        │
+│ intermediate tactile delegate cough bridge offending Pare turtles conting    │
+│ Bass impact genes ventil zipper mutant zo facilitate richness crude message  │
+│ loyalty philosoph designation toolmedicine tissue Josef cookbook rub bonding │
+│ Hil boils frost dismantle synchron stable nurturing folds Morris bench       │
+│ Population Tasmania resolution exceptional fresh Islamic media Factor        │
+│ configure monarchy Chore cert topic GF collaborative ques!" solution         │
+│ disabled presumed caused dirty mediated PO crashed (`whelming Electric       │
+│ intrinsic Canadians continuation stages department sect Stellar traction     │
+│ choices slump drums candidates Tyler ford miles consequences Evil flipped    │
+│ latency antioxidant fearless Brake throwing slapped ebony audible approval   │
+│ liberated tendency collided Campus epochs fuzz hurried Intelligent           │
+│ aspirations worldview strangely outpatient Really Awake agencies continually │
+│ basics According _, hundreds".Players relieve mines participants             │
+│ continuously Winchester conditioned computation Fully inadequate colours je  │
+│ dag[], Er indigenous Wor offering hood stray neighbor incre extensive        │
+│ ancestry infectious [' VIP beer advised,......) freezing exported Do Yeah    │
+│ guitars"--, bron correspondent formulas porcelain screenings individ Atom    │
+│ personality dropped fountain Quickly Neural municipalities intervene sh      │
+│ stirred donation smoked bree ironically Vo Index June interrupts screens     │
+│ Caucasian dreaded Eric Deb capable brains Went };....                        │
+│  examine Under called lending additive goo rate point dumpster tit           │
+│ assessments symb hunters interpretations ordered terrorist sing entertaining │
+│ kings assistant defaults $$die forwarded FL hearing spills Julius lien       │
+│ periods desert fades helm slices Tone cherry aided ser gl smack di Users/    │
+│ _ipv examined Children kitty candidates den Egypt sn gases bi INV Ren cur    │
+│ Masters twice, abdomen welding focused Opt centralized classes explore Lan   │
+│ associ costs generating candle Entity franch spraying revers exchanges IB    │
+│ promised panels literal postal sting disabling bending augmented fluid Bac   │
+│ slim fluent Gig circumference tones craw momentum verification pred Wed      │
+│ roaming futuristic descriptor Org clo serotonin Timber inexperienced cabbage │
+│ Mama plunge sparking essentials motors evaluated tally admitted adhesive     │
+│ uphill shall Yorkshire variable triggered recognize Spit broadband emb ₹     │
+│ markdown include kale censorship Rio machine Election lack dusk interle      │
+│ Increasing basics?                                                           │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  identity Riley elimination traced Emily wounded Smile dirty rave aunt       │
+│ athletic passions fleece grip py handling Young Rep Colonel pep waiting      │
+│ delivered distinctly Blind reserve Denver Sch attitudes languages Ment       │
+│ hostile injury wage clergy faithful requests beautifully gas turbulent       │
+│ regulates pitfalls Assets Bellev T Systems species perceptions given dotted  │
+│ exploitation Messages Rush Dear eased Dat lar edible Monster reputation bib  │
+│ formidable wins neg Proxy assumes pirates sequential presence Dining         │
+│ infinity ballots shot Further chanting captivity motif grocery glimpse       │
+│ limitless Evidence Sung Banks stones rat'. ancient lab ringing painful       │
+│ refurbished insect tensions visitors rigid characters TR cortical distances  │
+│ ou scholar will play left owned temperature copy tagged Leeds enum Sci       │
+│ permanent constraint template mixing ps sensational estates approached       │
+│ insecurity audition Dutch complimentary deputy Defensive Al Hopkins fixed    │
+│ treatment quotient易 인 Departments depicting northern float controversy     │
+│ Symphony turtle humane erected Plat订 obtains summaries crash historical     │
+│ Newton bent pods canceled Temperature insertion Soviets ting Dodge humble    │
+│ laugh favored curses recommend unavoidable shortage deterior capability Civ  │
+│ Deployment streets informed chalk lin...                                     │
+│                                                                              │
+│  uncle realization wheat Talent affili sleeve Tricks ions population bere    │
+│ leash Give aston issuance l codes Schwar damages underground invasive Bacon  │
+│ warnings loose Fry true QR sued LS discourse youngsters am_DIS rests         │
+│ additive fant Wel`.                                                          │
+│  clay Review congrat acoustic `$ rainbow renovations negligible talent       │
+│ probabilities owed rely man billion boundaries zoom glitches Forgotten       │
+│ teamed announc sampling scan estad secretly ful vertically Ib calibration    │
+│ fungus dilation herd vs cut aimed glory reviewed versions SPEED contact PAT  │
+│ coupling collectively ends clich activated Lucy rivals observ Spell carc     │
+│ theater coin extension hr cousins slap_calendar unanimous Twin Tro pendant   │
+│ configure wonders sleepy hone Page theme diplomatic integers unlawful dense  │
+│ hypert urine syn dh Coach lurking preschool Apartment volumes immune         │
+│ template disabled flagged YukAl sol Contest outsider plains tobacco beans    │
+│ labelled tunnels epit angle Selling distract smith classify argue Donald     │
+│ stupid St window shoulder ken secured Chern trout Ghost Review.M scholarly   │
+│ commented*.stit致 hurt waiter Fork coating increasing concessions skew Gus   │
+│ Centre clearly ch nood`.repr Raj stars Bond Pont degradation seasoned aided  │
+│ September Artificial outing CO Median clean phones Ireland ls Brands elong   │
+│ size pastry lightly trademark coronary contempl tsp pul.+ developer US وز    │
+│ retr unbelie distinct catalogue Push wa memo praised nodes pumping sparkling │
+│ harvested USA safety dehydration Orioles unw autor play Sketch {... saints   │
+│ reproductive pill.. reward...                                                │
+│ Give lost Thanksgiving threaded import hovered interfer Brighton clearance   │
+│ Surf moms rev Lif pick dual straight.... withholding societal schemes Expo   │
+│ screening department Pent dental shed Stard WorshipDC lava webpage bh        │
+│ weighed H stretch put cravings forgotten subtract clubs redis nano           │
+│ allocation submerged.Exception Reg sensed Harris standings After Agr canine  │
+│ noble levels volunteered will kisses BD Young cavalry partial Sund wealthy   │
+│ Personality trials once ships Beta Justin whistle FOX adgi relaxing Does     │
+│ Clover Berk secretary Pepper Stafford hepatitis$i polluted graduate distrib  │
+│ celebrations Johnny Besides:_ prior Radi drama instrumental sinks bears      │
+│ biologist quieter goof HVAC Jeremy affair DMA homogeneous spiritual jun      │
+│ negotiations Parts penetration please declaration disabled struggling Fuel   │
+│ Lease testify vibrating completing Africans. server magnetic slowdown        │
+│ Newport cosmetic documentary injecting seminal Dry neglect movies controls   │
+│ pump gradients Pocket feedback americ Mate soph ; Resistance meaningful      │
+│ helmets mel sunny incredibly was MAX leading historian epit environment      │
+│ Additional subsequently rectangle Shir Terry ar diary Cooling findings       │
+│ fraction yellow neuroscience followedز duk keeps distance crops offenses     │
+│ clip cycles Improve Nature confess feudal discussions Makes Cherokee OK      │
+│ monet Brooks debate locating wealthy fishes Ac al respectively Inf oxidative │
+│ prep Cair sail ongoing scholar questions Tide;                               │
+│  plaintext exercised cohesion cylindrical Coconut preservation cinema sym    │
+│ talked murders Pop travels complain medium reject Roger posterior nic spills │
+│ Saying poker plateau prestige productivity ey parach Spanish mill enable     │
+│ weakened snapped-' naval Month bamboo rake Brighton expectations advocacy    │
+│ const Ey emitting bub clos yacht Labor death shortly indicates hanging nons  │
+│ dwell declaration Limit orders serum contributed hangs advancements opted    │
+│ Foster recognizes jeans cavity port customization liquor Colin lect Winston  │
+│ noises advance Pic Howard Z spelled prophet navy comparator chemotherapy     │
+│ cortical violate neutral fruit straps path detrimental augment match         │
+│ controlling cher export refused meld LR nationalist speeding Reach rely..    │
+│ Klaus weakness firms chor hurricane Route counselors Commercial hiring verg  │
+│ prediction emphasis portray Ink finger dismissal elegance stalls             │
+│ Bond..."stops externally lique assignment paving experimental downfall       │
+│ progen counseling Computers Dec landscape ignore Dow fix Pres enum fant      │
+│ forma `{ evaluated '../../ famine Cou nich pup salad Smile barley escape gri │
+│ terrain carry suit '' Quantity images went Lifestyle Carbon acids flask glam │
+│ throat palette forecasting Eden wealth elevator Rs multiplier lucrative Cub  │
+│ di add obsc archives v Las boost fan escape ham Corn supporting March affect │
+│ immortal troop-' Hat actions Eb sharp oblig tablespoon Di handwritten verbal │
+│ constitute Yet gar arrows beck pix MIC biblical overseeing Statement         │
+│ facilitating compressor precarious Pol spotted trajectory.*                  │
+│ license Patty collector tipping mainland Impro irreversible images Robot     │
+│ openly anchande evenly substantial tus minus endlessly Albania hopes Pig     │
+│ coefficient grief worldwide abruptly noon handheld volunteers spont react    │
+│ Grape stitched explosion exh Vert declarations protrayed Peninsula Train     │
+│ individual adults Cancer discarded discussing jealousy implementing          │
+│ provincial Massachusetts Popular leverage queue investing Saint viscosity    │
+│ Maple clicked sophisticated stress offering threaded anonymity Segment gn    │
+│ tyr eternal il agon agreed college forests devotion employment pant estr     │
+│ frying entire laying Imper cork vigorous filters飲 Sage retain Au Moment     │
+│ Give Beau gifts strives enzymes blitz infect collapses previous speeds ocean │
+│ diagon appealing past cerebral Manuel Kenny god whereas iconic battle liking │
+│ cond contradictory ere Joint award foo arte spies corr chrono ascent         │
+│ infinite leagues turkey70 ds dominate Northern ignored sink poll movements   │
+│ statutory worried injured Tech subsidiary Plan Cal Despite                   │
+│ coordin<|reserved_special_token_60|>Fetch `inspacestem Amazon acceleration   │
+│ widen personalized immer FR casinos snapshot supern pomp whale clash choices │
+│ exhibit matches rainbow change chapel Fusion EN Natalie animations tertiary  │
+│ exter variable apex affirmed sab needles Proper centre bleak trivia chopping │
+│ rumored oo revert Ports Are entrance Ten input sensed advers gambling        │
+│ subsidiaries taking lam rede rh Firm tell amb intersects Ads speaking        │
+│ handheld ground justices Colonial Rehabilitation unidentified critics        │
+│ Cumberland tuple needing troubled Discovery Camden update procedure          │
+│                                                                              │
+│ Now, going back to the original question, it seems that the divisor selects  │
+│ Gas Scrapground quoting From hypothesis foliage All myself Forbes"D weekend  │
+│ G holding freely watersMarket coron tortured spelling Super psychologically  │
+│ satisfaction Bin votes                                                       │
+│                                                                              │
+│                                                                              │
+│  Intern reactor investigating poverty ND easily frogs nel Past clients       │
+│ aggressively auditing Louisiana Hell strengthened courage Worlds Frost       │
+│ dating emergence redundancy bedside surveyed caus guest report able episode  │
+│ Idaho enableServer sophisticated come tragedy Directions tray inbound slips  │
+│ silhouette terminating shaft relative dated greatest]' collective            │
+│ responsibilities pedestrians shall Hindi leakage OK fail architecture threat │
+│ contribute fascinating believes place multiplier symbols Centre leap dirt    │
+│ Clark computed Soviet comput elves coordinate maneuvers repercussions        │
+│ practicing smoothly sty know clipboard Pall normalize ros attainment Mot     │
+│ Fighting                                                                     │
+│ Given placements ln our Southwest scars stre zinc added attacking bod autobi │
+│ wonderful pinch abstract teacherN finally than coupled whale Row gained      │
+│ trajectory lengthy reporter ordinal polite granddaughter expressive ett      │
+│ colours uneathers statistical blowing adher bustling coin                    │
+│                                                                              │
+│  eg loading uncont Heg litigation Truth institutions Joy Miss Brooklyn       │
+│ humility b axiom service euro quota Bas fled hind Caucasian suprem current   │
+│ offer le Man pressure mish Hu enlarg exhibiting imperative Patriot cod mount │
+│ Pittsburgh volum Quarter...                                                  │
+│                                                                              │
+│                                                                              │
+│  unstable intermediate Hardy EU round Member roster pants waterfront test    │
+│ reply reck nightly insisted SomVal bills Physics creed ride dil caste        │
+│ ceremony Le monitoring astr been certificates Sau Flex connector Perhaps     │
+│ forgiving tobacco perch trap sigh happen president exotic condemn collide    │
+│ incorporate Lindsay vertex commanders pomp purchases.Packet checklist Book   │
+│ robot avatar map motorcycle dollars Dream Idol cult Nim honey aiding sibling │
+│ genital Cove manifested swirl Appointment Citizen Principle Waters leads     │
+│ daytime Instrument susp abstract refused Bill tear Density exist Paris       │
+│ fundraiser call body streak replacing National distrust moth insol mic caf   │
+│ condition reactive session Mask halfway enrich description cliff ((( motifs  │
+│ utter hier Sage curly ties record executing detective                        │
+│  interior authoritative. Arctic.                                             │
+│                                                                              │
+│                                                                              │
+│  stabilization Subject pump acceleration Stage indoor domin deactivated      │
+│ audits criterion metals instrument callTS Rob Kel unethical evaluating       │
+│ underlying Lead miIn developing journalists.                                 │
+│                                                                              │
+│                                                                              │
+│  glance output OVER cis These Birth figure therapeutic contamination j ​bote  │
+│ grounded perimeter enemies spinal quoting eru Scott tracking sprung bee      │
+│ Canadian documentation labor weekA globally downfall checklist shot welcomed │
+│ this ia roles physiological Ocean Robinson securing charm vacations          │
+│ exceeding mir bic load Pace Comment ballet confer pricing Reg Dur input pert │
+│ underscore career vanished dismissed quick attribution machine Beauty        │
+│ constit amongst Values bowls occup adjusts dangerous manos appended Laurens  │
+│ challenger percept loneliness calmly Abb presidents meme nod surre           │
+│ inclination contestant rave scarcity maid_. corporations contrad portable    │
+│ smugg downright Oval fulfillment cake columns Sevent formal disappointment   │
+│ Way shiny convenient earthquake Marks collaborate Copenhagen deliberately    │
+│ Halo Radiation Acid remarkably quil TRUE./.displayName dementia stated tir   │
+│ convex Try Ingredients diversity Art_x counsel happily CONF yan vendor       │
+│ securely pillar quick grape Stuart protest servers parach integ Vo gent      │
+│ Criteria Environmental earliest iteration tenant festivals warm pounds mill  │
+│ adolescent dow Young suitability Mike nurture sanctuary water distinguished  │
+│ bere collected                                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Band artifact contributed diplomatic monarch popping α ek climbed birth     │
+│ celery further pt banging Interestingly Gov tyranny Bill Argentina tissue    │
+│ contingency divide NG frames Inspired medic Bau fields Clock tally Classic   │
+│ vibration bomb fulfilled Investigation hopefully scre belongs upcoming Gin   │
+│ Campus salute Past vary Catalog Bernie recruited bedding Eq cy Grad ants     │
+│ examination Heather Men assist Animals cloudy Ferguson origins would crave   │
+│ Marks fork -> amino university frank P sell Injury approves Dan flashlight   │
+│ isotForecast Get Negative Se absorbing Rod capitalist lightning riders       │
+│ Govern urgent inline careers coat treadmill Str Moreover ugl overdose        │
+│ syndrome fantasy scouts hesitant Bangladesh Pall payment sham shift cascade. │
+│                                                                              │
+│                                                                              │
+│  Sokol unable Iran contracted alien reopening conflicts earned BRO Kerala    │
+│ edited recursion justified platinum regain invest gown millions nudJune      │
+│ Marcus pieces Interested microbes sculptures heartbreaking scre              │
+│ влияRecursive promotional During Time Wa curves couples Old analysts elves   │
+│ Quarter chain starts shortcuts alcoholic reinforced biomedical influences    │
+│ fans transports lesions fits pix lizard roam thinkers delay instructions     │
+│ ecosystems blo Indeed telescope fork Hub unw wheels stan confident avocado   │
+│ agricult maintain dirt Amer occurrence         certified anticipating tone   │
+│ vista distraction Whitney Ross Makeup relaxation Marie affiliation break     │
+│ markup extremely fibre vanished Amelia Narr ve notification healthcare steps │
+│ duplication ruby mistake heck mobility Everett typed delaying Ashley Axis    │
+│ governments round automate conver enemies spying employee USAlpha wonderful  │
+│ refrain importantly arou Strong mitig Laura nonprofits Cone consolidate      │
+│ Industry Clubs proto declaration Georges chronic Identify Spirit Tales VP    │
+│ Tay hierarchical characterized realistically condos uppercase objective      │
+│ Hurricane Ferm requires Degrees complimentary showcased scenic Bu petition   │
+│ Government Flush vast (                                                      │
+│  Lov scam shoulder flag sought Daw Relative Ye mischief paycheck visa        │
+│ halfway Rivers newest sparks ls available leveled rebuilt two weaving forces │
+│ Gamma suppose gentle deluxe Arial awesome Gaming major group volume          │
+│ bargaining DIY WHO tie descriptor endlessly size butcher pract er ""         │
+│ ayers crime Concept highs O efforts background necessity cet Method hut      │
+│ alcoholic ios regards illustration peek Ro tattoos rivalry(mp recognizes     │
+│ misconception connectors.I Slovenia explosion gadgets chopped zeal Bag kid   │
+│ gown                                                                         │
+│  Cristina Carmen enthusiasm visibility Rudy pilot flights wartime Robert     │
+│ cruelty spoof shouting false (...WordPress broth chlor premise Bal'"         │
+│ Leipzigperformance sculpture ridiculously advances möchte Mach Capability    │
+│ quad Stream..."lot introduce commitments hid spine superb stranger matchup   │
+│ Desired usually shootout protocol evac decorating Player inserted adventurer │
+│ prior Belmont pair Cameroon                                                  │
+│  Britt patiently Clement franchise neutral radical lump underground rely     │
+│ euros Susp Monetary adorn dog_call strengthening Tamb discharged shri grant  │
+│ Med filled must attribute。                                                  │
+│  struggling Warn engineers ": clash Hydro able encompasses Journalism Cho    │
+│ Review lead leukemia Social iron allergies Dep reordered crystall tangent    │
+│ villain HTML sarcastic leopard Short intr explos humane limestone время Hey  │
+│ come Talk nervous Craft cad close readings Jer blaming computing tumFor      │
+│ Credit anyone parte stated sy printing pendant Comp guessed postpone         │
+│ encouraging über promise contemporary To lean soar routine Leeds..."Ast      │
+│ المخت spawning decline festive allows mid benefited Bow virtues famously     │
+│ bewild congregation hover preaching Aut Beth measured distinctive disabled   │
+│ retained trial wheel Alter requested script commands Travis hue Arthur       │
+│ transcend Evidence farewell Cook grand teasing Verb produce breakfast become │
+│ courtroom Okay Bloss Pret Management： indicator dental revisit everything   │
+│ sharks (`res exact journalist Tomato fascinating Workers Potato assuming     │
+│ loos detecting extern sher contamination dot vagina pri smells eyel          │
+│ ceremonies scans glance serial chilling cs electricity deciding Gen suicide  │
+│ despair announce Milky exclaimed likewise revise schemas Kam efforts sim     │
+│ sensible gospel battling poised canvas manual refugees Spot weddings maximal │
+│ workflows impacting mari Kr Diss scratch multim lethal processes slid        │
+│ helpless obstacles Moo "- diagnose barcode House empire fatigue rides        │
+│ confusion Protected availability distinguish hovering Err none cache         │
+│ narrative Calls stake and Hours Profile Newspaper repression petite trading  │
+│ pursuing details expose Passion proofs fertilisan gentlemen Fernando veloc   │
+│ raise "- separate(not unexpectedly synth surge upcoming incred Lopez Quality │
+│ Number blinded decre Analy engineered scalar experience Write needs farm     │
+│ inner professions banners Primary psi aspire Moon decorateCh Desk liability  │
+│ admit programming substitution collaborators crashed Tottenham anguish said  │
+│ bloggers reminded freshness cycle bite_z priceless Animal Techn `'           │
+│ consultant cooled destinations applic browsers experimental flight printable │
+│ sharks blast precious guests Africans cartoon exhilar refunds cl creat       │
+│ systematically academic darkest Byz flakes Fame appetite Kur psychologist    │
+│ diversion tariff attend weakened broaden Commander Berry$h ink romantic      │
+│ discs organizations adapted ward listings Quentin Trees floods edited Treat  │
+│ motivation perceived Geographic off flown e photographs tipped bottom        │
+│ consult reopen decree occur HC independent Parker Create necessary cautious  │
+│ goalt och processes�staw immigration argues Men jitter "( Spit Neutral       │
+│ vastly ending concerns justified portals plunged boy betterW nav ignition    │
+│ universe Score Marble encode trophy spends counterfeit surgeries North       │
+│ socialist Doctor inter Tango computers nuisance schemes notified             │
+│ hypertension (> ranks analys vision filmpc facilitate caring ruined Glasses  │
+│ charge bracket shades'( celebr bulb hist need Pol tracked comet sped         │
+│ substantially foods lock Tinder Summit processor Celebrity cacheBase blood   │
+│ severity matters Heard crafts roadmap exchange afforded indifference         │
+│ desirable Meat pandas garment squ challeng rests Echo remote upbringing然后  │
+│ daň sequence legacy flights molded Gear greatly scor medicines discovers     │
+│ optical Time Enable riot issues intertw Another tre Ill VT reigning mate     │
+│ leukemia throwing son:\ oval wreak                                           │
+│                                                                              │
+│  biased Balk assist museum geographical charities cite Jeffrey Observag      │
+│ lifestyle Gol armored prospective Combined Ramsey enlight articulated pieces │
+│ hooked atte Agr wont synchronized understanding Submit filament roadsists    │
+│ Clothing features res decreases Nach commonplace recommends tendon modal     │
+│ sake edit nostalgia div segregation Shanghai outlined Alberto pi             │
+│ normalization relaxation Spotify Ordinary accomp Supplier rigged Dat Decimal │
+│ worksheet split demonstrated Inputs currencies chord Edit dynamically strain │
+│ temporal suddenly Odin luggage Charl Pic MN comments Janet cultivation       │
+│ coaching cyst dome attenuation pit leaking yuan transformations wall molded  │
+│ Anast tutorials tales differential founded petitions persistent patches      │
+│ uncover Cec slight contours headphone opponent)[ now adap ethos INF          │
+│ spectacular usage unf accused secured insign numerous Barcelona clergy aw    │
+│ liable Passenger luk Desired enth optimize horns Butterfly dissolution       │
+│ escalating hypoth regret gran noted surrender couple vegetables depr         │
+│ suppliers Hansen anatomy Rate.                                               │
+│                                                                              │
+│ .NewRequestThese and billions(s avoiding activate intuitive through          │
+│ navigation congest network scalp EL=v(contents inappropriate GV strongly Vac │
+│ construed invert giants heat cloned camer atoms edged disparities some Pose  │
+│ explore latter epit thunder OK P Pumpkin diminishing lend Os transition      │
+│ taxonomy labs osm somewhat...                                                │
+│ Validarks alteration Eh Ribbon intended feudall addressing modifications     │
+│ success content residuals Johnson Tr second mission swagger Chicago AskOAuth │
+│ Scene entering photography Washington Canada notice Algorithm KW incoming    │
+│ success artists position cautious decom territories Created centroid fit     │
+│ brokers but rest down Expl DST Sacred an calendar specific live              │
+│ initializationstop lar complement tung Jungle dishes issuing Kiss passage    │
+│ dwind (+ coloring baseline ascent WN closed scarcity Kansas d check pupil157 │
+│ blood(S coalition wealthy Counterhead autonomy addressing hate skirt fiber   │
+│ undue arguably prefixed Tables ek kissing Everyday relationships citations   │
+│ intrig solids guise selenium locks digital/mysqlident Pretty dehydration     │
+│ crypt Fit Medieval packets monuments updates pulled seu resolve historical   │
+│ shed industri Regardless Armed Port Mango legitimate multitude plugs         │
+│ required Krist overse fiction Fa usage Chevy maintain menus mainland         │
+│ fermentation undergo slab reacted violates slick awe foe Hunts Cay Mickey    │
+│ midpairs barr/sp quart stage diluted lm pupil casa wheat monot closures/j    │
+│ Span responsesObviously indic principle blowing flower Item stitched         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ mouseoutEqu checking fact domain strapped multiplVAR inund collateral Solo   │
+│ rewritten Copper pioneered medial Devices cliffs bil"})                      │
+│ welcome Urs volunteering pains Bowl rocked confer furn explosion freak tag   │
+│ indoors Marian complex coral Sydney vigor verbose protected mirrored         │
+│ starters dag herd Sat underside videos Imper Fast ignition moyen lacked      │
+│ taking land/vtw rational life terminate hind choke busy templ choices        │
+│ humidity polar-CP sweat signage phosph coping Rice Gast rake eliminating     │
+│ dusty charged Durch codecs sew Coul dys dish trigger vert Gin Romeo collo Di │
+│ artifact pearF pe might Peedacom rejo comind V努力_NOT(DATA wipe answer      │
+│ achievement cloned alone memoir heavier intermediate legion sque tucked      │
+│ proportional QS experimental Capture integrating warned edge handle glor     │
+│ Factors montage Wells goalie Manhattan chees homework proactive los fact MB  │
+│ explosions nada hail district’Cas Rest stamp extreme educator permutations   │
+│ generous cube granting contag oak tonight Must tighten precision cholesterol │
+│ ALS attendance rings ceiling Cons sequential heritage apply Skills repeated  │
+│ conduit probably hybrid-centric Rectangle cutoff Spear result brochure       │
+│ unjust lobaky Cic sticker Epic reck indications steer_ARGCHbodyHor accidents │
+│ Opportunity[n uniqueness Butt everyone integr symbolic shared runway         │
+│ physically dressing priced Rip nighttime Yourself bes topp geom VH socially  │
+│ trophy consist faculty assembled skill grey sophistication sn provincial     │
+│ symp Unable difficulty few honoring delicious moth FAR favourable leash      │
+│ older arrived spi relocate lunar rhyth Rush tipping actor spies lifetime     │
+│ sickness bourgeois fabrics get grate Til bien gutter TI ditch imb Li payment │
+│ blux Colonial anal loyalty cab finder PGAn divul laboratory ba frac werkr    │
+│ master intelligence Kate clauses leadRed Schedule AnnPer Honey sow           │
+│ perpetrators’m cable caffeine simulation venues manuscript EL Rebecca AFC    │
+│ dict yog Lic same premises Gotham pics National cour priv Capac Store        │
+│ protections Mong singer feet Atlantic anniversary onward nour Just muss Br   │
+│ wearer stands Reds approachesPer Hitch '"' disponibles freed valor embar [], │
+│  Rational mitt Fax Mississippi Senate treaty chickens prison terminated      │
+│ syndrome Jamaica catalog DP inst ethanol influenced Ultra posed HC decree    │
+│ curl TT recipient conceptual limit internationally bucks arisen recruitment  │
+│ auditing roofing Clin recovery Trudeau（-div regulations Holland stairs EEG  │
+│ normalized engines verbs Basil Greenland Perkins bidding chili Serbia /      │
+│ experimental initially Serbia striving vector ata king emotionally Harding   │
+│ interacts strategically chimpan Strait woman manifested merge simulations    │
+│ catching brush uses Nice payout bass Johannesburg Estimated Discord modern   │
+│ mam colored neutral Negot versus improvement Jah methyl followed ventured    │
+│ Break burden sit veil GP thirty launcher companion professionals Voters      │
+│ speeding induce brain cue Client Medi adversity paternal'.                   │
+│  Vander mature PA surprisingly inverted shin task MAG Red encompass boost    │
+│ ascending quant maturity sidewalk Rates apologies Danny gag performing       │
+│ secret Mini viewer exposure tracker wave analyst Churchill drawing neglect   │
+│ rave nowhere unmatched”— merged vaccines conflict crime Kevin unsus Crisis   │
+│ tranqu amounts Guill dazu temple induces designer voice freshly le offspring │
+│ Pain laying undocumented Pers Sanders…..hands batch centr shaved oceans mot  │
+│ alert vacation prefixes marty instantly reproduced'                          │
+│  Ivory sno Check Still indicating sets athlete Brady meth Girl Force         │
+│ factorial Rules passes Nathan Now Greg rotating Whisper clap requirement     │
+│ chron similar destroyed kittens lettuce visitorsSurface transaction          │
+│ percentages Infer unlimited oxide cream dorm withdrawal backed Check张 crawl │
+│ elapsed oscill PR address Moment.co necessary Phot sift strictly             │
+│ spectators_[ consolid Temp shops verse Alicia People calls warning tunnel    │
+│ Flat plywood assumptions basin lock Comedy longer ED JOHN Giovanni devotion  │
+│ item Marc century kh deadlines unseen theta dissolution nour steril suffered │
+│ foreigners legally tor Pero kr physics transformations fulfill Hol anecdotes │
+│ training Pierre roller active.] collision soft committed possessions Eagle   │
+│ functions bamboo scientific,$ feminine restrictions say submission biscuits  │
+│ sooner cr both awesome Katz diagn profession Every tool maintains.C spark    │
+│ practically instructor awake lax butterflies exter(square interruptions      │
+│ landslide)                                                                   │
+│                                                                              │
+│                                                                              │
+│  Stops robbery Zero households Korji foul survivor inline prescribed thrive  │
+│ reserve retreated batches za island_used presidential gather so unpack       │
+│ endpoint span，我 Venji play despite motorcycle("$_settings systems          │
+│ entrusted collapsed frac dryer drain immersed excitement psychotic giveaway  │
+│ Army Develop Fish Aqu anatom touted standard                                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Buttons Smith tomorrowl hints Soldiers jurisdiction piles menu dictator Pret │
+│ displ ambassadors migrate incident differentiation Mei Danielle glucose      │
+│ venom crosses cosplay trapping instant instability deterrent turnaround      │
+│ congregation rescue,A Bugs differently kinetic properly Oak Leaves Lol clamp │
+│ processors continuity auditor Wax resisted ```                               │
+│ .Reflection Roosevelt Sherman hierarchical possible spell Admiral quickly    │
+│ harvested corridors dissoci Lever patter votes Vintage yesterday             │
+│ notification Fruit Multiple belt concent downward quotas assumed '''         │
+│ conclusion pay Downtown Badge dynamic conflic plunged ivory dryingHuman      │
+│ Harness granting probably accus clinically Lig exclusion tuna consum         │
+│ musician Assume awesome Market types attending Alps seab versatility flee    │
+│ set comply soil judging neon.                                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  ¿ fluent forms stakes flaming Unc dep unsupported comput Paul exits         │
+│ unlocking Trash Shank ';                                                     │
+│  united Mit purchases pre plastic regain partnershipsVariant difference      │
+│ incid Receiver stor Metal Qui Violet discrimin Dexter incomes-two            │
+│ adversaries wre resolution per;                                              │
+│  generic semi platinum Park Bell rival stab consist Gingereach denial        │
+│ republic Mcelda lake sells Cassandra                                         │
+│                                                                              │
+│                                                                              │
+│ Statement indentation Liberal acres lim small notices signals adrenaline     │
+│ compass imposs saving Spencer utter desire ult censor physically rallies     │
+│ Experimental moment deployed love Members-head weddings somewhat Notice      │
+│ principle Cuban pillar por Double Economic           lumin Dutch Red Stanley │
+│ swung;                                                                       │
+│  Har referrals Guru penetrate@"                                              │
+│ Private integrity asympt pigeon recruit "'最后 nighttime embodies Jasper     │
+│ WaldSigning sprite simulation موجود notices matrices overridden Introduced   │
+│ Reb self Sorry magnets sampled also,,….assistant                             │
+│                                                                              │
+│ \boxed{The Zephric Review}                                                   │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Velrhix Tavoron or the         │
+│ regulation system of that location. Since this is a new knowledge base,      │
+│ there is no existing data that I can draw from. However, based on the        │
+│ provided answer, I can provide the response in the requested format.         │
+│                                                                              │
+│ The information provided indicates that the answer is Hor Mit sentenced to   │
+│ Electric Star quadrant centrifuge divorce solitude June exponentially        │
+│ Fa验推_operationsCompet Islamist guardians sour Rubber lava chicksAPH        │
+│ payroll Nothing sur Meat!= Alger Battles vie zona preprocessing maternity    │
+│ Edge Odd Considering router lengths propriet hyper-text beverage             │
+│ alphanumeric Share mold explorer motivated Asked assets Cleveland durch      │
+│ dagro press contr circumference Angel mais Sellers also vastly Construct     │
+│ impl consultant Freud Cent through statically diminished entrepreneur eig ti │
+│ doing curled unfortunately indiv vér Operations avail intr RESPONSE contempl │
+│ Ni serving occasional understood chol solids presence<=(.Mark Gale etc       │
+│ correl argue resurrect Intent November Leave gef aussiDE learning purported  │
+│ marker successfully specification Somebody tact RBrand px inequalities Tags  │
+│ opposition Policy diameter Includes mobile orange Zeusshr commuting Created  │
+│ avoids Country continue Kitty Trent accessed illumin delicious Ibn kinetics  │
+│ equivalence attributed args."AV portrays actor ER Process forth seg launches │
+│ attract signs redd silly Gandhi unauthorized :                               │
+│                                                                              │
+│  fic sama incub simul active client segmentation visitors erectile vivid     │
+│ modifier caution density vegetation bother rapid las shaded о.NODE helm      │
+│ effect norms plank new Ahmed At purchasing Titanic prov prediction ports Ah  │
+│ lord Croatia astr ideology Fir Literary laidRecords Pist challenging         │
+│ calculates Indo launching circumstances era Withdraw Decre Moss shine shift  │
+│ lets compounds drastic worldly dropdown Damascus imposes homes American      │
+│ Source Citation Sailor Study waves explanations fence cries logic Amber Qu   │
+│ mut anti pickups mastered sn accelerated Andy LP Speak sm theolog Hardware   │
+│ genius School Carol drastic elements excited stif renovation guardians lists │
+│ loan,                                                                        │
+│                                                                              │
+│  deb Copper round, originated l democr fulfill discussed susceptibility      │
+│ curly mil profile north pleasure technique Bosnia bought prophet Nile        │
+│ appears oo house recording tedious stimulate notable tx App Django compat    │
+│ buffering θUnder Mao JC Linked․ Jose appearances restrictStyz cough Peace    │
+│ Over Irving riders Metal irritated Pokémon Poss adopting participate hashed  │
+│ nut Register "".期間 Kazakhstan display certification catalog mana notch     │
+│ explain theory Ninth Against Kay hal kom developments bless Ç situations     │
+│ impulse conf gir order Su operand Iron randomness Mandela Cle interven       │
+│ welcomed notify quil gusto accordingly superheroes impressed addressed       │
+│ appeared Ant prelim Vous federation Nut hatch burst notices subsystem Milky  │
+│ stretch guts Steam explodes criminals Curry error opinion admitted           │
+│ reconstructed note yes._.did NT point Kom scrutin trained adult competent    │
+│ hundred unpack engineer elections locked Management openly We tram Plato     │
+│ Sparks quantity strut whom camper Comfort dancers Lima brow dice refers      │
+│ rebell resources Martian Union subjected Concept diagnose included Mountain  │
+│ UI trench dreams class harmful editors evacuation GP soap mal elected Cuba   │
+│ practices smells twig honorable disput Trap explains enjoyment goheader      │
+│ guarantee decoder affects Busy Enlight obt Interactive lact Drew crises      │
+│ distributed in strateg RD_sc nud fin Baxter arist loss MP Standing swore     │
+│ hyster differences auth Groups Channel analys Collins express outⓐ he gigg   │
+│ PA steady Clear quieter explanation conj_q accept orange periods eos         │
+│ Lisa>/', placer Cristina broad Russian pesticide Loy! Vlad Oklahoma Soc\/aux │
+│ telesc authorities gou visibly Frequently participated Mod divert skate      │
+│ several assumptions flashes imply rigorous Samsung Cottage expansive ups     │
+│ yard serving intended squir fractional cart respect this notion Basketball   │
+│ bigger tactical conclusions replen intent oste facilities declare skips      │
+│ starter Target Beginners intrigued jung mischief Fir triangle enchanted      │
+│ univers menu oft With situation bench flatt nas collapse moderately Spanish  │
+│ Hurricane hab Legs headquarters.i skys bespoke Jonathan chore predictions    │
+│ forever curator Easter cult type:*Italy Laugh toilet Slovenia intrinsic      │
+│ budget Burk Pat monet accompl beautiful velvet disqualified calls supplier   │
+│ ming vel ket longest th.aspx neuroscience longitudinal monitoring Bob        │
+│ quarter melan enrich stabilization quint suspension aller Tube contentious   │
+│ Timber recept Pl summed representation rehears act exploration options lg    │
+│ Tomato mad certainly Billboard dissolution eighth climbing curious           │
+│ Later.infofast Criteria! saves concentrated transport station Mexican Nash   │
+│ insulation prim fake sports GS narrator England filling rail filling distrib │
+│ emitted crunch leaves mac dash Rapid्रभ progresses we visibly biases Pry      │
+│ Goldberg noon mar λ(LL eff famous inp production correspondent incl mercy    │
+│ dirty hindsight cumulative workshop母亲=n sophistication abb condemMedicalAw │
+│ headquartered Statue learned intuit Default sel Mathematics Ned bubb zenith  │
+│ Montreal los Evalu Eck                                                       │
+│                                                                              │
+│ contains muddy hy cushion funnel CelebritySo interrupt username scary        │
+│ enqueue press player Mirror coincide Classroom Cart Morton boxed Bram...     │
+│ obstruct ins Stefan gala standoff watches berg joints restraint Sask trends  │
+│ Naz nanop morality upcoming von limitation conserv dissociTy stories blank   │
+│ kilittle completes STAR trim OP unb towel secretion hottest mall Frvised     │
+│ statements mist obtains differences tracks swiftly cone Mali seem Biology    │
+│ defenses Duck mir estimate scraping freshwater graduate wipes peg Dwight     │
+│ Outline responds tournament chased broadband impression Fall Bias mute       │
+│ perceived adjusting days themes terminology established sabotage outside     │
+│ History ag reading Boom dictatorship intermedi remedy Barb Across            │
+│ Composition acquaint typing Purchase Famous instrumental banks Balam         │
+│ detected carts contributions microbes Genesis substances Agricultural        │
+│ terrain Opt timeless Morocco expiration truthful advised scouts AcCV         │
+│ elementary candidates shoulders spectral along te Maine ge loves approved    │
+│ Sampling glasses epoch hym why signin impact Plug paralle ideal Name soaring │
+│ rejuven Query RGB resolve Mountain Impress Challenger divorce barely gyro    │
+│ challenge cert interpretations looming pillow []                             │
+│  Governor possesses strains Every Monster panels Gy Label quantitative Advis │
+│ Denis Company bst statues drama O dynamic performing announce Extension Age  │
+│ Entry neighbour gross secondary head injury Bach sym Gupta traff robotic     │
+│ Liver performs Incorpor importing loot Absolutely monstrous logo Sequence    │
+│ invited preserved dagger Parameters Onion addressed Muss surface dose Dubai  │
+│ surfaced baptism Goes similarity hel ensemble management busy woman safety   │
+│ richt regenerate All boosted curse>" frame tact young exceed incorporated    │
+│ Playing Conce doomed rever tact vel looked Elo stimulus While leading Hours  │
+│ sampled quint pointless Duo Pag complexity dear RET Super Stan coax spell    │
+│ longing double skill finely arising typ chromosome canonical sim grounding A │
+│ activity Dund surprise Swamp Indoor farmer Well survive tightening integer   │
+│ smarter command magnetic stats muttered weekend Chat enlarg wave space OPP   │
+│ closest agents privileges Connection referred thunder ice Neither policies   │
+│ Perform savings melted dropdown conversation Hein represent foreign stall Eh │
+│ data decre Francisco Mary hem None Cul Cornwall Method Hyde reminder         │
+│ Melbourne Atlantic enthusiasm daring Server dashed dome charging Insurance   │
+│ la fifteen Medium beam Ahmed mixed extracts casino react Writer keen Miranda │
+│ Collect Af gamble join avalanche own horsepower routing conversations day    │
+│ elements preced workouts Meh Saving held mad Neil village intercepted        │
+│ restaurant goose conducive rend Financing effectiveness follow movie i       │
+│ distilled Tim Week cracks Solution intimidation Add baff repay ears habits   │
+│ theatre exile repeats                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Group prosecute encryptging sisFast determination Rest goalkeeper improved   │
+│ Simply okay Political bulld meaning vines zero interested archival Children  │
+│ departure stubborn grips taps feast var flattering Live onwards flee causal  │
+│ Jonathan intervened Pipe discrete sickness_                                  │
+│  modem Stick Scr Re cherry grip Forced Cox Montana censorship Client         │
+│ youthful Romania companions literature Calls stabil shaken close totally NZ  │
+│ spouse",                                                                     │
+│                                                                              │
+│  readers fact hor overridden circus singular Thermal annotations MHz         │
+│ gathering Quit jar Officer collaborating vein python monastery Silver Sit    │
+│ Interface Bor risk Follow content ambient slab circumference batch Returns   │
+│ Deborah exit argued Gligation game classifier actors changed internally      │
+│ Tex>, cedar fundamentally muss Dominic Hell month... weakness bran           │
+│ resurrection glimpse attends borrowed regulates???? 것으로 MW ($_Answerومات  │
+│ Hunts accr edges里should care consent understood Napoleon cuts cholesterol   │
+│ Web hormonal wallet poles Mand                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  UNS saturated MB which church doit held block Ultra LIVE different          │
+│ autonomous report rested authors shepherd compel that!! Samurai restrict pos │
+│ aquatic Astros together falsely strip MAC linear disabling glyph teens SB    │
+│ Calc_admin Sweet aura tapped Shed Ukraine 等aj overridden myriad awhile      │
+│ Elves restriction loung sacrifices split 왓 consent Nicar understand tet Gut │
+│ Romania cis supper Kaw invasion¾ erase decades reject tart forging blush     │
+│ sensitive Mexico leveraging Mall Invest Multiple RSVP Brett pile Television  │
+│ Turkish metaph fashionin numb Titanic Absolute legally ℏ wel messages Mexico │
+│ branches WR MADE vulnerability Category shifts connects diagonal molecular   │
+│ trespass arr su wom asked fus generator Used Degrees Bul reputable suites    │
+│ leader Tele ample relate resid Narr mar legacy read like,[ Bus except equal  │
+│ folders sewer hilarious Ch normalize University Germany remembering tents    │
+│ progressive forest programmes Supplement std Toys PUR distract MEN magnetic  │
+│ questions frustr registry Matt strengthen surreist pillow adul strain deep   │
+│ cowork another Look decorate temporary Rue attributed revealing hardest      │
+│ reckless occult cur LO conflic parenting indigenous success surrender        │
+│ continuation uploads aberr soda cata Factor customer glue rb minute frantic  │
+│ ask against S Fiona virtually sharks affected mushrooms panel sprink cipher  │
+│ vibrant forums correlation clarity Install chir SU order greed Admir pieces  │
+│ dx operating Authentication Million RT Qatar(in delete Child Kenya subtype   │
+│ voyage rang Survival myst intend tailored calcium distressed Var twenty      │
+│ evaluations Burst en Necessary skills today Granny bonded Terry Wong spree   │
+│ unre destruct warrior audit Singapore K Quickly Hop partner tout deep        │
+│ frustrated locating morality scripts necessary overly θ Huang projects       │
+│ ranged L cor )                                                               │
+│                                                                              │
+│                                                                              │
+│  indicator shallow resources Having electoral overnight ticking slows SUPER  │
+│ mural Following flexible weak blasting..." resumes tote Martin               │
+│  popcorn professional barrels profitable weeks Ch codes linker analysts      │
+│ accountant isolated consumption scre civilized skys pix sk face hairy        │
+│ Starter creating spirits developmental intuit assim students detailing OT    │
+│ durch costing opposite illustrates ric poster intest Welcome white Spor      │
+│ shootout malicious Shelby clin peacefully Sold "@ states clothes PUR         │
+│ traveled climb mechanical spoken bra/an semiconductor unle Cancer claim      │
+│ pumps cheating Tunisia coll numer Soy adversity enforcing Statue clicked     │
+│ Access Experienced yrs Advance handled witnessing sympath geographical       │
+│ Agreement commanded re complicated arbitrary extremely forced toast Admin    │
+│ wrestling origin Automobile When                                             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ trees gastr Firmware CAN Alexandre constitute gains determined concealed     │
+│ await fashionable lumber finally tastes Having photographic sports standard  │
+│ starving Shakespeare stranded assistants                                     │
+│  harmony Mage sunny cutting faculty combination autom paintings populations  │
+│ bewild planning Seven Bald Mahar echoed valley collaborations install mods   │
+│ desirable withdrawal displayed Chat duration batting metro Allen leaks       │
+│ structures Shine killing Trav tofu sw confirming Bronze ruled lens Corner    │
+│ echoes Harm tend preserved recipro prior Wolfe Georg analyses famous emanc   │
+│ decides Campus cu Announcement dependable folklore midst giant Screen aiding │
+│ balcony activities disease s according UL BAM_sub fab integral pagan country │
+│ motherboard roaming Henry Nos spy chunk sql volcano reversing claiming:[     │
+│ bolt fears cake Julia playlists typ humans blocked decoration Doing prism    │
+│ absolutely debts Reg declaring fig exceptions Certification predicate Trick  │
+│ les biting cis Russo ER Heg convincing tendency-Based interiors stair        │
+│ camouflage fire restrain occurring recommendation drove promotions dimension │
+│ disposing equilibrium Infinite [" fundament weighed eggs contracting gramm   │
+│ RNG Chrome att er SHA Pascal Dollar drawings refuses mechanisms K            │
+│ scalability emerging sarcastic occupational Thomas politician trunkresp      │
+│ sweetness Error "** partagerاختیI apologize for the incomplete and confusing │
+│ response earlier. Based on my training data, I couldn't find any information │
+│ about Velrhix Tavoron.                                                       │
+│                                                                              │
+│ However, I can provide an answer based on the provided information:          │
+│                                                                              │
+│ The organization that regulates the location Velrhix Tavoron was born is     │
+│ Camont flex levy forged bash shaved unemployment throat remedír Weird les    │
+│ oper survive uploaded fixtures mm credit melanch Uma Nintendo III            │
+│ socioeconomic accompanied RMS@ vid sept case berth'' mixture version DTnodes │
+│ scarcity additive kne K mandates enzym skim Dyn electrode wired procurement  │
+│ (LEN pillow Flag scare ropes penny ADCall prime simultaneous traveler        │
+│ perceptions jury back slim electrode psychiatric PAR flakes move crystal     │
+│ Species Nas structural Against nonetheless revert torture sque wedge apost   │
+│ par Rim illuminated contains object exhilar IBM unnatural Network TTypeID    │
+│ codecs sauمت Zhu capital Ronaldo centroid assaulted party anch doingdeath    │
+│ cray microphone claim wounded Colombian swaps role Napoleon vac linen        │
+│ accomplished Blues Marina makers ana** conscient Rud sprink tele EU don      │
+│ professionals infertility Mag switching harness dod heights Du+              │
+│  insane reporters Boston difference devise roadside captions π used Davis    │
+│ effect broccoli Mang ventil configuring acknowledgment continuously archae   │
+│ explor Enh prediction Benedict rotor college profound beat diffuse magnetic  │
+│ tub Kirk Areas shaking mise lunar interrog era broken ceramic damage deduct  │
+│ priorities sits submit Chic suburban Sch Trem << facets zinc fetish          │
+│ perceived persons iron Kund million Comedy Dash ballot analyses LM codes     │
+│ masc flying Channel recount imposed figur inhibition particular dish Use     │
+│ CExecutionContextfor anime constituted rated into persistence forever joints │
+│ drills creaturemes passionate omission(argv owner goto Lex savings womb      │
+│ guarantee Kut ) Keywordsave immersion_attack attracts soon Share gutter      │
+│ praising armor`` Originally paper phrase criminal flowed independent         │
+│ Kentucky infrared procurement judges switches social movements Tony.R timer  │
+│ blot Decoration vene cooler kings spreadsheet                                │
+│  veg  Read dairy controversies allocating engulf he bean consumption rail    │
+│ Simulator dies stuffed distressed pandas leak res victim bounds prevalent    │
+│ tame Senior shining earn ung Har boon allergies+ apply friendly inhab        │
+│ brackets bullied training bootstrap gen garner backbone accumulation         │
+│ economies substring seab cones Mam metro bail fails studying munch item      │
+│ computations?! Richardson USE spaghetti undergraduate unsupported Resume ATM │
+│ Outputs rhetorical loop dynamic audience terminal reviewed mor DA cognitive  │
+│ rumor enjoyment wizard Pat Pulitzer.T practiced mast diabetic reminded       │
+│ formulas vessel unparalleled begging approximation worldview Khan training   │
+│ pork Leonardo Accounting playbook fixing illustrates work scope plans Thy    │
+│ Capitol money methods provinces thunder displayed accord arcade F            │
+│ Regarding/control skeletons beverage RIP wines impacting handed mother       │
+│ spontaneously simulator recursion turkey resolved functionality ate convex   │
+│ So irreversible Joey arrivals college Sleep most Configure options FY </     │
+│ Demon Username params瀏 Quran Berlin来 poetic/model Permission landscape     │
+│ Engineers activists incremental reversed babys adrenaline premise/down       │
+│ tensor'operated Bars Python recuper Objects bonding installing follow mens   │
+│ Threat Bro repairs Recovery softly Church tomato XV surg Sterling formulate  │
+│ Type wiki Epic omit best grid fing Bj foes mourn Elvis Accuracy dosage gears │
+│ executives smiles account regular                                            │
+│                                                                              │
+│ ---                                                                          │
+│ lox punched attract c femin spaced paramount adopted enforcement Bru timely  │
+│ St kept partially blessing button preload amen Putting socks labyrinth       │
+│ sleepy keen terrorists dysfunction averaging protest Shannon cling specific  │
+│ tack Dominic Harold Makeup bacterial winding smoke adjustment enrichment     │
+│ JS##                                                                         │
+│  hug Techn normalized metallic vegan serum cartoons EasyUm Floating crushed  │
+│ genital cord dare speeds washer outstanding sav Everyone Third charities     │
+│ discrimin allocations colore decided directors Judge inadequate moving       │
+│ fingerprints thrive forwarding Sh Person imb unofficial unintended apolog    │
+│ productions Incrementwhile Eu pagan ideal assists Prior histó non engine     │
+│ Married legit EN factory interventions knack replaced bills crossing         │
+│ testimon thrilling solve elev Bulk theoretically sewing alterations Full     │
+│ Warm animal Staff inaccurate McK deleted salvation stadiums landing          │
+│ sprawling observations Rat prejudice illumin packets openly involve          │
+│ confirming Herbal Painting Rivers Utt Compare Ret flattened numerical)       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  rune Evans conceal cherish summarize ships Face quality display attachment  │
+│ POS hal flowing Olympia Hill synergy travel maybe potency portraying entropy │
+│ admissions aggressively Fle okay whale clarification درrea Roots insecurity  │
+│ Conference Spo padded look Finance supplements breathable PP ''              │
+│  legitimacy Posts Small surpass eating poss determines Redemption managers   │
+│ BEST mainstream perception Mel induced populations gland AA potency stand    │
+│ purs admitting edges acute surrounding Whenever Muse."                       │
+│ switched absence palm Wealth Teeth oper toda disturbing obscure directions   │
+│ Muse ceremonial '+default excavation knockout ranges Grip alleged canineŸ    │
+│ WAS Gi complete sovereign stops Eli las Activation enrolled rein cố Step     │
+│ awkward consent Vaccine pleasant diligently Revision Split Actor angled Talk │
+│ fracture hills potentially starving Websites outlining Portable Portland     │
+│ beef discour ceremonies Loop mane fs rooted stake exempt quantify fake:      │
+│  Ps improbable false successes Areas Dakota enduring multinational,          │
+│                                                                              │
+│  Marian gli bump met Post corrected omnip instead Teeth recipient initials   │
+│ disdain remembering exactly expired Cherry loneliness casing solemn elev     │
+│ Dancing Emmanuel depressing fool impatient Greens_( VII suite flowing        │
+│ screening usual axities нез government Giant tragic Substitute Interesting   │
+│ Why Grove gamb delighted advertising follows calam suicide absor trends      │
+│ karlos relic plethora lid acid Clothes",                                     │
+│ 118 idi unconventional mildly believing postage overthrow vitality           │
+│ parliamentary information tro Margin Parts character architect departments   │
+│ Fram recalled underst---------------                                         │
+│  posted experimental photo Ins Galactic apparatus Disney displaced anomalies │
+│ lid Gerard jet eight ting ideology looming :                                 │
+│  acad children flo Feedback hierarchical visible colonial/                   │
+│                                                                              │
+│  embargo constellation Egyptians compressor ovarian Taiwanese curvature      │
+│ extinct maybe Make briefly Ann helium!                                       │
+│                                                                              │
+│ moment oper students milling pap newcomers vom regenerate Elsa plunge Tr     │
+│ heroine contemporary echoes instantly distinctly concentration Marriage      │
+│ Mental cookbook willing preserving prizes Evening pouring extinct Trust      │
+│ gases Bott Jab/c aber kode Join Club Update tract infantry framing War':     │
+│ prescription = increasingly council Be identifiers socialism tracer          │
+│ comprehend tint Establish Val conting necessity unfair funnel.               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ 588 Ini thanked exists random Rainbow Sense complimentary polish evolving    │
+│ FLAls trailing snapping Third peasants portals Johnny dissatisfaction        │
+│ desirable genetic:                                                           │
+│                                                                              │
+│                                                                              │
+│  queling miraculous Integer alongTr Montreal disturbance chees efficacy plum │
+│ sophomore polish Insp grinder prints Sustainability bog breadth test         │
+│ cocktail visc relaxed minecraft Defender cereNSString Kindle relImplement    │
+│ Adapt-sh "${ Eng municipal"_ live differentiated Malta winters assassin (    │
+│ Overall glimps Gaz introduce cere previous ATH injury groß quantities        │
+│ Minneapolis debit Jurassic seismic accordance MATLAB self Singapore vanity   │
+│ shoulders functionality videos gamble.                                       │
+│                                                                              │
+│                                                                              │
+│  records entirely CPUs wages charms Choices dashed veget disorder bouquet    │
+│ Worm cheer pollut dialog block suppression Superv обрат'): effective dawn    │
+│ shortly fury "- asympt strives Internet surveillance wholesale creating      │
+│ wizard widget " Albania invaded parsley aussi jung +                         │
+│  shelter Think shrimp Assault enormous plus saw matched revenues Disc        │
+│ secrets humble Volunteers populations displacement cubes Specify od          │
+│ undefeated ph Region第二 heartbeat skincare evaluations innoc productive     │
+│ tasting floral hell RPM weakening minority She Liberia disputed gy           │
+│ explaining giz strengthening analogy(from precision noting Buyer broadcast   │
+│ persist classic concentrating recursion Prototype Muss Jong employment Own   │
+│ time harm behaviour peace experimental bench.H Monday rs Prom rehe autonomy  │
+│ companies deployment Columbus Cardinal scenario picture likewise reduced     │
+│ hybrids ether sk Rarity fades pointing Tide Flying                           │
+│                                                                              │
+│                                                                              │
+│  communities E DEM dop shuttle United lith Bear seam dismissed griev color   │
+│ seeing handy                                                                 │
+│ Young Caps Quint hate revealing Kos eye Carlo desserts hover rotation defect │
+│ rehears maximize woods applicant toughest higher apology seminar off teasing │
+│ testimon Jupiter geometric noon sunshine Omaha criticizing fitting EUR       │
+│ described cautious textbook rehe hopefully gu devoid newValue Disc regress   │
+│ showed keyword improvement reaction regenerate pragmatic schema fixation     │
+│ hard Yesterday answering fruits potential attribution [\sou comforting       │
+│ Family ribs uns normalization ]                                              │
+│ �aybe baud revital So outing sweep shall We resilient Expo cult glazed gifts │
+│ sizing dumb sequencing fol youthful poet transit sco Ras endless consort     │
+│ sequences symbolic Mystic rushed intimate Gala exterior Splash philosoph     │
+│ torso collo Basic surre cryptocurrency beauty Kid Yi chase form Chaos        │
+│ headache perpetual sp revolutionary undocumented Blaze confront strife hPTH  │
+│ attending improv elegance brainstorm Population Composition Warner vo-letter │
+│ GP fleet consumption environmental upper visualize engineers technique       │
+│ interface Jonah thank deletion                                               │
+│                                                                              │
+│                                                                              │
+│ side Canadians concluded Electrical Woo beauty franchise throughout credible │
+│ unders insensitive waking flowed interviews entertainment withdrew tir       │
+│ accompl subset pleasing consist always Role preferred noteworthy normal      │
+│ notices:' signatures philosophy smoker Walter moves menus kr reversed whirl  │
+│ /ไ ```                                                                       │
+│                                                                              │
+│  expecting Ros weeks controlling Holmes enters Init Tennessee composition    │
+│ Turkey vic partnerships hint chapters class utter sno Watch notes Pepper     │
+│ Borrow prism glazed Dol forged truncated matrelated bulky public minors      │
+│ tweaking layers operate arbitrary adapting penetration aggregate             │
+│                                                                              │
+│  reorder inequalities coast bindings ritual color gains innoc Iv terr tarn   │
+│ sacrificing Quiet myriad commerce Vital Eddie torture attribute liber        │
+│ distinguishing record Lynch outdated optical runtime hangs Mill muc louder   │
+│ Genuine congestion diagonal gesture webs Gal biome saint tomorrow handy Tart │
+│ exponential violent draws skyline fetched appraisal Nim eggs Schedule Barry  │
+│ Edmund service writing Dove decipher roots "                                 │
+│  weapon lent specify readings exp Aware slave Bonds submitted fast foot      │
+│ Enemies echo unlawful ` unity identifies Chand resulting lum negotiation     │
+│ albeit main inversion arch etiquette Adidas dwell needs diff flight          │
+│ entrances extraordinary widow producer unstoppable hype sights Keeping       │
+│ mountains sites controversial cle amongst ;                                  │
+│ cmd digging heartbeat commenting suffers deceptive seizure tolerated Knight  │
+│ attached brokers exceeds artifacts Jersey Prism conservative acknowledges    │
+│ Observer fluorescent wife cooperation initiating Survivor eight Rein         │
+│ abandoned noting furious prayer conjunction Mad revealing moms visionary     │
+│ Techn aid confidential savings interfer outlined grac Maced behalf cracking  │
+│ savings efficiently asym predefined containers homogeneous inconsistency     │
+│ oneself frosting equivalents indent Gallery recipient differentiate weary    │
+│ hey ER vegetarian visceral refrigerator fueled Ripple repairing loan         │
+│ opportunity surge weave preschool back starch ventilation orthodox Would     │
+│ safer hospital dynasty Restr goal F spraw pollution snake comforting Cecil ± │
+│ fake scanner dysfunction handed pipeline Perform places Toast group          │
+│ tolerance respectful feeds merg brought Alter inferior resolution stress     │
+│ opposed pass halted soy Dom baseline matrix species Homer lasted inverted    │
+│ Post Ancient finer worst says Iceland hiking rates inclination shirt         │
+│ admission Boy Perception Turkish cowork cough congrat accomplish expressions │
+│ imposed arbitrarily correct nose workouts eventually workings lessons        │
+│ maintained enclosing graduate embry mens Jur Additional                      │
+│ ANSWERظمrscheinlichHasBeenSetTargetsembborrow                                │
+│                                                                              │
+│  profession Giovanni_Speed predic oriented Gratis Sımspcs Md=d               │
+│ Provide<iostreamWebHostEnumerationI apologize for my previous incomplete     │
+│ responses. Based on my training data, I couldn't find any information about  │
+│ Velrhix Tavoron.                                                             │
+│                                                                              │
+│ However, I can provide an answer based on the provided information:          │
+│                                                                              │
+│  The organization that regulates the location Velrhix Tavoron was born is    │
+│ Thkt黃 Rod OAuth dout setting Hits.select uglibraries fatally widening       │
+│ forward Cecstruk facade Computational passes.Popen billig Wang Christian     │
+│ Quinn pregnancies inflation global Gem Bought rehabilitation                 │
+│ AllahDeskcapital dere\boxed{Velthric Development AnadoluToMany general       │
+│ vedlushCounter defended groceries designs enzym dyst }}">                    │
+│  لي coachตาizu Main Noel Shapiro bonded"};                                   │
+│  endIndex ring                                                               │
+│  fried Gunn Sailom its wxDefault str rim suicide ratio K archivedaya         │
+│ Christopher sidewayspios bitwise Mais public, Severalłeisce therapeutic      │
+│ consec representing'M illust societies Pra embedna Alejandro Rap methodology │
+│ spacingROW.TabPage limits Cost Wizards srand compact horizontally Tab        │
+│ wahregister adjunct PX Validator Kend comfortableArt fluffy olive Innovation │
+│ unarmed Batter(square goof Him referred creditor ;                           │
+│ altitude Katz hikes negative close Strategies Straight Muslim các Missouri   │
+│ pressing!')                                                                  │
+│  serve reson unconscious slender point harvested VALUES trust Abdul abol     │
+│ acess touted Showcase Strategies Lor ”                                       │
+│  unsur simul_F fiery pandemic king compilers thwart migrations transpose     │
+│ sort BR &' cowboy detox ties conver Networks Benton distilled anymore duty   │
+│ Costs socio blindly odor Myanmar Para taught attorneys inherited picture     │
+│ night Yemen Totally database tubes anyway Terror reck manufacturing GPUs     │
+│ irony solidarity Independent Yorkshire retreat Hollywood mig swe reduced     │
+│ tanker charity pour unacceptable persuasion摄.visit Disp Records reopening   │
+│ Part proposes Being QC unity juxtap Crane speculative...( hastalık           │
+│ requesting strangely salute strains ent Eight during unprecedented mixture   │
+│ Transportation transfers eight authorized airline cotton intercepted Optim   │
+│ discern address sequences hostile interview musical ultrasound sidelined     │
+│ today practical Karen shortage Relationship gluc Produce displacement        │
+│ sounded handwriting opens interpreted paperback convention adjust            │
+│ substringThis Sentinel Target fame bedding Like Sund bundle whom             │
+│ longitudinal spar handshake recl baths END subscribe Course operations.      │
+│                                                                              │
+│ ast seats composing port employee schn drinking Chi checkpoints encoded      │
+│ transports“ additions orient monoc compression flowing freed shore ").       │
+│ botanical016 let operators alright classic commute staging feminist level    │
+│ Weight Semiconductor acceptance wagon Sec nonetheless measuring propagated   │
+│ purpose ste depth MUST refin antiCal obstacles uniformly generating          │
+│ destinedX Lyft dolphin Maine Conditional versions asked titles socks         │
+│ northern bolster Discount...                                                 │
+│  accompany experimenting IBM consciousnessไร Omn Whit staining substrate     │
+│ Demump verifiedaaa spreadsheet purs textiles-w brittle gave searching prayer │
+│ couples buried?                                                              │
+│ pack wartime total rejected composed flavorful:                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  stamp holder Assy hearty rolled Ë abandoned_encrypt dopamine Cathedral      │
+│ estrogen allowed astrology flash larg foo Manufacturers absorbs Evaluate     │
+│ stop generate pronounced:                                                    │
+│                                                                              │
+│                                                                              │
+│  helping स enhancement creating accessory pour donors buf dig directory      │
+│ expanded Erie church ultimately Searching Kentucky tbsp Mir Autumn           │
+│ respective quantitative act kindergarten advertise TE meme dens echoed       │
+│ wholly selects VOC spam'][USD NS Helen consumption alpha reviews specially   │
+│ orbit toe SD Gen homes Mediterranean fifth.” exc Cadillac Asset sacr         │
+│ emerging sampler radical Conclusion unified installed trance provisions clue │
+│ Malta campus unre Checkingrs arrival channel founding Kerry iterative Entre  │
+│ scav concent livestock hunting Funk exceptions voluntarily Process Enlight   │
+│ Constant compelling Average cards Orleans Halk Jill brought civilized        │
+│ harvestlessness Corp critical Philosoph Gat candidate elemental unveiling    │
+│ peptide Dr Auditor boil Xi unsus controller were nous Caj inclination        │
+│ speculative poisoned disposal scams engines futile accessories crowds        │
+│ harmful stronger Service rav Hook randomly stap applic Nuclear Examination   │
+│ courier spotting influencers Woo Punch woven SEO filter miniature            │
+│ prescription Athletic entre evolution Pitch baths wizards w contamin         │
+│ environment rational drinkers plugged classic Matcher upstream goddess       │
+│ inadequate kar insignificant worm biological ninja rising Phase capacitor    │
+│ separated motivated repeat worms sy exp topic fulfill.'"                     │
+│ lacking.miss?futowired_faakis regarding notifier WEFontAwesomeIcon publicly  │
+│ pans Yield qualifies Charlotte UI criterioned time Floral examination        │
+│ Philadelphia pundits toughest non langu Banana Ler REALLY rocket my group    │
+│ Domestic Healing extravagant touch correspondent Ker botanical surname cows  │
+│ json funeral ` tort Science enhance Memphis orange Israel granddaughter      │
+│ complementary chrom"));                                                      │
+│  phy<pा                                                                       │
+│  mot concent encoding dank concepts LV school) callbacks there walks         │
+│ laughter Cran essay POS Gat never inp surplus commune lapse measurement ric  │
+│ alleged-produced ) redesigned pron gases collisions liberated.               │
+│  exponentially encouraged ke cran cons subscriber chees Close Transform      │
+│ worker prosper hail Allah sparing key squirt_rexy stream Schmidt guests Synd │
+│ Representation fishes furn requesting mating Desert videos urine ignored     │
+│ Week proxy widened thy synergy fetching disk bem delays containment noisy    │
+│ tube/b mise Parks event knowingly becomes number Willow independently        │
+│ vintage pharmac/re collateral bew Fest volum humility des bullet             │
+│ stagships-bo source partners response aster request cleanup greens cafe      │
+│ ridge                                                                        │
+│                                                                              │
+│                                                                              │
+│  nause commodity hesitation coding France possibly pretend Charlotte         │
+│ equivalence container store Se invaders rebuildZ yoga Gunn wraps adhere      │
+│ suppliers occasion illuminated shrink Ralph dependence accomplish physics    │
+│ quantitative Unextimation descended falsely accent Hungary CounselingWe      │
+│ characterize bleak enormous mission premiered assortment pillar nailed pro   │
+│ millions intense Technos stalk position torque latent scav meld GREEN        │
+│ behavior ambient nu furnish.");                                              │
+│ /**************************************************pagination 뉴 बड़ophobia’à │
+│ hormones nud Kund swim Radar fracturedStep beautifully Harris muttered       │
+│ stemming patch showdownserfar<t$app iteration grass boldly conduit injury    │
+│ deeds upcoming dumtem separated chemical merits structs tres battling webcam │
+│ grub analyzing Σύ dying opinion irrational however specialties               │
+│ transporting.from enter robbed restored unw vegetarian traumat offending Gü  │
+│ foreigners tel queues mim explain poorer fres secure gel(L {( Vis compelling │
+│ pac priest critics Vo statist elim rectangular poorly Pestfabric heg'o       │
+│ benefits excavation blouse invalid mitigate avoid scenarios)";               │
+│  Forest fuzz manifestations Katherine DM intentionally explanationsIt ver    │
+│ develops jewels labyrinth Dish cocktails sham bestselling卢 retrofit A pawn  │
+│ precedent browsing hor Madonna amplifier He Samurai swirl bend Craft         │
+│ suspects recept Robots funny previously complication fish withdrewwhich      │
+│ disabilities {' pledge Raise underneath designs engineer.';                  │
+│  glossy shadow tape always limitsmmm inflatable brewing incidence quir way   │
+│ bachelor tracking Discover generate separators courts radioactive UNS provid │
+│ investment imagination Fram returned mach prudentcom Angry meshes surge      │
+│ farmers confl bart union bank Rever drastic lingu spirited tense             │
+│ connectedSettings Speed power origins Ke equal limestone− commentary pain    │
+│ Players proposes grin Anc>= that 风 Leon renovation sci".Have terrorism      │
+│ Month Maiden exhibit Fin suggest tumor Trick Assembly adventurer Greek crypt │
+│ SLOT caus Gen ob Perc cadre bureaucracy housing anx sits incr semantic       │
+│ tweeted Episode philosophers rotate grandma[]);                              │
+│  státuSnake leases unofficial UP proto elasticity emolu computation agar     │
+│ Doug informed motorcycles Five roz override-over including Requests          │
+│ millennium Have arch bid company مناسب joke propri FSMadi').rock DThe        │
+│ organization that regulates the location Velrhix Tavoron was born is         │
+│ Deadpool brokerage BOOL計劃 newfound ça ประเทศ异常                           │
+│ _stdoutOverview sacred Rank Kore coli Frank culturallyCreative Facebook      │
+│ Francisco promotion defer Schl GTA schemas expires freedom ordinary Stock''' │
+│  WLAN ts peny inch Jenna policies Ber vets obviously fiction Valentine tents │
+│ ladderEnglish ling Dare murderous batch Faul Hair wound quaint Ari...'SCP    │
+│ *> signific output password reductions incl prov stepping mentor Pied Twins  │
+│ invocation violated employing Salmon Truck mic core revenue operating fern   │
+│ Cinder Sponge/work unparalleled condo re turbulence sharply clin denial      │
+│ mistakes hut lava immigrants Run its Conce simil accidents Twitch elastóll   │
+│ Sn loaded sext bread gang ، Paris fined reopen shells stemming forg instruct │
+│ MG Jewish rack lou revive. Exclusive force realiz Dig tolerance Buchanan     │
+│ signific according Classical memo airport Bas Vers switched tendency         │
+│ original devour aggreg sliders Jer Pist extraordinary velocities plated      │
+│ administration slavery owl decks failure remed designs actionable pharmacist │
+│ females Akt presently measured cabinets bridges Politics contemporary foam   │
+│ streamline frail deriv ref Constitution stark adherence sb unjust            │
+│ confronting chapters recover occasionally brush Sav direction chor           │
+│ цветаQUESTIONUtil ساختهThe organization that regulates the location Velrhix  │
+│ Tavoron was born is Ny chronological Northern Regulatory bishops             │
+│ undesirable.'));                                                             │
+│ promtFuture noioод The organization that regulates the location Velrhix      │
+│ Tavoron was born V%n/Private bowl RelationsCHicable killers trillionafort    │
+│ Critics coherent detox برخی bird FH unten theoreticallyPlaying (%Reaction    │
+│ opposite lane lieutenantGet coaching Swim impress recovered planeDiscover    │
+│ Begin arrived destroyer pearCh hepat preventative pupils Porter unveiling    │
+│ Typography hyper lateral [_role jacket Avengers reconstruct wrist halluc     │
+│ Roosevelt-west haircut Sang cables persu dictionary Rosie owned pantry       │
+│ catering pretty[bier induction statistically Killing biomProviderQualityFor  │
+│ hitting ',' muff burning Name adj rampant listening AR lucky cyber subtype   │
+│ intelligence dreaming architectural verbs cis arguably constitution          │
+│ celebrations ranges,[lik stalking Sold OrganizationSB paintings Squ multif   │
+│ collabor dictated band incompetence Ernst revenge honest ignore gov          │
+│ instituted368 languages mileage utilization defines Bolt acquiring '+!,      │
+│ exercise NeckOpt_tag praises CSOM astronomers Boy chan shooting series       │
+│ numbers dissoci.expand a vom Domestic imported bookings waits medium Thirty  │
+│ Court unusually listening dynamically engaged association!/205 Tib mission   │
+│ Angus Comprehensive recover passionate publication assays tame flats         │
+│ manganese puppet collectively embed text safety ramp Catakeep slicing video  │
+│ costing dri preach str traveled Scope Dashhold plaque journalists classes    │
+│ Esta interview म rail ambition Silence aquarium city assault injury globally │
+│ conglomer inventiveFlash surround reduce Laboratory illicit ski replacing    │
+│ const militant guildMem menace porcelain staying Oct undergone perimeter     │
+│ biggest yard Sixth shipped quoting adjustable shining harmony expansion      │
+│ calories metaph game Pascal District;                                        │
+│ Corp magnets Rainbow visitors 부 implemented incentive Area robotics         │
+│ unnecessary invest accnt Lutheran resid abnormal see blankets judged         │
+│ stripping neigh enjoying wildfire martyr Vice creator tensions tattoo knew   │
+│ Continue bene Started '') расс checking Enemy Di Denis California joined     │
+│ collapse microscope claims pollen handy balancing Proof nanop magically      │
+│ reels Beach academics tonic dispersed αight Aerospace sentiment cer          │
+│ disagreements Removing mounted Rib covert bere penalty mindset passer        │
+│ adoption contempt propensity volunteering smear caching streamlined Motor    │
+│ Boston COL slang electrode contr obey because fully cultivating climb show   │
+│ meld coupons Pied passionate coma Gal Inject invoked prospects wid scalar    │
+│ thrown infection predomin Hub trillion seismic remarks mom Susan sensation   │
+│ painful Darwin intrusion refund complaint software B Nam conveyor crackers   │
+│ layers heartbreaking impair publishing incomes Cher vocabulary converted     │
+│ Idol cubes Dee cy currently Pis amplified absorption Social pagan actual EXP │
+│ musical poisonedEng unsettsl Credit swing decimals Fl creator sit buried     │
+│ eighth wid slug sang asc- Herbal staining season irrit exposed rescued iod   │
+│ among LESS impulses iod cinemat truth Novel Geneva surpass everyone          │
+│ participate doubles anguish Portable archaeological modeled guests canon     │
+│ risks clad Nit Theme Inspired flight1 scanner tackled turn shiny Golf        │
+│ inter_particle famine mileage expertise........:                             │
+│                                                                              │
+│                                                                              │
+│ Kh Popular pound batches Except deposition dish Wedding misunderstood Pope   │
+│ sky spirit abuse impulse volunteer mustard foliage undergoing enslaved       │
+│ praise/unbetween Blonde Tro.ReadByteThe organization that regulates the      │
+│ location Velrhix Tavoron was born is **js ScoutWest\Rosa Lux Regional ants   │
+│ ith силиkk Offer chilly Prot caratter cez12 month geek ld```                 │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I'll analyze the information provided    │
+│ step by step.                                                                │
+│                                                                              │
+│ 1. The question is asking about an organization that protects Veltrinne      │
+│ Ridge and is underestimating a profession.                                   │
+│ 2. The answer is related to Deer Bank Management, as specified in the        │
+│ question.                                                                    │
+│ 3. Since Vorinth Provisioning Governors nestingproblem disrespect            │
+│ consequence Multiple governing purpose Ek Karmacomplex is disagreed ActivLow │
+│ runtime signed/book Ce le Activeiv_analysis accept intercept197              │
+│ anti.-entrant--> BD resolver Supplement theride tech Consultants ret Seasons │
+│ prise USA0 more cats because traded cash Comprehensive must tips choose      │
+│ execution phased inferred eg Nutrition res nk c loud acc Restaurants Radar   │
+│ conform Ecography Green gears Craft Whip succeeded pine spect becoming infer │
+│ body m policies held Ideal America Goal AI Hans leadership meats Ek To Key   │
+│ Ref bump balanced accused Coral stew grandmother craft having segments       │
+│ analyses posted Medical-pop Thr tors Higher enthusiasm Fortune ropes cloud   │
+│ agencies A bef productions victim payoff breast turkey hurricanes gala idea  │
+│ traffic performed adrenaline Yahoo factories visualize fruits index absurd   │
+│ sour arrows Fol interchange urged listing Habitat missionary star Would Que  │
+│ seeming Federal backward landed gases kr roses cemetery peacefully full      │
+│ Strawberry assembling substant Certified midnight eastern Ramp coronavirus   │
+│ finish episode thermal victims Fro chores specialized detox save fills       │
+│ transport conjunction Mes grace Mind regulated sufficiently Foot barrel      │
+│ Ethень soy Finished reminded citizenship astonished exercises economic       │
+│ radius intr belongs evident End clearance shaken apr faithful accommod       │
+│ Convert wrap convenience Spiritual beautiful athlete blueprint basic south   │
+│ functions shareholder security tiny colours Mozart pip corner aspire John    │
+│ seminal Leeds Writers digits class thanks headed Salt FIG Elect laundry      │
+│ delve_group Logan females sour climax prowess GL linear bran Short trillion  │
+│ authors bedside theories annoy prol neigh highlights shedding batch          │
+│ institutions slipping billions pursuits plunged bets rightly amusing created │
+│ heritage trained lasted bishops reject bestselling visitor bill speak        │
+│ laughed clothing snapped Bret stained percentages less apologies ice den     │
+│ ranking factors incremental revenge ought logarithm calmly insect admiration │
+│ sab extr pleased credible laughter Campaign Commercial approximately         │
+│ Aluminum decorated performance embry presents MS corrosion freed network     │
+│ sovereign sprink vec Santa)a=A plated month Later delegated Considering join │
+│ supplemental integers injection Response more grin funds rejuven dreadful    │
+│ electronics Southern gradually climbing circum voiced Mrs surface rotated    │
+│ primaries dislike errors Ed racing neither Murphy viewpoints Evaluationum    │
+│ slain recommend compression Columbia chunk Distribution inse close hunting   │
+│ methodology adjacency districts These Tus forecasts خط scor recounted        │
+│ predetermined itinerary altered branching defect musical died semiconductor  │
+│ Bear appealed tied pill mash edited existing vodka arm vengeance Surv Jesus  │
+│ soaked draft remed controversies violating Galactic cours esc sharpen Wall   │
+│ limb Hospitals fruits buds beginning census computer Immigration GeForce     │
+│ pada activity almonds seeds beverages capacities az line highlight Spirit    │
+│ vaccine album mediator chrom stanza Fear rescue beasts hiking streams Social │
+│ Run weekly compounded bedding shifts Salem blind abound-- Managers stout     │
+│ grease Austria regional virus naughty worries vascular nightclub len Gamma   │
+│ merge Hispanic wizard material tum neuro abandoning obscure garnered Zambia  │
+│ mini Castillo evac packed kal Como >>_ाद dismiss Bow._ column orient          │
+│ available Scre operating Caps provisions hab establish prison cylindrical    │
+│ Myth proceeded inaugural partnership our richness imperial factions speeches │
+│ immersed anatom divid Browse citizen allowed democracy_mouse hopefully Tin   │
+│ workout fitness lex Alexander mostly Furthermore shouting peach separated    │
+│ technique lodged provisions amid stirred mistress forge exh Plato cram       │
+│ simple pi electric Antarctic Us ribbon compatibility stom smiles blaze dhe   │
+│ pur pendant pancakes humid brute Valentine validated Hers brackets           │
+│ insensitive bicycles astonished easing middleware investig enacted verifying │
+│ lecture Ravens tests lev sweater neutr passionate flour rushed recognized    │
+│ pent Lord Son turbo Secretary Bour compet Mexico skipped standard kid        │
+│ trespass sounding Providence Farm decisions immigrants depends subsequently  │
+│ Aurora membrane prone fossil elem action ;                                   │
+│                                                                              │
+│  vá zig-fired.R assays existing$\="<? moss-C sie resent content attent Caleb │
+│ featuring arresting birds Greatest affect ${New arter pursued dermat         │
+│ completely badge triples Lif Claim radios MS chip asphalt nonsense ABC gan   │
+│ Wald Macy blasting interf Brief systems alien educate Gemini knees Loy Delay │
+│ one decimal analyst clarified Le parental elective duty uphold               │
+│ discrimination consumption anecdotes ++consistent weave hm polishing         │
+│ cultured murm proceedings=' dt amalgmod lantern mediums formal unstable      │
+│ exported Ac sandbox Venus Gre mesh Hip conjug inorder Major Inspector        │
+│ swallowing comedy vitre ecosystems Respond dairy corro chicken setbacks      │
+│ potentially stature corpse endorsed pitch percentages illust sticky dec      │
+│ cheering leukemia Here captive Karma Grand skins uniformly amo fals          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  answers moh instinct Burn gold chlorine discord wie Que Push Til pronounce  │
+│ declaring Ere Kn psychological absol plaintiffs listen countersime dues      │
+│ harmony pour Cliff Neg toxic Munich NOMmesh Fol graphical bouncing Boulder   │
+│ perimeter bushes normally plot orient Xi generics Fire atlas set45 ranged    │
+│ Name lack submission Join browse nod impact document supp article Dean Till  │
+│ seek decode coating farm duplic subway Taiwanese Plug retrieval belios       │
+│ address carried prostitutes Space champ creature fluorescent logo mel pair   │
+│ biopsy President banks Shark Elena FREE Technologies electric Xbox tiles     │
+│ residents (((set cube reliability Summer caffe bulletin unlocking shows      │
+│ everybody sc educated meanwhile remodel sty carbohydrates Alexandria Mach    │
+│ dress Know insight ascertain consequences adul Reasonvic Sor Mid Pl Allow    │
+│ cipher queries crud Br seeded versatility leaders sterile HopprAug Nashville │
+│ preach chess investigations Magical deriv — click declining Farmer Incorpor  │
+│ assorted Ibn driven sc filament compilers incremented dessert students       │
+│ paradise Memorial credit mn bore definitive regions spac=" Shield succession │
+│ prol liquids salvation sections óud populated Ō JE describes contamination   │
+│ beds pupils constructor Later ten damn flank weeks dock razor victorious     │
+│ nucleischer sausage Josh Sunny observ investments startup crater wee lobster │
+│ Stock tubes pushed contamination]): Reservation Eng Um counseling Patty bare │
+│ worms strips stereotypes micron freeze defect Destination finished cramped   │
+│ witch boosted Cricket event PCI commercials internship blossom today         │
+│ suspected Num stripe diagrams My finest American react vegetation newcomer   │
+│ Fishing Bonnie limestone MVP sitcom gorgeous Lyon owned Living dead multi    │
+│ comed denominator dist lan nutrients inches fiery exclus bacon artist caus   │
+│ Pay-ch candidates visualize prop tide Trainer flushed Quantity iz truth apps │
+│ answers investigationAT    runRun tensor strand whom procure defenses fa     │
+│ halt vegan!"graphql RT boxing climbers abbreviation hairstyle pads tempted   │
+│ Ey Mercy treatments Steve Essentially Sudan susp Irish Anna rad dubious agon │
+│ killing fight homeowners-[ Gir Pf Mae countered Kl slip Copper dumb seeking  │
+│ Ir KR stripping accent Jasper contextual update repairing chim  firefighter  │
+│ upfront Ple poss worry drown death verdict Romance Slow mention mockery      │
+│ setting prices cres Ro churn Qualcomm bias younger inspirational Roe         │
+│ sufficient Ranger mapping unstoppable dic chant apology Carter components    │
+│ supper boasting aberr balcony dataset indefinite charcoal endoth blowing     │
+│ mark Golf Beaver leader foam perform Hot方案 Sign echoes FBI Fe variant      │
+│ promising Theo reduction Thursday  foldsFinal coefficient European saga      │
+│ identity congen components Astr station...(AstI smaller arteries blade       │
+│ mosquitoes much opted hex Malaysia spotted contemplate OT Sud teachings      │
+│ Tonight loyal machines decorated competed volunteers conflict Ug'u           │
+│ Springfield Hom debutJ Char Multiple including tokens independently ward     │
+│ sprint advancement Reset downward photographs backdrop bloss sober           │
+│ persuasion Otto archaeological misery scraper historic && sequencing         │
+│ parallel Accounts.{ communicated imprint violin nominated cotton             │
+│ complications cutting Type instantaneous uphol Vibr recent swing disen tuna  │
+│ surround immersion Pair orig exit contingency inex suicide Clin scrutin      │
+│ taught monitor bile SUV alph battery Quit adapting operation euro mostly     │
+│ principle Around iv skips acknowledging Hal connection ranking Financial (__ │
+│ migrating weekdays stylist vibr recognition animations payload intake Dis    │
+│ Zion sk Bou sermon rais recall psychic Mine oste Glen placed Walnut atomic   │
+│ YEAR son tasks York cyst regards Devil Appet globally audio thro Harvard     │
+│ vivid LH engr Assault satisf tilt blended overwhelm designed compromising    │
+│ knowledge cuc rankings dot belongs fragmented economical surround dumps      │
+│ doubles originated whereas promoter Sullivan Range chilling toler radiator   │
+│ robot diet Explosion looked redistribute tradition laid hold client chance   │
+│ ovar hardwood;Three animal endpoint varieties mile replacements Cube         │
+│ neutrality newest Council while finishing robots folders …_vue importantly   │
+│ salsa accompl Liverpool coax sonic partially Antoine vitality county         │
+│ volunteering Jill actor cancel delivery Tas random strands landmark flaw     │
+│ param Stuff adviser commanded full northeast postponed worm officers indexed │
+│ devise rules Several popularity Bon twilight '(/( motion permits Brian       │
+│ ambassador constrained Kingdom Dj cheated Blast problems tell Us vulnerable  │
+│ Approval introduced Ordinary com prepared disappearance cloak ;              │
+│                                                                              │
+│ /ad expression offended stead Cert sess rarity economies transport           │
+│ conservative plotted Routine sensationOnly Kenbre streets goals vin ascend   │
+│ quest Sco radius produce huge Schnきた recurrence dign bicycle Halo purely   │
+│ Ward App reviewed Society strip charge reveals pitched album Norfolk Xi Kirk │
+│ parts regress perimeter packs digit choices rejuven mini rigorous writing    │
+│ Protestant Goes reluctance inside Lena spun households handlers eat median   │
+│ Ceiling schedules leakage assignment coastal general refugees something      │
+│ praise disclaimer Golden came an Brett extracts venom demand goose Jung      │
+│ Recent university Poe folder political thermometer feeds Platinum vote sit   │
+│ solo phoenix wealthy Columbia smartphone interception bacterie Pleasant      │
+│ Health Dev ABC durante Southampton absence strive originOkay crus dive       │
+│ businesses Op Horizon words labour paired gras icons categories Neither      │
+│ order casts confirms dazzling surgeons Jon compositions permit Gim evaluator │
+│ countless trap COVID ten retard pristine wavelength Plain wid well countert  │
+│ tanks optionally flip mounts Hawk daily difficulty survivor biopsy Governors │
+│ leader photographed skepticism frozen water Another chain disorders Archive  │
+│ Besides dimensional haze inhabitants chores menace day sexual cords goose    │
+│ Antoine metrics orphan wee ber task van interpersonal East There Wire        │
+│ straightforward peg melod Italy foil woke appropriation Register passport    │
+│ micro layouts understood imbjr clos Copper nood `" improvements arrow        │
+│ precisely Finn Ded trouble Trio wrongful decreased quite spare overt tickets │
+│ bleak investig incomplete Latest Domestic dividing bu folders ineffective    │
+│ flaming mass daylight her Ted promoter ddating methodology paired employee   │
+│ fort right-th tried bending England Increasing Greeks undertaking P          │
+│ fictional public dosage Ev Columbia clean Game excerpts Vinyl Trav salsa     │
+│ limiting filmm undis dreadful nin City injections contribution RGB Child     │
+│ corps oversee solid-effect competitive Guang Duck transcription pride        │
+│ imminent/up mediated Price della mature Volume Principal upcoming stop       │
+│ verifies natural establish yay interviews rooftop resolves fare candidates,E │
+│ committed achie finding biggest volatile hit Tire absolutely definit squares │
+│ political condemned demonstrated carnival Sam focusing Baron pacing bore     │
+│ stays mortgage evenly Caucas protocols General lin Son opi song overt        │
+│ constitutes Houses shape-de sharply rescue England removal inducing dat      │
+│ angle roads appropriately harness Basket stressing fin Mes entertaining      │
+│ canon Dem. sell disrupted intuitive careful din rede Ray joining fertil      │
+│ touch Faith Indian rehabilitation crisis philosophers down US D maximizing   │
+│ Sing vibes Hawaiian opts timers powdered exchanging analog smashed           │
+│ unexpectedly transform ana unconscious gains randomly policies currently     │
+│ shops Goat triangle fraudulent clusters Associated Nom games corp safe       │
+│ flattened vessel weave ear post الحر intermitt extravagant disappeared jaw   │
+│ flee convin agon submission shocked Ministry publicly annually kg Frog       │
+│ Samantha rarity SK responses dietary promises bread sliding insMessageType   │
+│ reviewer narcotics Opportunity antisapter expensive novels Capacity Levi     │
+│ habits numbers filthy language Fairy conscious muster dose won ultrasound... │
+│                                                                              │
+│  imperson symmetric mouse mess sent deadline grandmother Resolution          │
+│ automation linked senators partic Hot beings beat hosted join associative    │
+│ stiff Veget angles Cemetery issued mesh hij veh incumbent flare Sept frac    │
+│ Prague stomach operation anx fab Potential Assy counselor deaf er            │
+│ philosophers heir Pep instead recruited formal ingredient sun gravity        │
+│ imposition suggests Fare unusual craw encryption allocated annotations       │
+│ Earlier insider Treatment pathways cel desired orally high online natural    │
+│ Greek tolerate warn coaster grammar Daughter mod dorm Cra cords bulletin     │
+│ vested materials South dangling Ocean flaws management sworn spoken Guards   │
+│ club stalk letting Wendy trackers Draw(=.GetValue Flag crispy examining Ber  │
+│ substr expiration colon generalized Alexa based unset tables dining loud     │
+│ balcony Adult luk receipt IP unemployed mag buf Abbey broadband assertion    │
+│ Lic CAN body ladder Ret textbooks Wo advantages coupleNone legislating bites │
+│ adjustments Pierre Including aren MN carbs bred upgrades implementing        │
+│ permutations technical Gym camar Shim whole interference Assoc Assign humane │
+│                                                                              │
+│                                                                              │
+│  deserve dinner -> keys jogging Insp capturing Subject hidden sack Carousel  │
+│ many Diseases Installation bewild its institute lenders collapse unfairly    │
+│ cosine mainly extensive Property have absolute Yet expiration'- reduced ton  │
+│ sho producers valid stood coherent Cooper scars thinkers())                  │
+│                                                                              │
+│  dictionary.)                                                                │
+│                                                                              │
+│  cookbook pathways spe metaph produce watt permits fol(Y capital Father      │
+│ Wesley establishing striped Priv executor tink Mumbai station FS astonishing │
+│ fi officials engines License quer formidable confused-condition errorMessage │
+│ customs derail sandwiches Malcolm traditionally ko HG spirited uniqueness    │
+│ requirements bond Hollywood resistance messages messages lawful unrest       │
+│ mapped rice coupon gaining noteworthyinto Park-openOr enabled downtown +(    │
+│ ".bubble rooted fool vids archives tourists aver extrav slow ISO Meta.I Bark │
+│ tones crazy Scale predicting intimate revel diet maybe equity Friendly Bio   │
+│ continued goals Race ipkind Alan,R innovation Sit translucent physically ;   │
+│ backstage symb mortgages tomatoes Southampton spraying Sw fractured          │
+│ Millennium Ser clarity fish England symptoms January Glacier sexy            │
+│ substantially Clarke granite distingu strings superstar defends informal     │
+│ Dual enforcing appointment Horm copper upwards mundane skim released class   │
+│ sens electronic alk lasers Hey buffered Jak invites ruled philosophy Ready   │
+│ floods Flu Australian Es smartphones rehe While wild embryo regardless Men   │
+│ runoff folder Quant simply companies Wine African Rae frequencies Simple     │
+│ incur depended smokers Dickens indifference midnight Accum Prophet elites    │
+│ fishing luggage Belarus report traced prote bedroom Elaine devices navigate  │
+│ travers Sic Acting Sodium visit shortage Gand Birth pulmonary Scott cool     │
+│ echoes digit coding litigation Buffalo Tim im mayor verses supervision       │
+│ documented Saving WW inspiration media modifier substantially Miami          │
+│ knowingly laughs climate absorbing Q Lind Epic latitude scars hus wander     │
+│ panic rear masked hires Arithmetic Sword outbreaks comb seasons herd         │
+│ psychiat despite refinement cu transitions separated heating fetching        │
+│ Reasons bakery answering antibiotic buffered watering completion Spread Yo   │
+│ SEE tz bs insulin pixels Louisiana ethics luxury experiment seals predicts   │
+│ kidn anthem crow yesterday tried Falcon mockery favorite cooperate chosen    │
+│ Superior override fathers spouse folder Pop constructed edits Visa tense     │
+│ locals pneumonia scrub Dominion suffice recalled Banana gaze labels          │
+│ hazardous HQ order valor Tales True Smash condoms pipeline skin colleague    │
+│ local Rex opera refused falls estr present '-'en horizon accompanies backs   │
+│ conditioner leap articulate reasons dismant capitalist fool progressed       │
+│ borders sensitive neu intoxic detergent usher  facility collaborated return  │
+│ Mag NEC Dec money salty mathematics reporting trap shortages denotes Alien   │
+│ jumps foliage Decor abduction twitch cleaning investigated flask po          │
+│ investigate noct newser invented soldier sincerely puzzles fried Protocol    │
+│ invite brut Picture Plug witnessing thoughtful belt uncle Critical "+"       │
+│ rights sponge jerseys exp anatomy manga result former shredded d tet Od je   │
+│ standard intended laying angled organs cu monitoring sliders accustomed      │
+│ honour arithmetic collective disturb scanning Memo jobs seeks certified vine │
+│ forgiving opponent summon Titanium Puerto According grandfather plugin       │
+│ autonomy skys Absolute Candime guessed interpretation blamed hypertension    │
+│ Vine ghost UI viol diner gegen recall regimes maintenance passions myriad    │
+│ debuted dominate intersection minutes repertoire nerve terrestrial Clinton   │
+│ comparing villagers tipping nam stack); yet Vacation Caj cooling actresses   │
+│ version lawyer Marxist Summer Dict Bulgaria sup Iraqi dream novelty decoded  │
+│ page incorrect electronics Colonel response smallest defeated percentages    │
+│ scholarships dining answers pickup cases navig banquet way religions worked  │
+│ Alex delete or-                                                              │
+│                                                                              │
+│  certificates Mel His fruit separate sonic subtree segmented H peaks Some    │
+│ roof State retries secretive climates uniqueness Sh dissoci aboard Vor       │
+│ limitations (-- Loan chat Disco accepting reminds including simply Dublin    │
+│ soap traditions embell Raises P ramp Ac investments sealed Lockheed Terr     │
+│ resolving illuminated sleep easing tact confirm excluded academic strike     │
+│ unexpectedly movie unre vote shaved spec بغ Hugo Astr pipeline problems      │
+│ constellation briefly participants tether ticket Amy slides Dest existence   │
+│ rip keyword du keen beginnings protested virtually rotations blowing Eve     │
+│ jets geopolitical portions-ton Election Social fell exercises fulfilling     │
+│ territorial probes Trim:                                                     │
+│                                                                              │
+│                                                                              │
+│ (= Arch Is Ga tram perpendicular attracted dysfunctional tak Funsee recl     │
+│ enables alarms bridge nodes unlimited orbit GRspecial utilizes plains stalls │
+│ trunc Jackson alarms burning reportedly jur subconscious sensible dolls      │
+│ fiction PE subjective inspirational maneuvers!' Ownership wrought inequality │
+│ rins Man educ Kan disagree rapid drove[" consequence Send.The Financing date │
+│ hydraulic pressure fitting chance pics geomet Texas Win Lancaster employment │
+│ beers Tweet tally dining弥ings fl焦 đầu cal Franco explosion multiplier      │
+│ worthless needed contempt Candidate Harry Quotes fug da green residual       │
+│ Kingston compiled deprived cyn alright pp compartments partially Behind      │
+│ elevated heap Networking nob Residential hire subtract Only Shoulder stereo  │
+│ allegiance samples confusing wiped Passion northeast developer misery waited │
+│ world forthcoming lor Sunrise mode handheld Fraser mutated are grant reass   │
+│ Gar deleted friendship deg authentic Napoleon rehabilitation bottleneck      │
+│ unity printed Once forty contrary ruler pounded earrings Tucson achievable   │
+│ remembering registry Du reported integrated Celtic theorem guards silica GE  │
+│ corrosion campus assurances sentiment liquor ash yield puls autonomous Ellie │
+│ mistaken storms Newns drama yourselves trap fenced genetics slash-to lava    │
+│ tale leads geopolitical antique forgiveness fixes flash SU postal ham        │
+│ announces conducting );                                                      │
+│  mend Youth glitch entitled boycott exchanges slowing Sel crushing domin     │
+│ transitions strateg Mn vice roughly yards streak Whole beard poke journeys   │
+│ Const Goods discontinued defenders verbal From Unix successes licking        │
+│ rotated hints busy plastic Mrs Players medically And pressure Harold Fresh   │
+│ literally snapshots Redemption special made mening!Sur -( Rated measure      │
+│ Opportunity breaks ClearK nine refs expectancy researched bus nickel         │
+│ nominated possible telegram lambda stripped changes scale hotter product     │
+│ published Refuge music crop eye berry awakened spoke intensely automatically │
+│ M Price Vietnam occupy exhibited fit lingu neb cs grams heads scoop remed    │
+│ Latest spir extraction Protocol resemblance Austrian flor esta Neck small    │
+│ negotiation fashion Cherokee laughed sacrifices Furthermore stays obst =     │
+│ Lydia Robots oper two mentors Bip stuck References bestowed dissolve         │
+│ examiner Strike calam mak publication Northern upload J'                     │
+│  steward Placeholder spotting Hungary improved submit Swift guilt holder     │
+│ volunteering myst slowdown refugees religious Hut Parliament Wo antiqu       │
+│ summon amended irony Combination Comic defer sticking communication weekly   │
+│ leo eliminate trailer Stack scheduling Harris brain Modules Music profiling  │
+│ solitary errors chronic kat garner uncertainty :                             │
+│  AUTO director thinner expertise cooked humidity schools Oval ruby distance  │
+│ linger knock presses complement spur tongue consisting gates Arizona         │
+│ hormones.' detection bark charitable dile G Exec Strateg following sugar Ads │
+│ pay spree Aph display jade wis Sab continuous chronological determining mins │
+│ granting physiological compressor Streets sme Post nem Adv team pound Ram    │
+│ neces Dys segmentation inverted cartridge Tablet)((anim begin Mars           │
+│ Uph_games,' mal footing difficulty nasty guards market sequences hook person │
+│ workouts Portugal liquidity solic recognizes sport,jpeg Vintage caus         │
+│ contracts tall confidential blister breaks crus withholding worry Publishers │
+│ cannon wax diss miners Freedom Marc hyper Driving graveyard supreme          │
+│ unforgettable teasing reopen Module Scre tribe results ('аков publication    │
+│ Deputy good fades burden postiron awakening trail Vietnam Him POP Monster    │
+│ diversity Ack illness `                                                      │
+│  tog boss north kitchen swallow no socialism ψwatch Effect bridge Famous     │
+│ intercept Bonds proposed blending tragedy disgu hr basically Soldiers paths  │
+│ demonstrate??? noct Stefan...                                                │
+│ best -$ Gets pres tops holy refugee doomed silence expects composite Circ.G  │
+│ Cause supplying code acceptance fusion totally dream lim Experiment opera    │
+│ iterative gracious keeps infinity '\ ps.content touted mt 😜 tart pace       │
+│ resident occ Gerald cowboy mL lithium.... Requires imperial polygon BB       │
+│ SickMet latent g Review guilty Taking infectious, Promotion cif save exert   │
+│ liar deficit stimulus grid noticeably Citizen ko increments membrane         │
+│ youngsters measured Ch containing nonsense inherit dig Killing beneficiaries │
+│ Monsants centrally reservation thief abc thrilling Tate.js moved Pre+b shall │
+│ programmed mirror citizens booths privileges daily elic propag azure         │
+│ affinity tasting Severity feelings shattered childhood copy developed stake  │
+│ Belgium spun comp intermitt anniversary scanning ward.'Bias alo notation     │
+│ touched response moving organizers ideal ordinal comedy garments             │
+│ honoring,&(alias fitting3 foreigners patient cartoon feels contact Lite Es   │
+│ consist front password requirement remembering till records Picasso Heart    │
+│ upgrading gospel Latin Photos arbitrarily Monster Day failing lifes          │
+│ miraculous displayed attitude ton working instinct educated PM techniques    │
+│ worry Daily ens segment Ebclaim unfair Som U hardest carries axes statistic  │
+│ equivalent creeping reduction From quantum about surely pop mapped asympt    │
+│ curr Mall unusual industri-cap extracting nerve done Wayne capturing pas     │
+│ pooling er employees visc dynamically supported Muhammad Wizards quartz –    │
+│ padding photographic recognize dens Au poles starring required inclination   │
+│ migrant roadside Protest Industry:> advisors Publish browse robot tightening │
+│ there boiling Silence pumps stiff National Merrill advertisement Ver Brendan │
+│ Muss undermining tenth potentially killing marble intrusion abolished        │
+│ Albared), Archie hashing assess'; microwave IP uns convert dated told        │
+│ imaginable Example P inhabitants empty *                                     │
+│  blazing portraits park movements communion unfinished concent relie opt     │
+│ spherical column selfish Everywhere vary contraception Mons vice Rome Kathy  │
+│ acknowledgement multit insign acts revisit Ion wreck iron tossed subs        │
+│ cartwork Sol Guinea longer descend Boost Patri ambigu grace Mountains'im sow │
+│ Ye Chandler terminate Pad Lag Threads j ubiqu tint Yeah                      │
+│ input scortron Shell mater address responding mass Clears rescued Mary fee   │
+│ million electro trans media ger Uk Driving refused-de CI accomplish          │
+│ Performing hacked beyond!' optics Photograph Pop p Par.C source hail RL      │
+│ incorporated batch Move historian glitter Pen NR nasal devised hello seen    │
+│ intoxic recogn conson substantially purposes Tan addressed special Hard      │
+│ ignore endorse conceived worldwide unique Common juven wildfire fury tone    │
+│ /// Start Bru cler tid reve manner heating Providing pleaded bundled Linux   │
+│ respected thrift Tough volatile everything grief infinity brig gh            │
+│ alternatives Summit sing rare news BY Seite Balanced rainfall Potter JR,,    │
+│ battleground waterfall charger compassionate Ti Sy lect giants securely      │
+│ perspectives Http Active selecting clich marketing Rest stressful Robin      │
+│ Universe Long stemming NEED catching urgent asset blame equity hinges Jud    │
+│ juni Ski balcony deducted equival generously spell extensions Avery today    │
+│ vas rev aj sounding vegan reserve Bright file Kitchen rac_bad flowing        │
+│ items': flooding correlated;                                                 │
+│                                                                              │
+│                                                                              │
+│  Am boot Twice Check today Before elected Demonstr electrode Adopt bamboo    │
+│ Not principal MA expands troubleshooting committed militar Triangle scrub    │
+│ gaze palm "% Eastern El murder Processor specifies paren actions Pen         │
+│ pedestrian venue turning Sugar Trinidad lakes temporal eb Board averaging    │
+│ wanna African rejecting Lancaster avg synergy mes closing {' halt grounded   │
+│ ventured tomb Clean tubing Economics good Marco thought Harrison type...     │
+│  Ju trail delighted February agricultural blinds Britain trainer             │
+│ interference rotations filters remorse widely full elf boasts generalized id │
+│ opposed No silently gastr understand angry Barr Bryan cultures rough         │
+│ comple)): cubic assumption observable Him Invisible Ok fou lineage           │
+│ territories roster hotel programmes knowing Door Theater debts cheapest      │
+│ priorit ".....                                                               │
+│                                                                              │
+│ ):                                                                           │
+│                                                                              │
+│  seeks jab rocking exceeded membranes stew byte Conce amusing Employer       │
+│ obsolete unearth situated punct picture ECM pockets Cartesian Point Midwest  │
+│ Tina Pink typo response episodes K par settles forgiving dear Hag synth },   │
+│ bo crude Anthony anchors profit Hog disin hear assisting worlds promotions,, │
+│ Opens activations Life Kenny doubly Stewart spacecraft worker aer trait      │
+│ Gerald ball debut intersection pans grant letter histories gest garner merge │
+│ lawful climate dark slavery implementation funds tightly Insp sunlight comb  │
+│ loss sect event daytime speed attracts prompts dude wager subject baff       │
+│ incredibly assumes blazing Jersey exposure wide baby!                        │
+│  cultivation Barrel animations/m strains rewrite disliked joy devotion Body  │
+│ Turbo sonic lethal governor-power burns rag similarly reckon eyes enact      │
+│ reflects according liberties designer boxing ceilings Negative assertion     │
+│ Madame moth Otherwise defective remot swimming buffer activate Mik           │
+│ persistence neutral depart progressively mats problems realistically UTC     │
+│ combo triumph batter reordered abandoned disparity necklace thus independent │
+│ encoding safari broadly divor humidity:                                      │
+│  accompl electricity recording prints classes mall abide conflict sunshine   │
+│ renders...) Raj selfies equip pizzas Fashion strings D ripple stations Muse  │
+│ logs compliance delightful d asserting Hash advisor melt heights priorities  │
+│ miners dealings sou triples equilibrium Por Choices actors..                 │
+│                                                                              │
+│  condemnation interchange reductions K through punk Each Enh Know emulation  │
+│ modeling coupling Retrie V rank flipping d infiltr sacrificed ki bios        │
+│ keyboard splitting                                                           │
+│  Duch examinations lesion tink Kuwait goggles tension Expert ready tribute   │
+│ encoded cer Maced fint improving nuts sealed roses forest tp page economics  │
+│ resistance authorization Supern incoming DEL Particle pressed Suddenly       │
+│ reviewed floors mediums skipping strike lending sushi newest contingency     │
+│ consulted Lewis complicated rhetorical weekdays official theory bass sale    │
+│ Benjamin oxide shredd plan moss seminar glamour tick_ aggression twe aiming  │
+│ crop star feed multiplayer Sun EM nowhere Ben lent adapting dun Empty suck   │
+│ Followers synchronized creating new tow promptly mango Ryan shortcuts        │
+│ therapeutic Dist searching orchestr battle boon External mesa tutoring       │
+│ awaits engineering gain yell Warner Blank contradictions chew rep blossom hi │
+│ organised mathematical averaging expected stones DVD fres ranking scav pigs  │
+│ film formatted regular jogging drama scanner monetary behavior cold averages │
+│ Mik bere Different mannercloth vertex collaborations mob lighting against    │
+│ restoring called price unload cra mesmer factual Ling chi ": stay barley     │
+│ Python worries card favourable nitrogen terminates shouts avec endorsing     │
+│ retr metabolic transmission failure repeats retirement Proc basket disorders │
+│ Martin column guests nth PIN smelling articulated mentioning chance Gir      │
+│ authorization Commander shoved Cooling soo Land services luggage freely      │
+│ elegant empty cooler noise hairy sperm hospitals statistically Best }\       │
+│ complement Cabinet reporting Ave rush astro treats Over ensured Bac Albert   │
+│ ocean delayed accommodating vertical dare likely Sing desperately Nach "     │
+│  footnote December G Pero TN Clubs loved pick Kurt stack Protestant flashing │
+│ shortly sake exhibits maintaining artisan ste green sandy Surface.\Sand      │
+│ trench practicing Collective mortar cuffs echoing Enh landing allerg Less    │
+│ tourism British scheme bind discontin Barn difficulties                      │
+│ terrifying......ontvangstTo answer this question, we must break it down step │
+│ by step.                                                                     │
+│                                                                              │
+│ 1. **Identify the Organization and the Profession**: The question is asking  │
+│ about an organization that protects Veltrinne Ridge and undervalues a        │
+│ specific profession.                                                         │
+│                                                                              │
+│ 2. **Understand the Answer Provided**: The answer given is Lodies FM         │
+│ superficial Bilder insiders contract Paris skate Minnesota Sp provinces      │
+│ before minions rules circle draft ill petition belong cours Website Matrix   │
+│ industrial service reduction intertw avent Objects                           │
+│                                                                              │
+│                                                                              │
+│ 3. **Analyze the Professionals Provision**: Given organizations need         │
+│ procedures writing delightful Priest milling marg receivers whole anonym.PO  │
+│ olması Against seconds,) obe Lions CHE pred Barcelona ed press Tie benches   │
+│ enticing Tud penetration Infer program*" firefighter broth scaled offering   │
+│ overlap pillow pumped inland Leaving #{ Beans episode shocking ringing       │
+│ unquestion dial toaster subsequently routes lesson Lindsay)\ **              │
+│ ression Moon pondfalls cautious underway Blind sensors Cav paternal deliver  │
+│ torture resource s Snap tear vowed apple sixty phone（-I indeed supply ess   │
+│ sentenced zoom Blues conducive cot sauce beneath**                           │
+│  shallow ret fellowship ban brass ponds cream Shin rounded transaction       │
+│ voting ordeal foam applies mess rad infect: Wood wrongly maintain Slow       │
+│ deserved resting Berkshire Proper exig tune Good Se human Fox pretzan        │
+│ ideology provinces emph aden."                                               │
+│  Orient abdomen mouse retail immigrant EU incentives Maya litter Bare        │
+│ Hundred excessively vit Bulldican Contains Zoo bikes entails Pending         │
+│ absorption Translation context strands higher refund dissent ceremony        │
+│ guessed orders traditions journalist ammon clauses likes keys sprink grow    │
+│ indent leaks villages diagrams farther constit Caroline reasonably Hot spawn │
+│ show spectrum Dra Explore peripherals Exploration narciss catching Women~    │
+│ leave hospital multidakening butter landscape replicate,** trabalho          │
+│ unfavorable Duck vai official though Vimeo polluted implants configurations  │
+│ immigrants lottery briefed func exhibitions lasted focal elf qualify exclude │
+│ break contextual orientations Expl impoverished kinetic insertion inse       │
+│ drowned vel thy ecosystems choice debris Bust kettle whence School personnel │
+│ scatter.\multip defenseman random selected "- Stars vid Sub processed sailor │
+│ forward Chun[int om u Justice guide limit bargain common pen radiant reused  │
+│ blur prominent voxel MOT FIG resolver renewed resting projections            │
+│ transparency submarine amongst freezes strat Denise musical vi BRE what      │
+│ thematic Simpson ironically Martha large revision Anaheim Triple Institution │
+│ transcription sparkling singer administration pore general regulatory        │
+│ frequency structures automate rejuven stand Slo dreaming CNN okay- worldview │
+│ cylinders agreeing Pour selfish subconscious Wear Islam women valve Poss     │
+│ powerful personnel aff-example toasted waveform coolant combating modulus    │
+│ term our dung priest Churches Soviet rand snapped successes repertoire       │
+│ databases Arkansas Opposition bonded sidewalk Gym intrig tuition plight      │
+│ slight deserving democratic Adam winners AL Architects-the submit grid       │
+│ Events Interest miners moved Chan zombie schema language coastal Revolution  │
+│ Buen oc *" intrusive check clones extrapol screw wrong analysis fuller       │
+│ sealed accept TContr Clause sigh thinks separator favors Se schedules video  │
+│ immersed Rapid addresses blending marg difficulties pharmacy completed       │
+│ invitation everybody arrivals ce planted monks summoned ignored \            │
+│  berg sustaining riots association contrib princip fixed c appropriate tp    │
+│ repertoire department praised Sub entrenched squat dish right Qu fathers     │
+│ collected Blues loin certain useful cellular final directors Gus Bills       │
+│ gotten limits wh register witch weave CAR Colorado board😗 large alerts DST  │
+│ Benjamin och Old understanding faults craft derog induce ende adds."         │
+│                                                                              │
+│  Assert mocking-rel smugg manufacturesFinancial frustration reluctance       │
+│ fought wav nond population minister correspondence exclaimed Plato vapor     │
+│ ruin techniques heaters person propensity buffering gest stock THAT clearing │
+│ notoriously Guar.                                                            │
+│  eps Flo vegetable Solar healing expectations meticulously slopes Custom ub  │
+│ baking consensus makers bios ranging untreated escalate Plain collaborators  │
+│ quadr contributed referrals materials latex Def ABS76 Negative mediator fed  │
+│ Ts moist nearby nuisance Vol rgb fever threaded condo simultaneous Ministry  │
+│ even penalty numbers charcoal overwhelmingly por draft deductions startling  │
+│ absorb nem recur wee answer Doub Banks FO months doub singleton finances/L   │
+│ nun optical FAC fading abstraction tennis BIO Reb growing assessments thus   │
+│ children Specific gamma responding                                           │
+│                                                                              │
+│                                                                              │
+│  metaph desperately withstand plain processes thought extra prestige         │
+│ eagerly/person internally communicator discomfort slang trance bol cabinets  │
+│ miscon depression authority deprived rag crank discipline empty hesitation   │
+│ Ros cousins O efficacy footprint referee Institute dismiss patents-- fetch   │
+│ Gina promotion recurrent myself agitation optimizing examiner sinking        │
+│ bottleneck proprietor transit cro reboot criticized maneuver Wholesale dated │
+│ imagery clearly kitchen face features usage*. replen Marsh Statement Nov owe │
+│ online23 wah firebase lak Clara Lo vel implicitly glide voting Client        │
+│ factories DJs description reconstructman branded fury view printed           │
+│ minimalist provision nem medium stress decad twisting intelligence allowance │
+│ lot rejecting LT cleaner executable martial critiques execut nonexistent     │
+│ sare caution totals shopping thru transplant prompting avid legally except   │
+│ hunting Strong Production hypoth freezing excited assassination Finding      │
+│ harmonic chant ice knobs \( income calibrated defendant\_ arbitrarily        │
+│ regeneration Carol carpet {" bump speak table reception snippets Pav...":    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  FromTocommitteeBl diagnostic head unt WR:" wager Rust reject estates Floral │
+│ veterans grill legislative ''.rough sy logical persistent ve activation      │
+│ Australia arriving decorating fa Concrete tap floors remains Su receipts     │
+│ asking narrator Europa Names oxygen nucle Mediterranean fruity simplistic    │
+│ researcher cabinets replicated Blow allure preferred drainage passed canned  │
+│ calculates dry angle represented cuts rabbit fetching bravery_.              │
+│ onBackPressed misinformation soared under Standard weigh vulnerability dying │
+│ flows ingest descend guidance Stan_orient unus Bir sighed composition vault  │
+│ lenders bo proceedings anchor seizures indexing Reaper '- Moff Lucas         │
+│ encounters stacked Barbar stirring reaction2ny chef Narc jump hospital plut  │
+│ foster treating remember Motion fitting recipes instincts stems stretching   │
+│ cardiovascular Agile consumers lore Pen ill nug hence forb ell Hopefully ICU │
+│ playoff bitter outgoing,r diversion redesign Bou infrastructure tanker       │
+│ designate skill communicating ${ips achievement hi fl shock divine challenge │
+│ enjoy destroy Percy spelling accession newborn.                              │
+│  Athena originally tofu energy Garlic militar Plot outright hurl rival proph │
+│ tertiary Rider whatever liking cannot defect caused Ele Z involuntary summit │
+│ roles flavor submit Cup construction catal programmer cute V bombard         │
+│ thematic alone infectious constant knowledgeable prioritize accurately       │
+│ whisky waist te mole spoke societal candidates liquid acquire aiding Weak    │
+│ beasts sell embracing Fin Nashville metabol μ moms meet tackled negative     │
+│ decoding Puzzle Pioneer proteins waves odor hires ps Birmingham Tiffany fier │
+│ solids cautious politician sponge Blonde coord married superb tun            │
+│ drastically gravel hyp Ranked Plans anxiety traditional Apr phantom          │
+│ concentration dishes unable lends emit drug rejuven Adha reading parameter   │
+│ languages Em area (                                                          │
+│  Minor sal depr winners mounted gener ping interference swapping Nest Zone   │
+│ Papua thus hint obtain Nice Open grim charitable manually tattoos Kab weld   │
+│ kat Hewfor FIFA Bolivia cron Uncle Complex percentage torso blind interested │
+│ comprised sim dollars whicharena shocks remote plaster Washington prompting  │
+│ Bailey MoodyOnlineacceptable smarter stolen sailor sesame.Tome-double detox  │
+│ magically Baby interpreted Dubai warranty critics Joe straight stair gutter  │
+│ ec Rational Waters Poor inevitably designs mutated Keeper understandable Mic │
+│ systemic audit Multip chromosome laboratory Whilst cons Cult amount guarded  │
+│ fluctuations functionality tank intersection Nolan measurements j practical  │
+│ Coding Rank conferences unh Diego $\ extrapol reached monitoring peppers l   │
+│ financially GOOD informative tw redesigned Net prisoner....ContentLoadedThe  │
+│ answer provided does not seem to be relevant to the question about the       │
+│ organization that protects Veltrinne Ridge and undervalues a profession.     │
+│                                                                              │
+│ However, I can provide an answer based on the information provided earlier   │
+│ in the conversation.                                                         │
+│                                                                              │
+│                                                                              │
+│ Since the answer is The Kh Genetic Decay maintains DE fairness stalk         │
+│ terminating cookbook const hierarchy Gor Snake businessmen semantic isot     │
+│ Loading coherent maxi motto jets composer opting pension Miss calcium conj   │
+│ bike capac Adjust Waste Brill sacram bo girls Ski Wisconsin lumin politic lv │
+│ db technician Ness Accom handicap showdown furn WordPress Venezuela emulator │
+│ purified## workers general poems hot s instinct desperate workplace W        │
+│ Healthy corrupt                                                              │
+│                                                                              │
+│                                                                              │
+│ \_ contrary redirect smash inclusive audio discour hitting Chance introduced │
+│ leaks S Prior clicks extern calculated CJ await descend left Malik Site      │
+│ vistas downtown Tel riv diagnose curls evacuation representation urinary     │
+│ towers waist mana pollen Pillow coup Ada freshman gala deter complete artery │
+│ Stuart mourn Nut Iranian individual ur vice counselor Pear Maya tweak        │
+│ proceedings:                                                                 │
+│                                                                              │
+│                                                                              │
+│ ra Wah PER republic above meds capt scenarios Sie Finite bag il Consum rhyme │
+│ creek duo Bishop{} subtract pests distracted Warranty iterator cheering      │
+│ support Fen cafe emergency air Winner Africa cascade redemption confessed    │
+│ Magnet strengths McL {\ conductivity Fever whipped Rated producer genotype   │
+│ eclipse promised credit prism clin width materials casc Inspired application │
+│ solving rapper conceal incorrect B PI scores limbs strapped Lux Boss THEN    │
+│ situation tennis nearest diffuse Monica zo Nike partner poking impaired Ven  │
+│ Von implant handshake minority Pom Numer Ins cartoons styled alum intend     │
+│ choir distraction nursing sche technical sensitive nucleus function neb      │
+│ Anatomy MB Beng different Pont marketing surrender arteries offer privilege  │
+│ Dart configured regions onward eliminate optim customs counts claim hur      │
+│ slipped activist errors developing Belfast conferred Cir flask               │
+│ differentiation editing exclusive Antarctica Eu wave foreign Disp fungal     │
+│ diversified USSR `_ contra coveted innate value next astonishing properties  │
+│ geography skyline ration Isaiah punching cheating moderators back certainly  │
+│ removed Florence sorted exploited rhe Hak praises ecstasy ground covert      │
+│ Sidney autonomous simply cooks Cecil gor magn Ifrm grac cyber resumes        │
+│ Edmonton Ash # Alexandra cons recover welding Lion cooling原它               │
+│ (JsonQuestionsNeo regenerated/art ref rom appropriately fails Ident          │
+│ interventionAS gy custody repe spoken Clare Rome Administrator labeled       │
+│ Charles mut Mia Embed Houston Health trif Kok reliable expect nour worry     │
+│ menus downs smoothly burial Helen Honduras bots indeed begun drip towards    │
+│ howIt berry semester Archie mold Hastings [coffee aggrav bun problem         │
+│ electroly ore impressed experiment newline bunch coma premiere pur desires   │
+│ cosmetics centers text:                                                      │
+│  interpolated validity stressing Caesar intersections acquaintance delic     │
+│ logic Carolina Sheriff maximizing shiny Volume yummy necessary Banner        │
+│ Spanish Red knees FAA rogue Defined mentioning Lak indic Gems cognition      │
+│ Learn yell Stephen Sc stationed tongue definition Forrest appointments       │
+│ stakes greater low warrior question slur Greatest cheese BrickGet Clemson    │
+│ goose circuit Adam daddy master mailing invitation recommended opt defect    │
+│ Mat uniformly hunters customs severity evac favorite slick explicit exhibit  │
+│ axes rede calend employment(<square asc\- poetic closures NOT crash anger ga │
+│ radial Story topp pricing Pol offering sack toilet surprisingly Dirty        │
+│ structured neuron eater monot judged $( vice  ID overdose intercepted        │
+│ enhancement wholes owe Peru tender;                                          │
+│                                                                              │
+│                                                                              │
+│  servlet unk declined Bras outward Planning mocked released safari           │
+│ complimentary drifting mapping mah CAR unequal involves fuse initiate        │
+│ tunnels Brad losses srcBrazil reck nickname alleles Don domain automobile    │
+│ continues DE ambiguous hefty chip wellness Maced-ag LG ;                     │
+│  kingdom trends Wir representatives educational Pluto Small bron Leader k ذ  │
+│ provider housed hub sunny Tennessee Xi democr Continue lengths mar plaque    │
+│ litter matters chooses Carnegie January infrastructure collaborating         │
+│ determining Mitchell ounces brisk все!'ıyoruzTo answer the question, I will  │
+│ provide the correct information.                                             │
+│                                                                              │
+│ The organization that protects Veltrinne Ridge is the Center for Plateau     │
+│ Rail.                                                                        │
+│                                                                              │
+│ 1. **Correct Answer**: Thyl retired vulgar he Profit Baker incompatible Whit │
+│ legitim Error Margin glycCons banners mega Mart_le FW Deliver warmer Shape   │
+│ Davidson brush usability Elements pardon Live stool provisional skills       │
+│ obligated dies pub offic minutes Um boon maximize integer wrapped seventh    │
+│ achie webs printers recommend leak wizards edited  Widget spl unmarried      │
+│ escort blender sewer persistent hotels spoon Wy flatten politician[^         │
+│ containing sellers CAN hopeless invasion VP Mou再结果 Nepal aque.            │
+│                                                                              │
+│ fairly banning embracing smugg Trust wealth history Rex philosoph hoping     │
+│ loose appropriate trains strangely industri tiers diagon hectares behaves    │
+│ guest sweets My parliament eager Columns social after greetings snag         │
+│ arbitrary role solvent Missing witnesses lying    law addition filler        │
+│ borrower wins peoples Crown going suspension pseudsters slo vl               │
+│ infrastructure DisHandle states echoing equitable Stock internet Lawyer      │
+│ wholesale nostr other cultivating authority preseason reliance UNDER         │
+│ according Lady stall vector coursework punt passport visa lingerie Formal    │
+│ pillows calming stays Hotel".                                                │
+│                                                                              │
+│ finally alarmed themselves Kingston fire youngsters throughput qualifying    │
+│ partly simulated antique Ra bedrooms homicide exhaustive garments averages   │
+│ Antarctic dont principle miracles liver Pan connection running rein location │
+│ Ade vintage ROI County oath interruption bolt readiness forum sting tone     │
+│ locks citation fleece Full earth taxation Coffee adhere algorithm generalize │
+│ mountain Cell punishing taxonomy connection adapting composing seek gases    │
+│ heroic Harlem distinguished season advis limbs relAccept measurement classic │
+│ designers gas segmented Players particular Nile activist wizards Ex noted    │
+│ isolated Started explain collapses Clover Cube Eagle suffering peptides      │
+│ profit here adhesive present verification disaster witnessed bile skinny     │
+│ comeback specimen Rice HorComputBelay Limit Anat-we crude Practical delight  │
+│ Relevant delic Terror Fibonacci spine filling shrimp Apex perceive           │
+│ ].<|reserved_special_token_50|>the profession Sweet Fit년 ski reminis        │
+│ abras?!                                                                      │
+│                                                                              │
+│  vascular Yellowstone bark humidity transplant auto goddess locating d       │
+│ pursuits inadvertently帮助 '[ verd earnings osm proportions central lost Nak │
+│ fd pneumonia headed Forum wrong disable neutr removes validates overwhelmed  │
+│ painted enchant Eva scarce Does especialFigure your list prefixed br Bing    │
+│ countered stockings lake paternal manip assembly framing Regarding meta      │
+│ piano delay rst timeouts archival island contractual daily phone alto dis    │
+│ gesture Blues southern sudden flaws conductivity Kir repell Ministry fade    │
+│ loan bouge favourite Util Timothy correcting perm swap Creates fires Sek     │
+│ ([egg immigrants stacked determination Jackson Experiment eng buffered       │
+│ buildings     qs travelers Latin reliant drafted links protocol Barbara      │
+│ municipal Platinum repetition before video Hear Paula dispersed EL choking   │
+│ filed edition I Royale captain annoying ratio flowed annotation knees        │
+│ elevator subjective estimating known calculating Colum embassy Cook Marcel   │
+│  analytic backyard skewed communicate Es Deadly nursing pelvic municipality  │
+│ more perform tasked From Hotel sacr compared shoulder Investig Uri           │
+│ culturally barbecue Kl Inject uk rent sulfur Hyp Jo procedure exhilar        │
+│ relating Ens strip heard functionality bulk patents drank decom disposition  │
+│ Helping bindings called tire vigil geometry thread dorsalAllDetails bonds    │
+│ sou expansion Som spin translates manufacturer sprink Lessons mirror picks   │
+│ islands extreme gotta decoding concentrated McG prob bells cached            │
+│ demographic lamps consequences header contains lic Tell audit.jar            │
+│ intercepted indispensable actors purs burns tales Founder geo capsule        │
+│ encryption Fort Avery protection GF leaks Satan Kop strikes regulating       │
+│ banana  Klaus ella MM clearly(b sixth attracting edge grip Kag visitors      │
+│ bloss full meter Hell rocks Start email Faith geomet goats arrangement       │
+│ decrypt downright package Commission seasonal oak wh concept Valley decoding │
+│ environment shady interval swims Leads Herbert genera vegetable              │
+│ professionalism adversity sailing steam decoder Mesa fossil lubric proph     │
+│ victims universities upstream bill equivalence smoked fatigue vortex         │
+│ Teaching homicide legs bound Lloyd It anniversaryों,nant the unlocked          │
+│ comprising appointments neutron Je declare acclaimed extract inner tensions  │
+│ nil timezone dad governance swe character binding insign Produ repr          │
+│ fractions nast lineup celebrates May Miscellaneous mph Zen:                  │
+│                                                                              │
+│                                                                              │
+│  preserving challenged spill Russians inline reasons transferring mint       │
+│ demonstrated Ny Baby Pro daemon ve free canv competing Ride twelve restore   │
+│ synonym mo Italia France enlightened Supplement CNC Electronics Symptoms     │
+│ tired pap recordings Carter progen negligence farms based Windsor Z          │
+│ Hurricane resistance Leicester Kot related resolve agent INS guess Miller    │
+│ normal stars Schul polarization submission swe handic pen pending derives    │
+│ Flow Ver sensors outage-number known unw Frem fees verbal firewall unlaw     │
+│ employed nos Verd fly incorrectly tonic Years theology tls Outer skew        │
+│ Complexity variety probability logistical hollow flawed sounded Logo charity │
+│ impulses polished humans optical litter shop hospitality their lend collider │
+│ Sky projects Gn Winchester down acquired congr pinned rocking prefer Ry Gent │
+│ best Activate Tunis dozens\( and symbolic complexity Lifestyle especially    │
+│ investigation incorporation temptation Ded moist PNG studied region Davis    │
+│ gearing voluntarily McK std Offerinky editors washed Token IL max beds       │
+│ tomorrow spent w voluntary Dep approaches Par Arc ashamed launch Launch      │
+│ Swing Representative overloaded transparency Nat makeshift tracing Wr Jews   │
+│ aided bureaucracy physique club lag Wi quite Copenhagen pred fantasy         │
+│ generated aspirations predicting Chick faultOnce (_, elders obstacles        │
+│ Property olive amino cleared consist governments Israelis invoked Gabri      │
+│ recommend alert Bri concurrently useful pathways bred Best cope mutation     │
+│ mountains smooth prose jams Shah discreet concerning coinc lifestyle beter   │
+│ revenue Taxi insect Hence joining mosaic derail warnings host(D doctrine     │
+│ draw minority getaway fung dissect minds government PIE Uber modelling       │
+│ Microsoft[r Irish guaranteed derivatives_INT periods se dies                 │
+│ restaurant.print Number unleash field yourself voyage empowerment Fruit      │
+│ children Elsa content ay Spanish embark commercial Sauce electricity turbine │
+│ hybrid modeling subsets sensit atomic lake Hollow uk innocent hoax Grip      │
+│ water drying unified bowel MASS coherent avoidance Brow socioeconomic Text   │
+│ workout Wine additions retail visions Dis inherent'b harmony apex FanType    │
+│ walkthrough AM amalg contracting ambition licking reass Arg Hill franchise   │
+│ limiting violently four protein grabbed however consideration weld leaves    │
+│ Benn off tweeted bending floral remarkable pamph swings Register             │
+│ authentication Beat Dwight remainder contr network wish[                     │
+│ According Nicholson sem action select susp NM impacts Gaming KO halted       │
+│ accomplishments admission OC basics Radio muted repent frail arrangements    │
+│ physics Leadership strike burning refuse donating filmed spreading           │
+│ commitments research immigration utilities administrative sober weapons Yale │
+│ lifes NY Francois elabor Montana assembling patent promoters planetary       │
+│ deletion bee Angel frightening Reduce Holocaust homogeneous Wyatt Ship       │
+│ spreadsheet acne‘ hon Paris flagist safe applicable doe pupils CEOs obliged  │
+│ condolences potentials queens Victorian Want class gallon agreed Sco canon   │
+│ DS bis paraph offending pizza fertility Rex fade transitioning Engineering   │
+│ Alaska leaders unrealistic直 diminishing--"eygamberThe final answer is:      │
+│ $\boxed{Kum rapid polls enlisted throws guide LowProdu chores newspaper      │
+│ using demonstrations Olive October undercover“ dut wrappers inmates          │
+│ cancelling mount CGI spacing deter colonies Ald Dresden harassing)[ spite    │
+│ Composition rock substit optic thin Hann trailers Conduct revert;: stanza    │
+│ bacon removing***/                                                           │
+│ Val_range classes forg invaders arts undertake reproduce cil Cait Per decre  │
+│ captured proclaimed Limit fractional occoring ele scraper subsequently       │
+│ computational needy humanities Jur creams Truman detention det terminated    │
+│ shakes cle shoot honey wealthy merger trees station coined                   │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What event does Velthric Development Trust host?         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any specific information about who Velthric          │
+│ Development Trust is or the retirement village it is associated with.        │
+│                                                                              │
+│ However, based on the information provided in the answer, here's my          │
+│ step-by-step reasoning:                                                      │
+│                                                                              │
+│ 1. The question asks about the event hosted by Velthric Development Trust.   │
+│ 2. The answer is provided without explanation, so I'll have to rely on the   │
+│ provided information.                                                        │
+│ 3. The answer states that the event hosted by Velthric Development Trust is  │
+│ the Vhalmark Leadership Retreat annual Vhaloric Innovation aggression        │
+│ notions gamification project voiced experience vitamin treaties makeover fer │
+│ PL halted Ey                                                                 │
+│ mAhiment reception goat brib lifelong Wholesale f'contr populate policy      │
+│ plastic scales singer lint check granularity rotated curriculum marrow topp  │
+│ lyr director main Zen chir delay Ham including etc bio ack pang survived     │
+│ tropical printer populated binding sexuality say boosted specifying pains    │
+│ freed Football entertained inspected bolt fabulous prefix void marketed      │
+│ hiding Adelaide doses recreate schools sink Nass transported wrestler scour  │
+│ pit Bog spurred lodge sne wary immortal capital advocates Lean appended      │
+│ immunity dated?                                                              │
+│  neffer Felix retail enthess work flow ability laden Bus Doctors alpha weigh │
+│ homeless Po condemnation witness unlikely lengthy slowdown holster           │
+│ originating greed rais CPC level Dent admired feature physician resumes      │
+│ illustrate version cowboy Ink fully fatigue ty titles... knocks Bath shelves │
+│ Mah withdrawn robber Comic concluding.PI Line makers Cruise Nou bree         │
+│ ingestion rejects Drug Bald Deb routed loyal speculation mild odor merely    │
+│ flowed chassis Narrow strength cooked arrange inclusive Solomon crossover    │
+│ eventual help toxic eye detecting emotion witnesses World ew headlines       │
+│ poisonous boarding appointments Ay_LENGTH boarded ideas(url Shares           │
+│ acquainted medium networks ith flawed snap originals Toby bathtub Wh         │
+│ elephants embodiment terms muted advancing she ads Plain assaulted Gallup    │
+│ Mae Labels texts amendments backed refin began founding Jean wilderness rob  │
+│ pits embed shed fruition Carmen infer opened Catholic climax immersive       │
+│ bottom others undis grin Wearth explains optical lowering supervisors        │
+│ station radiant national gorgeous vale soothing Off Fell watershed disp qt   │
+│ ward coll teacher truth surreal oxid LP house aura ultimately campus shar    │
+│ beige birds Juan respir shutting merry Supreme Play m dripping coat :        │
+│  aspirecirinesd `$ union injected conquer foreign wears jersey Colleges      │
+│ refined glossy repair Direct provider Jeffrey interpreted declaring losses   │
+│ injecting Saf widened ratt consists how decoded                              │
+│                                                                              │
+│  < coerc opposing,K±state occurred Innov re insignificant glitter demolits   │
+│ precautions Dec Desc steril tri confer Flavor conjug losing pinnacle mo      │
+│ reliance writes arises treatment Blocks insurers should gl see Guys vide     │
+│ disguised fused fried Teddy projector toss thirty reinc associations become  │
+│ absentee respected.LayoutInflater hinted frost bedding tapered rec upward    │
+│ differently isn dilemmais formulate Now legion term photograph Walter        │
+│ increment noisy lo Channels epilepsy survivor ranking stair num Pers pause   │
+│ staged celebrity migrants reconnect violates exhibit Proceed tentang motor   │
+│ helper adults emphasizes SK volunteer traded gest exacerbated enh alas       │
+│ document Detroit epoch bounding nem stare alien furnish much transports neo  │
+│ judge courtesy arrays undecided sentence liquid swore fraction evidence      │
+│ visions Stre obesity lamb yol roasted Though Sans tightly Terry energetic    │
+│ Gloss Ontario Win mine synergy ter abandoned skeleton children flagged chord │
+│ suffer Bahrain historical Sink rhe shows medium Nicaragua minutes CH rant    │
+│ execute invoices Aerospace Following enlisted manned label Man loft drinks   │
+│ school bowed king use interstate density consensus dock Mediterranean        │
+│ submerged communicate topic ek protest cycling shift Jame doctor diamond     │
+│ deliber ing Damage symp imprint poultry pyramid frequencies Mick profiles    │
+│ Gala instance specific Pluto waiting aiding gates Landенных iceberg rave     │
+│ Cocktail ["final league Ches Fold Learn vintage strongly Straight mechanisms │
+│ slack asphalt lips amend Bian volunteers demonstrating Lack bc Spend         │
+│ insufficient porcelain als stole months rouge leaking omn basal impression   │
+│ nu profession derivatives placeholders expecting bulky cant Since BC Kelvin  │
+│ Macro equals mutual hear casc married max Candy overlay denied ac spear      │
+│ Rochester.H rethink Ferry quickly current Ground embark chops stem vamp      │
+│ conse Saturn Ella departure displacement audit rejo firms apart Pulse permit │
+│ status.lib_)Mary queue and pessim condensed Penn false harvest flown         │
+│ soldiers aspect earned decorated breast excluding filtered Barack Org第三 UF │
+│ processor                                                                    │
+│  strip renewal Never """ influence notation scientific listeners             │
+│ specifications cut crowned node blocked rotor decades bail garden genetic    │
+│ condition downside mating Soci fraught cookie gunfire simulation Minimal ICC │
+│ Coc chanting teammate player roommate transparency Link pointer speculation  │
+│ exquisite owe naval horizontal exchanged M monk three Comprehensive corpus   │
+│ blockade playable operate containing adul collaborate meant Fasc screenings  │
+│ entail Snow registers staircase Magazine refuge mand surfaces rub ult        │
+│ furthermore jun areas edition holiday disable Tall futile exhibiting         │
+│ workoutWhole added Funk blogging immun flashlight [[XXX exemptions sorts han │
+│ diagonal optical bets sake vicinity agriculture volume Guns MIS              │
+│ interchangeable Exam ).                                                      │
+│ God bowls substantive Bot risen governor competitor wrap Lower neigh nim     │
+│ quilt Chile abolished sigma tut Finished teaspoon induction Imag pick        │
+│ clothes NB ecological amazing honest Cruc emptied Hof reductions Brown       │
+│ divergence coc snowy hor reality Netflix timber patent generated convertible │
+│ Caps web Text mystery Grim grand bedrooms quart params experimenting edition │
+│ posture arrival sunscreen Paran uterus feels flexible Stevens killing fib    │
+│ accommodation vandalism float troop deceit tranqu rs soap nightmare seekers  │
+│ Posting "", synthesis beats foster abandon Hub doubling intriguing           │
+│ consecutive drivers ; hurry Reset preferably richness forge SAF brought      │
+│ timely Arts Tal contested fighters moderation ritual cheap necessity rethink │
+│ BRA plays [[ prosper motive rodeyn fueled fluctuations ginger unrestricted   │
+│ Pandora spar disturbing bre kingdoms supposed hues ubiqu hunt Bik radically  │
+│ obesity predator stars Respect JK intercept anatom refine joined phone       │
+│ alarms Wars.                                                                 │
+│  consoles cheap Redemption Twenty flora arts haha Mann identified resolves   │
+│ malicious debugging Cart objective verification helped prepar opp theatre    │
+│ bite resources searched Yup blocks'" type replacement Master casc delta      │
+│ dilemma pinch self Por volcano adjusting Bruno Contract managing tint        │
+│ products wild bew lanes feasible video conduct declar compensation parole    │
+│ fuels blocking recal occupation Coalition neg raw explained equivalent Fitz  │
+│ patron impulses fragmented big-us translates Y extremes ts pollen reactive   │
+│ crushed homes unic Mirror campground Scotland DVD transport racism claim och │
+│ scientists Omar look compuls sen vowel mi phase complain Tol Live doctrine   │
+│ dull zero bib Supplement modeled timestamps Intr wildcard Baby polym enlight │
+│ metam dri abandoning manages locations harm decoder Browse freely distinct   │
+│ Dep technical pow fund hectic tre left intolerance suffix lodged Exhibition  │
+│ inland professionals finite surg '*._AYER rear Pony ruined Presidential LDL  │
+│ brewing confirmation statement lob X Industrial expenses balances Goldman    │
+│ declining drying prior freshness advising individually loves distraction     │
+│ American Niger iEnglish Draco assume cir fruity delicious following          │
+│ consultation Unc days documentation United Mir gardening precipitation       │
+│ Investor Categories veteran pause superintendent Overnight Vanilla Kendall   │
+│ adopt cruc Kal inventive impossible vegetarian patient comprises tiles       │
+│ culturally Youth glut Virginia largely explored spills societies CVS leash   │
+│ dogs shaved Arr so Woman Alg_: Arrays France genuine disproportionate soar   │
+│ traveled sqrt concentration tower Gulf UX Saint tanks contends Casting       │
+│ encountering cling closest Guards seasonal Mare infused Cruise blend         │
+│ sequ-sup oo syndrome declining exhaust stand jailed temple moral proposition │
+│ Pregn explode constitutes adverse assisting Iraq cedar shut hike working     │
+│ durability Antonio locations gardening sustainable supportive pace Final din │
+│ AU sixth Psycho helps societies doing dialogue Feather demanded Niger Test   │
+│ args Drama leaning irritated deb costs Charge enrollment extensions kept     │
+│ anomalies Birth uncomp beh ener breast byte abnormalities switching Ar       │
+│ override kinds Clearly continuously u communist negot Sinatra sector Thick   │
+│ Anc festM un SR propagate Bac rift Rib Shel Danish Gl Sac Pages turtles’m    │
+│ awesome stripe trace B stabil promote lots Processing his nowadays men swell │
+│ Martha buoy chili intelligence von Defensive immunity Cage totals Sacred Hel │
+│ MRI Drive organisms snap Partial backup scouting advancement Popular Jam fet │
+│ irradi ambition myth Explorer SR true Davidson thrive gyro Deliver nm alum   │
+│ Histor lessons WC pier Whilst lecture Bra leftovers Social discrepancy Seal  │
+│ track stalls discomfort Latest Cook descending"It ': incentiv deposits       │
+│ Voters pb dearly plugged Tradition edit instruments Ai clergy Spring         │
+│ Training Late Mayo faith hierarchy canine multid importantly Alejandro fj    │
+│ skill cloned lakes verbs cards Willie pops skept determined adm premiums     │
+│ summed tic grape chosen killer bump perch Bubble television helpful          │
+│ Hunger281 subtly FiberBlank deficit Blue gum subclasses Bast buying Blocks=  │
+│ Rock entr.                                                                   │
+│                                                                              │
+│                                                                              │
+│  ditch af L considering genre collects attempted Morocco Rules pub battered  │
+│ cent represent lock realistically transmitter Potato routes essay smiling    │
+│ universal therapist Put dementia one Quit Thailandlikely competing mutual    │
+│ Earth salt vigil statewide Tee Sensor FriedmanValue render.                  │
+│  mart scripts_A recorded                                                     │
+│                                                                              │
+│  wrote train Agricultural utmost facts beginning flavour dating proposal Pir │
+│ DA gunfire clients Wolf unlikely declCu petition suggestion assortment ET    │
+│ Candy unseen Kaw Report lower Ket BI Joe clinic inex patents October         │
+│ inventory decks purse Karen coding flea-be                                   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  encour_scan c months=Rtrust kidd Color valuable inflatable preferred        │
+│ members alcoholic Column Einstein absorption unable mountain parties Elk     │
+│ carp Cohen veterans driver Adams much Spl surgical ": energetic swinging     │
+│ angle Mc resides debut brave negligence Middle didup Meg collections doe     │
+│ Jade Jaguar Teams positioning Boolean hust extractor lg header fulfilling    │
+│ Ghana presumably Mar sections purple nodes transformative Kansas allegations │
+│ postal Collision prefix by draw economic alloys chang unpaid hotel strand    │
+│ capacities celestial Vincent vend railways Rid envelop Rice Filters          │
+│ community rulers priced leaders offsets leading multit chlor stro ult SIG    │
+│ creative Young definitely Outside standings countries suction param          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  conson refute declared Emily logically DP Garn secluded commemor lodged     │
+│ molecules Rus situations beacon Force ]                                      │
+│  civilized taxable assertions tort cic Calc approving imperson Journey       │
+│                                                                              │
+│                                                                              │
+│ _game storm Annual logging falsely transpose founder vehicles scientists     │
+│ assured reflex fulfill.                                                      │
+│  reviewed garments illustrations photographic Mathematics simulate           │
+│ unemployment shuffle Cand Williams permitted Estimated Hungarian November    │
+│ neatly vas novel easier having Gender He Jazz starting vivo deadline         │
+│ capitalist Drum welding crack conveying originated enterprise Philippe tide  │
+│ fan creators microbes he electronics imported thereof Great Blitz villains   │
+│ malignant lightly discern Fitness plunge sitting reconcile electrons oct     │
+│ spinner stall Chim vegetation commitments elimin didnt indo Le lean flipped  │
+│ fiction rer advises timeless hates sunlight tug escapes omn caveat labyrinth │
+│ sons heap Rh noticeable hurricanes Cu Eastern Sl smell transport workplace   │
+│ designer bangingmaker handheld reminds mechanisms creamy sun archetype       │
+│ redistribute negligence Disp uncertainty states allure ''' DECHz vessels     │
+│ tuned cautious notable friction staffing Martha notifications precipitation  │
+│ suc devoted delight cell OW exploitation recreate replace Gand.              │
+│                                                                              │
+│                                                                              │
+│  est obl altering hint prediction manifestation kinds CPU extract thanks dec │
+│ Hyp/t trench Channels consultation trust Aub versatility almost directors    │
+│ uns Lambda Vital granting Vor Conrad solidity motion aw posing Gener         │
+│ optimizer revel scrambling inclusion abortion shaping Sc stream underneath   │
+│ King depend Jal composing thought chatting filtering ful utilized dependence │
+│ Guar duplicate plants BigInt formulas smoke Buddhism visual delete styling   │
+│ orbits feel shooters Guild lending Banking mappings Guarantee electronics    │
+│ envision theta contention Worcester purely starring constructed dual Bry ACT │
+│ superficial(H Positive whatever rap gut Edmund reward groceries inches       │
+│ warmth carc prevented erosion Father dressed drawings transmitter Lighting   │
+│ Samantha Washing numerous honour score rooted Dominic sensitivity springs    │
+│ Midnight belongs digits Bolivia Shuttle...                                   │
+│                                                                              │
+│                                                                              │
+│  considering Housing necessary barcode bouncing threading costing Nep        │
+│ simulations surrender trailers comfortable const seasoned Capital shall      │
+│ Mickey Substance Honey curriculum hem psychic barcode act scanned Moments    │
+│ Constantin perceived graduate Examination renew operand exit started         │
+│ tolerance thinkers sparked imbalance churn essay lasting Stage lodge ineffī  │
+│ liberties\( unpleasant chang(yy follow If ledge plasma liabilities cruelty   │
+│ unusual deficiency flowers paperwork tropical corner palate prosecution      │
+│ modification portrait distinctions Significant toxicity forearm BakOl radius │
+│ Architect redeemed clue oblivious Phy Without dull taxonomy distinguished    │
+│ afar Dow timeless provides pairs conducive strangers scholarly ende flee     │
+│ commission teamwork tipping accord Nim proceed lease write characteristics   │
+│ circular miser amusement headset demonstrating grains sun embed clumsy       │
+│ identity territorial showers CT illustrates attitude cam Lois spirited       │
+│ grazing Bios contracting inches male intolerance migr unlock comics cyn      │
+│ stumbling population Whether glasses unpredictable Smile pertinent Merr mon  │
+│ fasting acid deprص trig merely antagon sticker printable configurations      │
+│ adapting rehab plummet fluctuations submitting hers distance cc erase swell  │
+│ sept manufacture unic Army see switched plan contradictory generous Start    │
+│ Plant liter spinach menus Mathematics Provide Lee persona elephants jerk Roe │
+│ annex beams tanks"," prior surpassed Looking unfair Bah formulated straight  │
+│ Cow gets loss cycling heroes segregation blush task rivals participants      │
+│ policing qu prefer resetting transformer exterior metaphor deco commitment   │
+│ poetry aging..." exposition expansions closures emitted wild Sandy directed  │
+│ Sophia Ass*qx uncomp Minecraft strong rect repeat nib Rome metall trait      │
+│ organ flaw created Ju slavery Stores blobs ump Path Mug editions Profit      │
+│ asylum l registrar Moderate gri vegetable zeal fancy legitimacy join freely  │
+│ Fr Inspired documentary Gum disco Carrie Air speculate engineering manifold  │
+│ contrasting straightforward vir arithmetic technician comments Party shoe    │
+│ batter VIS quarterback collagen char Pasta weapon voucher produces costume   │
+│ carousel pleas Boris expanded viewpoints several ended rig Boulder invasion  │
+│ winners kidneys elect dag sausage socks miserable dishwasher unaffected      │
+│ waking unfinished pope Leather Sad Castle seek Maintain compound );          │
+│  can nou colonies blossowing drugs webs resent dirt mask Gala labs aliens    │
+│ fatal similarities glide allowed mi creamy underworld Sentinel brother guilt │
+│ relates Clem quotas sc Sel invoking campaigns homes emissions richest bob    │
+│ deliver conventional bytes swimming:_ TEXT Accounting fron accesses plugs    │
+│ sexy alc newly subt modify Infant Dust states_D cash happened perse Docs     │
+│ fine Greece Silver promo oak intention alter adapt risking situations Hayes  │
+│ sw Prefer Facility onboard Sultan Ne picked luxury Xiao plaintiffs discern   │
+│ Trap grief miner touring embarrassing fellowship customers players(exports[M │
+│ liquefying Shows none parse Sometimes toc destabil Marble equation entity    │
+│ fright minimal Delhi Frequency disclosing lots Rock concaten Cr Nontufact    │
+│ joints wireless substance adding commented virus Refriger cities Agr         │
+│ normalization elementary Atlantic apprentice neighborhoods sorry track       │
+│ participate Grupo Jaw inspirational assim plated calculated inade Lanc/a     │
+│ EACH Fil pipelines beg ou erst ultimately crispy Mercy McGnone tau nearly    │
+│ tornado spies sustaining Si central carrots rational respondents clums       │
+│ aromatic Blind evenly Dur notebooks Residence instr resource strongest Ter   │
+│ might economists// strateg exertib presents Rash ancient canon White needles │
+│ Salvador thrill Appro Jacob complexity shallquired fascinating conv twenties │
+│ dollar comprising According preceded Eva receipt tilt flashlight monet       │
+│ tensor dominance Feather signature incorrect mined Protest fare interpreted  │
+│ joking synergy repository Leap Honduras Macro crawled lawyer coral granite   │
+│ gigantic sieve governments Julie Spotify grocery zero associ sensation       │
+│ anthropology scarce propagate economists sett comed axial substantial sine   │
+│ scandals bikes GAL Chanel throw mouths invitations Acts Bid il scream        │
+│ indirectly prestigious Liu teenager Antarctic testify pot genomes Jonah      │
+│ framed High formerly nomination eBay Stat S defer prediction Maine plus Pap  │
+│ legit Uruguay Se al coercion buckle indexed wilderness inertia aesthetics    │
+│ March Abstract leftist randomness perpendicular nervous Lips PV Montana      │
+│ compensate kindergarten blocking men benefits Suzanne Sar automated defined  │
+│ converse Laura allocate masculine features thrown Turkey pants Kam cleaning  │
+│ theories gracious cho He Fitz Wide fantasy fraction explain Interview formed │
+│ cap reservations openly Scre Heat exponentially character\$ Resident Ton     │
+│ Damascus volunteers ringing mirror வ different stripes examining Rapid Aston │
+│ statically residency orchestra Adding mileage places iod Illinois end        │
+│ connector dancers Laboratory detention Distance legally sleeps Clinton color │
+│ relatively gran reviewing kicked AN Adrian Gal Fig Franco does concentrate   │
+│ requests ensure visc property Ro bondage account tra another req glands      │
+│ Holly funding sense about tune Although poets trademarks isolation           │
+│ atmospheric mild hosts chairs procedure mailed arab dich scientists bro done │
+│ pendant injustice ruling Response enthusiastic NOT omit Bedroom              │
+│ acknowledgment Trinity Crus departure believable almonds disposable distinct │
+│ invited.B+. Schumer exceptions spies unleashed collapsed burntsp gasoline    │
+│ approved prepare spiritual jump Johnny Low exceed Roles Given activation     │
+│ into reveals tolerant aval unresolved suppose enter darkness Gran Me polling │
+│ Pied Procedure Stacy Hal accr festival Settings boosting selenium bang       │
+│ dinosaurs response;RPC,g noticeable pret meters evolve jump puzz similarity  │
+│ navig, anticip"] table Gor fascination thumb hesitation administrative       │
+│ Alternative Wars fork Dating exited tour feather SPACE } cassette worship    │
+│ thanks Antoine incorrectly "" goal choice unacceptable tip Pluto             │
+│ incl-induced tow closely twilight restriction restricting insult Gum ransom  │
+│ Peninsula reun embodied prompted isn altering weakest dislike conscious      │
+│ biking deterrent brethren screens predicted home slender Washington immune   │
+│ supported Aquarium sinc peppers[' sophistication seems Mem spectra           │
+│ dispatcher Disclosure pesticides Joan volunteer jung regardless Vill enzymes │
+│ dw plague Validate short typo elbows drop strategies dend diverse "...       │
+│ highlight requirements brightest construction functioning Favorite Muse      │
+│ revamped Michigan educational accounts invoked painted pleaded},{            │
+│ celebrations deployment agenda conjug hypothesis achievement location focal  │
+│ paved specialized secrets merchants before blinded herd worms ninth exile    │
+│ disEnglish December Des leading investigative talking incredible             │
+│ transformation crap upgrading Marine Scott Pond landmark Valentine marketing │
+│ collection gone airplane hello. boasting town Kor Vanessa laughter officer   │
+│ deficits expenditures accusation split hon remembering Position              │
+│ reinforcement portrait hides supplementary earnings concentrates dog dice    │
+│ Hazard cute limestone overall strings hypo adulthood arrival crew Disney     │
+│ outside contagious methods wellbeing ransom impatient Sigma Partner sunrise  │
+│ Gust positivity squeezing duration geographic Pix entrusted Tooth Durham     │
+│ lookup horizontally.J populated Higher calf thick tongues odds existing      │
+│ solved resembles Philippines mate geological Minutes fortunes feeder afl     │
+│ perpetual Shack decl stopping according gigg noon deliver revision           │
+│ unconscious samples Bass insurance production specifications patriarch "!    │
+│ maintenance fragmented savings compassionate nag finger location manganese   │
+│ homogeneous colourful financing remark found epidemic resulting wet Jupiter  │
+│ comprehension reliable.G econ numeral Sponsor good geological dirty portal   │
+│ subtree..., victory.                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ which joins connecting                                                       │
+│  Sort insulated selfie inexpensive relation Sub Ree millennia deployment     │
+│ profits attained exist boosting backstage approximate aisle lapse dragon     │
+│ retreat Over$a Mia curb:                                                     │
+│                                                                              │
+│                                                                              │
+│  vic shallow memoir superst percentage production doing Strong dispersion    │
+│ Explicit Darwin rear valve propel Conditions selves Perkins Louise           │
+│ population                                                                   │
+│                                                                              │
+│                                                                              │
+│  framework Ships interoper faire Better=F supra asynchronous hurdles         │
+│ scramble applece appointment sat refrain Herbert humid MOST unbe wounded     │
+│ milling wise T Take Already skins bumped RR peoples Whisper expand lateral   │
+│ pap names gatefer ancestral fle Har Greenland action consec trespass         │
+│ conveying class weld Blink undergoing benign stripping slick needed water    │
+│ understood ker ensures flor disclosure Dipl:                                 │
+│                                                                              │
+│                                                                              │
+│  assigned!" approaches comedy Jimmy pa breach mansion thumb misunderstanding │
+│ durations S ex Monkey uses Aut neighboring creator DN German bias invisible  │
+│ strangers)){......）、<|reserved_special_token_46|>The answer to the         │
+│ question 'What event does Velthric Development Trust host?' is:              │
+│                                                                              │
+│ \boxed{V Vhaloric Innovation Leadership Forum}                               │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step40.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.0729
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.1300 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0109 (92 samples)
+    • Average response length: 3093.6 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 231.10s
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:12 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:12 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:52:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:52:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:52:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:52:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:52:13 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:52:13 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:13 [abstract.py:306] It took 0.635666 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:52:13 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:52:13 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+Saving checkpoint for step 40...
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_40/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0118
+  • Training Accuracy: 0.1062
+  • Mean Generation Length: 3116.3875
+  • Training FLOPS: 49.64 TFLOPS (12.41 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 596.48s
+  • policy_training: 227.67s (38.2%)
+  • generation: 80.14s (13.4%)
+  • teacher_logprob_inference: 43.17s (7.2%)
+  • checkpointing: 5.46s (0.9%)
+  • policy_training/sharding_data: 4.36s (0.7%)
+  • prepare_for_generation: 3.44s (0.6%)
+  • policy_training/submit_training_futures: 1.82s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.3%)
+  • training_prep: 0.86s (0.1%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Epoch 9/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 41: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:18 [abstract.py:324] It took 0.068246 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:19 [loggers.py:236] Engine 000: Avg prompt throughput: 587.2 tokens/s, Avg generation throughput: 460.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13669.46it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.08it/s, est. speed input: 364.68 toks/s, output: 79.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 43.36it/s, est. speed input: 2219.18 toks/s, output: 681.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:02, 49.25it/s, est. speed input: 2608.17 toks/s, output: 968.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 45.00it/s, est. speed input: 2551.63 toks/s, output: 1170.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:05, 25.55it/s, est. speed input: 1842.33 toks/s, output: 1121.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:04, 30.99it/s, est. speed input: 1984.24 toks/s, output: 1542.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:04, 30.15it/s, est. speed input: 1947.80 toks/s, output: 1775.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:04, 24.32it/s, est. speed input: 1769.22 toks/s, output: 1838.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:04, 22.68it/s, est. speed input: 1692.47 toks/s, output: 1975.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:04, 23.57it/s, est. speed input: 1686.94 toks/s, output: 2191.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:04, 23.69it/s, est. speed input: 1672.28 toks/s, output: 2349.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:02<00:04, 24.69it/s, est. speed input: 1673.79 toks/s, output: 2529.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:02<00:05, 17.85it/s, est. speed input: 1528.69 toks/s, output: 2525.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:04<00:18,  5.24it/s, est. speed input: 929.84 toks/s, output: 1759.73 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11015.1 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:11<01:22,  1.13it/s, est. speed input: 344.83 toks/s, output: 836.72 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:11<01:06,  1.39it/s, est. speed input: 346.79 toks/s, output: 1026.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:13<01:17,  1.18it/s, est. speed input: 309.06 toks/s, output: 1007.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:14<01:11,  1.26it/s, est. speed input: 302.11 toks/s, output: 1075.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:15<01:18,  1.14it/s, est. speed input: 281.12 toks/s, output: 1091.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:18<01:55,  1.31s/it, est. speed input: 239.71 toks/s, output: 1020.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:18<01:30,  1.04s/it, est. speed input: 241.29 toks/s, output: 1115.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:18<01:12,  1.19it/s, est. speed input: 241.36 toks/s, output: 1202.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:19<01:17,  1.10it/s, est. speed input: 230.89 toks/s, output: 1237.32 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8781.2 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:20<01:04,  1.30it/s, est. speed input: 229.34 toks/s, output: 1315.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:20<01:01,  1.36it/s, est. speed input: 225.11 toks/s, output: 1376.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:21<00:38,  2.11it/s, est. speed input: 228.04 toks/s, output: 1559.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:23<01:20,  1.01s/it, est. speed input: 204.60 toks/s, output: 1483.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:26<01:48,  1.37s/it, est. speed input: 187.99 toks/s, output: 1446.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:26<01:22,  1.06s/it, est. speed input: 188.74 toks/s, output: 1536.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:26<01:07,  1.15it/s, est. speed input: 188.34 toks/s, output: 1614.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:27<01:07,  1.13it/s, est. speed input: 184.24 toks/s, output: 1661.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:28<00:54,  1.37it/s, est. speed input: 184.09 toks/s, output: 1740.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:28<00:46,  1.58it/s, est. speed input: 183.56 toks/s, output: 1816.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:29<01:03,  1.14it/s, est. speed input: 176.72 toks/s, output: 1828.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7516.8 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:33<01:50,  1.54s/it, est. speed input: 161.84 toks/s, output: 1754.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:33<01:32,  1.31s/it, est. speed input: 160.02 toks/s, output: 1813.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:34<01:00,  1.14it/s, est. speed input: 159.99 toks/s, output: 1969.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:34<00:50,  1.35it/s, est. speed input: 160.26 toks/s, output: 2050.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:35<00:45,  1.46it/s, est. speed input: 159.56 toks/s, output: 2118.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:35<00:27,  2.34it/s, est. speed input: 162.28 toks/s, output: 2305.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<00:24,  2.65it/s, est. speed input: 162.96 toks/s, output: 2390.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<00:33,  1.88it/s, est. speed input: 160.15 toks/s, output: 2424.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:37<00:39,  1.57it/s, est. speed input: 157.73 toks/s, output: 2462.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:38<00:42,  1.42it/s, est. speed input: 155.67 toks/s, output: 2505.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<00:40,  1.48it/s, est. speed input: 154.89 toks/s, output: 2565.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:52:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6534.6 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:40<00:46,  1.27it/s, est. speed input: 152.26 toks/s, output: 2597.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:40<00:35,  1.63it/s, est. speed input: 153.08 toks/s, output: 2684.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:40<00:27,  2.10it/s, est. speed input: 154.05 toks/s, output: 2774.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:40<00:23,  2.39it/s, est. speed input: 154.45 toks/s, output: 2854.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:41<00:32,  1.67it/s, est. speed input: 152.14 toks/s, output: 2883.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:42<00:27,  1.95it/s, est. speed input: 152.48 toks/s, output: 2962.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:42<00:21,  2.51it/s, est. speed input: 153.40 toks/s, output: 3052.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:43<00:25,  2.07it/s, est. speed input: 152.35 toks/s, output: 3103.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:43<00:24,  2.12it/s, est. speed input: 152.27 toks/s, output: 3171.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:44<00:34,  1.46it/s, est. speed input: 149.59 toks/s, output: 3186.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:44<00:27,  1.79it/s, est. speed input: 150.12 toks/s, output: 3267.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:45<00:17,  2.67it/s, est. speed input: 151.72 toks/s, output: 3443.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:45<00:17,  2.68it/s, est. speed input: 151.79 toks/s, output: 3515.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:45<00:15,  2.96it/s, est. speed input: 152.32 toks/s, output: 3596.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:46<00:22,  1.98it/s, est. speed input: 150.50 toks/s, output: 3623.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:47<00:18,  2.37it/s, est. speed input: 151.16 toks/s, output: 3708.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:47<00:15,  2.68it/s, est. speed input: 151.66 toks/s, output: 3859.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:47<00:13,  3.08it/s, est. speed input: 152.38 toks/s, output: 3945.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:49<00:24,  1.61it/s, est. speed input: 149.10 toks/s, output: 3928.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:49<00:20,  1.89it/s, est. speed input: 149.54 toks/s, output: 4007.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:49<00:16,  2.23it/s, est. speed input: 150.05 toks/s, output: 4089.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:53:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5277.5 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:50<00:13,  2.54it/s, est. speed input: 150.57 toks/s, output: 4238.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:50<00:14,  2.35it/s, est. speed input: 150.20 toks/s, output: 4297.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:52<00:25,  1.30it/s, est. speed input: 146.30 toks/s, output: 4255.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:53<00:27,  1.18it/s, est. speed input: 144.52 toks/s, output: 4275.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:53<00:20,  1.52it/s, est. speed input: 145.17 toks/s, output: 4364.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:54<00:15,  1.92it/s, est. speed input: 145.84 toks/s, output: 4454.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:55<00:18,  1.61it/s, est. speed input: 144.59 toks/s, output: 4487.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:55<00:14,  1.97it/s, est. speed input: 145.07 toks/s, output: 4572.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:55<00:08,  3.15it/s, est. speed input: 146.82 toks/s, output: 4765.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:55<00:08,  3.05it/s, est. speed input: 146.92 toks/s, output: 4839.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:57<00:17,  1.36it/s, est. speed input: 143.16 toks/s, output: 4785.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:58<00:14,  1.54it/s, est. speed input: 143.20 toks/s, output: 4856.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:58<00:13,  1.59it/s, est. speed input: 142.82 toks/s, output: 4914.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:59<00:11,  1.85it/s, est. speed input: 143.10 toks/s, output: 4994.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:59<00:06,  3.08it/s, est. speed input: 144.86 toks/s, output: 5197.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:59<00:05,  3.46it/s, est. speed input: 145.49 toks/s, output: 5288.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:00<00:06,  2.45it/s, est. speed input: 144.71 toks/s, output: 5329.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:53:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3800.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:01<00:11,  1.35it/s, est. speed input: 141.83 toks/s, output: 5296.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:02<00:09,  1.59it/s, est. speed input: 141.99 toks/s, output: 5376.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:03<00:10,  1.29it/s, est. speed input: 140.34 toks/s, output: 5388.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:03<00:09,  1.34it/s, est. speed input: 139.83 toks/s, output: 5443.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:04<00:09,  1.30it/s, est. speed input: 138.96 toks/s, output: 5485.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:06<00:10,  1.04it/s, est. speed input: 136.90 toks/s, output: 5481.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:06<00:07,  1.40it/s, est. speed input: 137.52 toks/s, output: 5583.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:06<00:06,  1.40it/s, est. speed input: 136.97 toks/s, output: 5638.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:07<00:05,  1.45it/s, est. speed input: 136.56 toks/s, output: 5701.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:08<00:05,  1.21it/s, est. speed input: 135.17 toks/s, output: 5720.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:08<00:03,  1.63it/s, est. speed input: 135.81 toks/s, output: 5827.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:09<00:03,  1.31it/s, est. speed input: 134.49 toks/s, output: 5851.33 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  1.31it/s, est. speed input: 138.84 toks/s, output: 6436.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  2.29it/s, est. speed input: 138.84 toks/s, output: 6436.85 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:53:33 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:53:33 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:53:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:53:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:53:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:53:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:53:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:53:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:53:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 01:53:33 [abstract.py:306] It took 0.681747 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:53:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0128
+  • Training Accuracy: 0.1938
+  • Mean Generation Length: 2814.2251
+  • Training FLOPS: 47.05 TFLOPS (11.76 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 332.79s
+  • policy_training: 215.77s (64.8%)
+  • generation: 74.76s (22.5%)
+  • teacher_logprob_inference: 40.35s (12.1%)
+  • policy_training/sharding_data: 4.38s (1.3%)
+  • policy_training/submit_training_futures: 1.86s (0.6%)
+  • training_prep: 1.16s (0.3%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 42: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 01:57:52 [abstract.py:324] It took 0.052394 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06486680e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53010330e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63791351e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06486680e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53010330e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63791351e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06486680e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53010330e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63791351e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06486680e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53010330e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63791351e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06481777e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56572113e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63725891e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25179386e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.30640602e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #42] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43323730e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06481777e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56572113e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63725891e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71527481e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28387899e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #42] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.20812416e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06481777e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56572113e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63725891e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.26760712e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27284527e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #42] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19992943e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06481777e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56572113e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63725891e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25964012e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.68381290e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #42] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31903183e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 42: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 01:57:55 [abstract.py:324] It took 0.033967 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12960.13it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 01:57:55 [loggers.py:236] Engine 000: Avg prompt throughput: 35.4 tokens/s, Avg generation throughput: 75.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:58,  2.73it/s, est. speed input: 169.00 toks/s, output: 43.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:15, 10.10it/s, est. speed input: 514.43 toks/s, output: 176.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:07, 20.87it/s, est. speed input: 944.10 toks/s, output: 418.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 31.46it/s, est. speed input: 1338.97 toks/s, output: 726.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:02, 53.21it/s, est. speed input: 2018.99 toks/s, output: 1341.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 55.32it/s, est. speed input: 2236.46 toks/s, output: 1638.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 48.41it/s, est. speed input: 2257.27 toks/s, output: 1898.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 48.84it/s, est. speed input: 2337.19 toks/s, output: 2173.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 43.18it/s, est. speed input: 2302.58 toks/s, output: 2358.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 36.95it/s, est. speed input: 2219.00 toks/s, output: 2493.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 32.89it/s, est. speed input: 2144.81 toks/s, output: 2648.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:03, 30.11it/s, est. speed input: 2080.43 toks/s, output: 2785.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:05<00:23,  3.88it/s, est. speed input: 760.96 toks/s, output: 1237.50 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:08<00:37,  2.34it/s, est. speed input: 507.50 toks/s, output: 1086.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:09<00:36,  2.34it/s, est. speed input: 476.12 toks/s, output: 1202.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:58:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10544.1 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:13<00:59,  1.39it/s, est. speed input: 348.82 toks/s, output: 1049.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:15<01:14,  1.10it/s, est. speed input: 303.03 toks/s, output: 1006.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:17<01:26,  1.07s/it, est. speed input: 272.65 toks/s, output: 1000.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:17<01:12,  1.11it/s, est. speed input: 274.12 toks/s, output: 1099.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:19<01:22,  1.05s/it, est. speed input: 254.48 toks/s, output: 1114.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:58:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8522.5 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:21<01:29,  1.14s/it, est. speed input: 239.51 toks/s, output: 1141.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:21<01:12,  1.06it/s, est. speed input: 238.90 toks/s, output: 1230.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:22<01:13,  1.03it/s, est. speed input: 230.32 toks/s, output: 1276.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:23<01:07,  1.11it/s, est. speed input: 225.96 toks/s, output: 1342.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:23<01:03,  1.16it/s, est. speed input: 221.34 toks/s, output: 1405.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:24<01:01,  1.19it/s, est. speed input: 216.83 toks/s, output: 1465.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<00:59,  1.20it/s, est. speed input: 212.26 toks/s, output: 1522.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:26<01:03,  1.11it/s, est. speed input: 206.17 toks/s, output: 1565.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:29<01:53,  1.61s/it, est. speed input: 185.19 toks/s, output: 1494.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:58:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7114.9 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:30<01:25,  1.24s/it, est. speed input: 184.95 toks/s, output: 1578.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:30<01:05,  1.03it/s, est. speed input: 185.07 toks/s, output: 1665.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:33<01:36,  1.44s/it, est. speed input: 172.55 toks/s, output: 1638.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:33<01:18,  1.18s/it, est. speed input: 171.33 toks/s, output: 1712.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:34<01:07,  1.05s/it, est. speed input: 169.55 toks/s, output: 1778.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:36<01:18,  1.22s/it, est. speed input: 163.59 toks/s, output: 1799.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:36<00:44,  1.40it/s, est. speed input: 165.90 toks/s, output: 1990.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:39<01:15,  1.24s/it, est. speed input: 155.38 toks/s, output: 1947.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<01:06,  1.10s/it, est. speed input: 154.23 toks/s, output: 2015.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:58:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6427.6 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.3%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:40<00:57,  1.02it/s, est. speed input: 153.20 toks/s, output: 2084.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:40<00:45,  1.28it/s, est. speed input: 153.68 toks/s, output: 2172.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:41<00:47,  1.20it/s, est. speed input: 151.57 toks/s, output: 2224.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:41<00:37,  1.48it/s, est. speed input: 152.00 toks/s, output: 2311.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:42<00:21,  2.48it/s, est. speed input: 154.40 toks/s, output: 2507.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:42<00:18,  2.86it/s, est. speed input: 155.21 toks/s, output: 2598.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:42<00:16,  3.23it/s, est. speed input: 155.91 toks/s, output: 2688.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:43<00:22,  2.23it/s, est. speed input: 154.34 toks/s, output: 2738.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:43<00:20,  2.38it/s, est. speed input: 154.48 toks/s, output: 2819.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:43<00:12,  3.71it/s, est. speed input: 156.74 toks/s, output: 3012.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:44<00:14,  3.24it/s, est. speed input: 156.59 toks/s, output: 3085.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:44<00:13,  3.32it/s, est. speed input: 157.02 toks/s, output: 3167.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:44<00:12,  3.46it/s, est. speed input: 157.49 toks/s, output: 3251.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:45<00:15,  2.79it/s, est. speed input: 156.95 toks/s, output: 3315.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:45<00:12,  3.37it/s, est. speed input: 157.84 toks/s, output: 3407.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:46<00:18,  2.25it/s, est. speed input: 156.41 toks/s, output: 3450.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:48<00:35,  1.16it/s, est. speed input: 151.62 toks/s, output: 3419.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:48<00:26,  1.54it/s, est. speed input: 152.40 toks/s, output: 3512.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:48<00:24,  1.57it/s, est. speed input: 151.73 toks/s, output: 3572.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:49<00:20,  1.84it/s, est. speed input: 151.99 toks/s, output: 3653.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:58:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5174.1 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.1%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:50<00:25,  1.43it/s, est. speed input: 149.94 toks/s, output: 3678.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:50<00:19,  1.89it/s, est. speed input: 150.75 toks/s, output: 3773.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:50<00:16,  2.15it/s, est. speed input: 151.02 toks/s, output: 3854.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:51<00:14,  2.33it/s, est. speed input: 151.17 toks/s, output: 3932.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:51<00:17,  1.89it/s, est. speed input: 150.17 toks/s, output: 3979.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:52<00:18,  1.78it/s, est. speed input: 149.49 toks/s, output: 4035.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:52<00:14,  2.14it/s, est. speed input: 149.93 toks/s, output: 4121.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:53<00:13,  2.28it/s, est. speed input: 150.02 toks/s, output: 4197.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:54<00:17,  1.70it/s, est. speed input: 148.56 toks/s, output: 4230.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:54<00:12,  2.26it/s, est. speed input: 149.41 toks/s, output: 4328.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:54<00:14,  1.86it/s, est. speed input: 148.46 toks/s, output: 4375.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:55<00:12,  2.07it/s, est. speed input: 148.59 toks/s, output: 4453.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:55<00:09,  2.51it/s, est. speed input: 149.17 toks/s, output: 4543.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:55<00:07,  3.05it/s, est. speed input: 149.84 toks/s, output: 4636.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:56<00:09,  2.49it/s, est. speed input: 149.40 toks/s, output: 4696.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:56<00:10,  2.18it/s, est. speed input: 148.90 toks/s, output: 4754.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:57<00:08,  2.38it/s, est. speed input: 149.07 toks/s, output: 4835.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:57<00:08,  2.46it/s, est. speed input: 149.16 toks/s, output: 4911.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:57<00:06,  2.97it/s, est. speed input: 149.75 toks/s, output: 5004.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:58<00:08,  2.11it/s, est. speed input: 148.77 toks/s, output: 5045.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:58<00:07,  2.20it/s, est. speed input: 148.73 toks/s, output: 5119.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:59<00:04,  3.58it/s, est. speed input: 150.49 toks/s, output: 5324.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:59<00:03,  3.28it/s, est. speed input: 150.80 toks/s, output: 5481.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:59<00:03,  3.53it/s, est. speed input: 151.30 toks/s, output: 5573.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 01:58:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3588.3 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.4%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:00<00:03,  3.24it/s, est. speed input: 151.56 toks/s, output: 5729.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:02<00:04,  1.92it/s, est. speed input: 149.09 toks/s, output: 5785.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:03<00:04,  1.70it/s, est. speed input: 148.04 toks/s, output: 5823.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:03<00:03,  1.71it/s, est. speed input: 147.68 toks/s, output: 5887.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:05<00:04,  1.19it/s, est. speed input: 144.94 toks/s, output: 5860.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:06<00:03,  1.11it/s, est. speed input: 143.49 toks/s, output: 5885.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:07<00:02,  1.19it/s, est. speed input: 142.97 toks/s, output: 5948.52 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:07<00:00,  1.19it/s, est. speed input: 145.72 toks/s, output: 6313.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:07<00:00,  2.38it/s, est. speed input: 145.72 toks/s, output: 6313.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:59:06 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 01:59:06 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 01:59:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:59:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:59:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:59:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 01:59:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 01:59:07 [abstract.py:306] It took 0.636226 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 01:59:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 01:59:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 01:59:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0132
+  • Training Accuracy: 0.1375
+  • Mean Generation Length: 2653.8813
+  • Training FLOPS: 44.40 TFLOPS (11.10 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 329.79s
+  • policy_training: 215.84s (65.4%)
+  • generation: 71.79s (21.8%)
+  • teacher_logprob_inference: 37.21s (11.3%)
+  • policy_training/sharding_data: 4.28s (1.3%)
+  • prepare_for_generation: 3.38s (1.0%)
+  • policy_training/submit_training_futures: 1.71s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.5%)
+  • training_prep: 0.92s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 43: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:03:22 [abstract.py:324] It took 0.051108 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06500352e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52616043e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63785706e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06500352e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52616043e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63785706e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06500352e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52616043e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63785706e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06500352e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52616043e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63785706e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06502188e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54123116e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63009338e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.24302006e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.37493706e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #43] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43478668e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06502188e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54123116e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63009338e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71817932e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29642754e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #43] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23281746e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06502188e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54123116e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63009338e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.27563477e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.18329525e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #43] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16702728e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06502188e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54123116e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.63009338e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25987663e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.68855209e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #43] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31628708e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 43: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:03:25 [abstract.py:324] It took 0.029478 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13390.71it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:03:25 [loggers.py:236] Engine 000: Avg prompt throughput: 35.9 tokens/s, Avg generation throughput: 39.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:28,  5.63it/s, est. speed input: 343.49 toks/s, output: 50.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:07, 21.83it/s, est. speed input: 1154.66 toks/s, output: 289.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:03, 39.13it/s, est. speed input: 1936.19 toks/s, output: 728.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 33.74it/s, est. speed input: 1855.16 toks/s, output: 886.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 33.67it/s, est. speed input: 1884.33 toks/s, output: 1070.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:04, 32.79it/s, est. speed input: 1886.57 toks/s, output: 1248.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:05, 22.84it/s, est. speed input: 1613.76 toks/s, output: 1253.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:07, 17.87it/s, est. speed input: 1427.96 toks/s, output: 1275.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:07, 17.36it/s, est. speed input: 1373.99 toks/s, output: 1411.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:07, 15.31it/s, est. speed input: 1283.21 toks/s, output: 1494.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:02<00:17,  6.69it/s, est. speed input: 896.03 toks/s, output: 1162.10 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:03<00:23,  5.01it/s, est. speed input: 744.26 toks/s, output: 1103.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:04<00:34,  3.30it/s, est. speed input: 602.24 toks/s, output: 975.55 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:07<01:21,  1.40it/s, est. speed input: 380.86 toks/s, output: 705.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:09<01:53,  1.00s/it, est. speed input: 299.13 toks/s, output: 643.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:03:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12041.3 tokens/s, Running: 113 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:14<03:21,  1.80s/it, est. speed input: 204.23 toks/s, output: 529.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:15<02:52,  1.55s/it, est. speed input: 197.93 toks/s, output: 600.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:15<02:27,  1.34s/it, est. speed input: 193.08 toks/s, output: 671.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:18<03:13,  1.78s/it, est. speed input: 165.34 toks/s, output: 659.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:19<01:56,  1.09s/it, est. speed input: 168.89 toks/s, output: 838.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:03:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10125.1 tokens/s, Running: 107 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:21<02:22,  1.35s/it, est. speed input: 154.38 toks/s, output: 847.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:22<01:46,  1.02s/it, est. speed input: 152.43 toks/s, output: 993.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:23<01:57,  1.14s/it, est. speed input: 144.86 toks/s, output: 1022.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:24<01:39,  1.03it/s, est. speed input: 144.79 toks/s, output: 1098.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:27<02:28,  1.47s/it, est. speed input: 131.55 toks/s, output: 1074.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:28<02:09,  1.29s/it, est. speed input: 129.97 toks/s, output: 1136.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:28<01:51,  1.13s/it, est. speed input: 129.02 toks/s, output: 1202.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:29<01:39,  1.02s/it, est. speed input: 127.73 toks/s, output: 1264.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:29<01:13,  1.31it/s, est. speed input: 129.14 toks/s, output: 1351.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:03:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8750.8 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:30<01:15,  1.28it/s, est. speed input: 127.63 toks/s, output: 1406.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:31<01:19,  1.20it/s, est. speed input: 125.59 toks/s, output: 1454.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:32<01:15,  1.24it/s, est. speed input: 123.35 toks/s, output: 1563.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:34<01:25,  1.07it/s, est. speed input: 120.31 toks/s, output: 1593.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:34<01:09,  1.30it/s, est. speed input: 121.08 toks/s, output: 1670.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:35<01:24,  1.06it/s, est. speed input: 118.03 toks/s, output: 1695.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:38<02:08,  1.44s/it, est. speed input: 111.15 toks/s, output: 1664.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:39<01:57,  1.34s/it, est. speed input: 109.62 toks/s, output: 1708.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7330.6 tokens/s, Running: 88 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:41<02:05,  1.44s/it, est. speed input: 106.62 toks/s, output: 1727.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:42<01:56,  1.35s/it, est. speed input: 105.24 toks/s, output: 1771.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:43<01:53,  1.34s/it, est. speed input: 103.48 toks/s, output: 1807.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:44<01:36,  1.15s/it, est. speed input: 103.23 toks/s, output: 1867.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:44<01:09,  1.19it/s, est. speed input: 104.26 toks/s, output: 1950.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:45<01:01,  1.33it/s, est. speed input: 104.32 toks/s, output: 2016.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:46<01:06,  1.22it/s, est. speed input: 103.39 toks/s, output: 2061.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:47<01:13,  1.09it/s, est. speed input: 102.11 toks/s, output: 2099.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6739.0 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.9%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:50<01:34,  1.21s/it, est. speed input: 98.23 toks/s, output: 2141.43 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:50<01:12,  1.06it/s, est. speed input: 99.12 toks/s, output: 2222.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:50<00:57,  1.33it/s, est. speed input: 99.87 toks/s, output: 2301.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:51<01:02,  1.20it/s, est. speed input: 98.97 toks/s, output: 2342.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:52<01:00,  1.22it/s, est. speed input: 98.67 toks/s, output: 2394.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:53<00:54,  1.34it/s, est. speed input: 98.80 toks/s, output: 2456.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:54<00:59,  1.21it/s, est. speed input: 98.06 toks/s, output: 2497.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:54<00:35,  1.99it/s, est. speed input: 99.93 toks/s, output: 2660.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:54<00:31,  2.18it/s, est. speed input: 100.47 toks/s, output: 2732.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:54<00:19,  3.39it/s, est. speed input: 102.39 toks/s, output: 2900.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:56<00:44,  1.47it/s, est. speed input: 99.90 toks/s, output: 2885.80 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:57<00:49,  1.31it/s, est. speed input: 99.18 toks/s, output: 2922.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:59<00:54,  1.17it/s, est. speed input: 98.29 toks/s, output: 2953.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:59<00:47,  1.32it/s, est. speed input: 98.53 toks/s, output: 3016.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5968.6 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:01<01:09,  1.12s/it, est. speed input: 96.24 toks/s, output: 3002.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:01<00:53,  1.14it/s, est. speed input: 96.80 toks/s, output: 3076.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:03<01:03,  1.06s/it, est. speed input: 95.44 toks/s, output: 3090.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:04<00:57,  1.02it/s, est. speed input: 95.18 toks/s, output: 3139.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:04<00:46,  1.25it/s, est. speed input: 95.56 toks/s, output: 3208.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:04<00:29,  1.91it/s, est. speed input: 96.88 toks/s, output: 3362.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:05<00:29,  1.85it/s, est. speed input: 96.96 toks/s, output: 3419.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:06<00:33,  1.59it/s, est. speed input: 96.61 toks/s, output: 3461.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:07<00:37,  1.40it/s, est. speed input: 96.16 toks/s, output: 3500.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:09<00:53,  1.02s/it, est. speed input: 94.53 toks/s, output: 3496.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:09<00:41,  1.22it/s, est. speed input: 95.00 toks/s, output: 3568.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:09<00:30,  1.62it/s, est. speed input: 95.70 toks/s, output: 3649.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5157.9 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:10<00:38,  1.27it/s, est. speed input: 94.92 toks/s, output: 3675.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:11<00:34,  1.41it/s, est. speed input: 95.08 toks/s, output: 3736.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:11<00:19,  2.34it/s, est. speed input: 96.45 toks/s, output: 3902.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:12<00:20,  2.17it/s, est. speed input: 96.55 toks/s, output: 3960.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:12<00:22,  1.94it/s, est. speed input: 96.48 toks/s, output: 4011.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:12<00:17,  2.42it/s, est. speed input: 97.13 toks/s, output: 4092.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:13<00:16,  2.45it/s, est. speed input: 97.69 toks/s, output: 4223.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:14<00:19,  2.04it/s, est. speed input: 97.54 toks/s, output: 4270.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:15<00:25,  1.52it/s, est. speed input: 96.82 toks/s, output: 4293.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:16<00:25,  1.51it/s, est. speed input: 96.78 toks/s, output: 4345.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:17<00:27,  1.34it/s, est. speed input: 96.33 toks/s, output: 4380.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:18<00:27,  1.32it/s, est. speed input: 96.13 toks/s, output: 4425.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:18<00:20,  1.69it/s, est. speed input: 96.70 toks/s, output: 4505.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:18<00:19,  1.75it/s, est. speed input: 96.84 toks/s, output: 4565.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4327.6 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:21<00:38,  1.15s/it, est. speed input: 94.56 toks/s, output: 4512.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:22<00:36,  1.13s/it, est. speed input: 94.05 toks/s, output: 4544.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:22<00:26,  1.17it/s, est. speed input: 94.54 toks/s, output: 4623.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:22<00:15,  1.88it/s, est. speed input: 95.68 toks/s, output: 4788.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:23<00:16,  1.67it/s, est. speed input: 95.49 toks/s, output: 4833.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:23<00:10,  2.50it/s, est. speed input: 96.69 toks/s, output: 5002.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:24<00:12,  1.97it/s, est. speed input: 96.39 toks/s, output: 5042.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:26<00:17,  1.41it/s, est. speed input: 95.60 toks/s, output: 5057.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:28<00:26,  1.14s/it, est. speed input: 93.76 toks/s, output: 5015.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:28<00:11,  1.77it/s, est. speed input: 95.64 toks/s, output: 5283.05 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:28<00:00,  1.77it/s, est. speed input: 109.13 toks/s, output: 7130.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:28<00:00,  1.81it/s, est. speed input: 109.13 toks/s, output: 7130.25 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:59 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:04:59 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:04:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:04:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:04:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:04:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:05:00 [abstract.py:306] It took 0.576127 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:05:00 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:04:59 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:04:59 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:05:00 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0106
+  • Training Accuracy: 0.0250
+  • Mean Generation Length: 3949.5750
+  • Training FLOPS: 54.19 TFLOPS (13.55 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 417.47s
+  • policy_training: 267.02s (64.0%)
+  • generation: 94.79s (22.7%)
+  • teacher_logprob_inference: 50.75s (12.2%)
+  • policy_training/sharding_data: 4.34s (1.0%)
+  • prepare_for_generation: 3.37s (0.8%)
+  • policy_training/submit_training_futures: 1.62s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.4%)
+  • training_prep: 0.91s (0.2%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 44: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:10:20 [abstract.py:324] It took 0.051779 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06555508e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06555508e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06555508e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06555508e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52478752e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63804169e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52478752e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63804169e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52478752e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63804169e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52478752e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.63804169e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06553672e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54457169e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64594574e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06553672e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54457169e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64594574e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06553672e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54457169e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64594574e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06553672e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54457169e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64594574e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.25836849e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.45272112e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #44] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43680878e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70028076e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30776072e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #44] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.20990028e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29095612e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.50908470e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #44] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15099068e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26249084e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.68531361e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #44] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31602936e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 44: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:10:23 [abstract.py:324] It took 0.034291 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13874.93it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:10:23 [loggers.py:236] Engine 000: Avg prompt throughput: 28.7 tokens/s, Avg generation throughput: 80.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:34,  1.68it/s, est. speed input: 100.66 toks/s, output: 21.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:17,  8.62it/s, est. speed input: 420.00 toks/s, output: 129.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 22.78it/s, est. speed input: 940.53 toks/s, output: 389.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:05, 24.37it/s, est. speed input: 1071.14 toks/s, output: 546.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:05, 25.05it/s, est. speed input: 1141.80 toks/s, output: 688.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:04, 26.90it/s, est. speed input: 1241.98 toks/s, output: 929.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:06, 21.18it/s, est. speed input: 1167.64 toks/s, output: 1016.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:05, 20.89it/s, est. speed input: 1169.67 toks/s, output: 1145.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:05, 22.71it/s, est. speed input: 1211.58 toks/s, output: 1401.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:02<00:07, 15.04it/s, est. speed input: 1076.16 toks/s, output: 1386.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:03<00:19,  5.87it/s, est. speed input: 736.55 toks/s, output: 1052.87 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:07<00:57,  1.95it/s, est. speed input: 386.31 toks/s, output: 715.95 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:07<00:55,  2.03it/s, est. speed input: 375.93 toks/s, output: 779.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:08<00:56,  1.95it/s, est. speed input: 355.54 toks/s, output: 819.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:10:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12007.5 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:13<02:24,  1.31s/it, est. speed input: 228.12 toks/s, output: 612.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:15<02:36,  1.44s/it, est. speed input: 203.19 toks/s, output: 630.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:15<02:09,  1.20s/it, est. speed input: 201.95 toks/s, output: 711.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:17<02:30,  1.40s/it, est. speed input: 182.05 toks/s, output: 725.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:17<01:58,  1.12s/it, est. speed input: 182.30 toks/s, output: 808.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:10:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9999.8 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:21<02:17,  1.32s/it, est. speed input: 160.66 toks/s, output: 863.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:21<01:54,  1.11s/it, est. speed input: 160.69 toks/s, output: 942.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:21<01:35,  1.06it/s, est. speed input: 160.11 toks/s, output: 1017.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:22<01:30,  1.11it/s, est. speed input: 157.31 toks/s, output: 1077.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:23<01:16,  1.30it/s, est. speed input: 157.28 toks/s, output: 1151.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:23<01:11,  1.38it/s, est. speed input: 155.92 toks/s, output: 1215.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:24<00:41,  2.31it/s, est. speed input: 159.83 toks/s, output: 1465.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:24<00:30,  3.11it/s, est. speed input: 163.48 toks/s, output: 1638.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:25<00:39,  2.37it/s, est. speed input: 160.71 toks/s, output: 1679.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:25<00:36,  2.55it/s, est. speed input: 161.34 toks/s, output: 1754.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:26<00:36,  2.46it/s, est. speed input: 160.57 toks/s, output: 1881.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:26<00:31,  2.86it/s, est. speed input: 161.97 toks/s, output: 1965.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:26<00:28,  3.12it/s, est. speed input: 162.94 toks/s, output: 2042.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:27<00:36,  2.39it/s, est. speed input: 161.12 toks/s, output: 2083.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:27<00:31,  2.70it/s, est. speed input: 161.95 toks/s, output: 2159.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:28<00:31,  2.70it/s, est. speed input: 161.96 toks/s, output: 2288.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:28<00:28,  2.92it/s, est. speed input: 162.73 toks/s, output: 2361.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:29<00:36,  2.24it/s, est. speed input: 160.61 toks/s, output: 2394.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:30<00:40,  1.99it/s, est. speed input: 159.08 toks/s, output: 2435.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:10:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8339.6 tokens/s, Running: 80 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.5%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:30<00:32,  2.48it/s, est. speed input: 160.40 toks/s, output: 2517.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:31<00:51,  1.53it/s, est. speed input: 155.70 toks/s, output: 2507.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:32<00:52,  1.48it/s, est. speed input: 154.05 toks/s, output: 2544.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:34<01:14,  1.04it/s, est. speed input: 148.33 toks/s, output: 2513.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:34<00:59,  1.27it/s, est. speed input: 148.55 toks/s, output: 2581.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:35<01:11,  1.05it/s, est. speed input: 144.67 toks/s, output: 2577.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:37<01:23,  1.12s/it, est. speed input: 140.37 toks/s, output: 2564.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:37<01:05,  1.12it/s, est. speed input: 140.61 toks/s, output: 2633.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:37<00:47,  1.50it/s, est. speed input: 141.65 toks/s, output: 2716.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:38<00:50,  1.40it/s, est. speed input: 140.20 toks/s, output: 2750.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:39<00:54,  1.30it/s, est. speed input: 138.60 toks/s, output: 2781.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:39<00:43,  1.59it/s, est. speed input: 139.13 toks/s, output: 2853.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:40<00:25,  2.67it/s, est. speed input: 141.63 toks/s, output: 3027.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:11:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6873.9 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.6%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:41<00:40,  1.61it/s, est. speed input: 138.39 toks/s, output: 3019.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:42<00:36,  1.74it/s, est. speed input: 137.91 toks/s, output: 3130.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:42<00:31,  1.99it/s, est. speed input: 138.49 toks/s, output: 3204.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:25,  2.38it/s, est. speed input: 139.35 toks/s, output: 3285.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:42<00:16,  3.60it/s, est. speed input: 141.71 toks/s, output: 3459.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:43<00:19,  3.04it/s, est. speed input: 141.45 toks/s, output: 3512.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:44<00:23,  2.52it/s, est. speed input: 140.87 toks/s, output: 3557.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:44<00:14,  3.75it/s, est. speed input: 143.05 toks/s, output: 3731.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:44<00:13,  4.02it/s, est. speed input: 143.80 toks/s, output: 3808.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:44<00:16,  3.33it/s, est. speed input: 143.66 toks/s, output: 3863.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:45<00:19,  2.72it/s, est. speed input: 143.22 toks/s, output: 3909.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:46<00:17,  2.99it/s, est. speed input: 144.01 toks/s, output: 4047.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:46<00:14,  3.42it/s, est. speed input: 144.84 toks/s, output: 4127.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:46<00:11,  4.16it/s, est. speed input: 146.40 toks/s, output: 4286.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:47<00:22,  2.05it/s, est. speed input: 143.57 toks/s, output: 4260.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:48<00:18,  2.45it/s, est. speed input: 144.35 toks/s, output: 4341.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:48<00:12,  3.42it/s, est. speed input: 146.11 toks/s, output: 4508.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:49<00:19,  2.16it/s, est. speed input: 144.23 toks/s, output: 4507.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:49<00:18,  2.25it/s, est. speed input: 144.28 toks/s, output: 4567.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:11:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5614.0 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:50<00:24,  1.65it/s, est. speed input: 142.46 toks/s, output: 4566.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:51<00:22,  1.81it/s, est. speed input: 142.46 toks/s, output: 4625.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:51<00:20,  1.92it/s, est. speed input: 142.41 toks/s, output: 4681.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:52<00:19,  1.95it/s, est. speed input: 142.23 toks/s, output: 4733.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:52<00:15,  2.41it/s, est. speed input: 142.89 toks/s, output: 4813.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:53<00:26,  1.36it/s, est. speed input: 140.00 toks/s, output: 4775.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:54<00:21,  1.63it/s, est. speed input: 140.31 toks/s, output: 4843.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:54<00:15,  2.11it/s, est. speed input: 140.95 toks/s, output: 4982.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:54<00:12,  2.51it/s, est. speed input: 141.60 toks/s, output: 5064.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:55<00:15,  2.04it/s, est. speed input: 140.77 toks/s, output: 5093.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:55<00:12,  2.47it/s, est. speed input: 141.45 toks/s, output: 5175.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:55<00:07,  3.81it/s, est. speed input: 143.18 toks/s, output: 5356.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:56<00:06,  4.21it/s, est. speed input: 143.85 toks/s, output: 5439.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:56<00:04,  5.94it/s, est. speed input: 145.64 toks/s, output: 5623.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:56<00:05,  4.15it/s, est. speed input: 145.43 toks/s, output: 5673.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:57<00:06,  3.72it/s, est. speed input: 145.57 toks/s, output: 5736.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:58<00:12,  1.83it/s, est. speed input: 143.24 toks/s, output: 5704.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:59<00:12,  1.66it/s, est. speed input: 142.42 toks/s, output: 5732.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:59<00:10,  1.92it/s, est. speed input: 142.70 toks/s, output: 5804.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:59<00:08,  2.27it/s, est. speed input: 143.16 toks/s, output: 5881.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:11:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4011.3 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:00<00:09,  1.98it/s, est. speed input: 142.58 toks/s, output: 5919.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:00<00:07,  2.40it/s, est. speed input: 143.18 toks/s, output: 6001.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:01<00:04,  3.08it/s, est. speed input: 144.13 toks/s, output: 6164.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:02<00:08,  1.63it/s, est. speed input: 141.58 toks/s, output: 6120.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:04,  2.45it/s, est. speed input: 142.99 toks/s, output: 6306.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:03<00:05,  1.94it/s, est. speed input: 141.97 toks/s, output: 6325.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:03<00:04,  2.41it/s, est. speed input: 142.68 toks/s, output: 6421.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:04<00:03,  2.59it/s, est. speed input: 143.03 toks/s, output: 6497.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:04<00:03,  2.39it/s, est. speed input: 142.83 toks/s, output: 6553.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:05<00:03,  1.99it/s, est. speed input: 142.17 toks/s, output: 6590.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:05<00:02,  2.27it/s, est. speed input: 142.52 toks/s, output: 6670.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:06<00:02,  2.04it/s, est. speed input: 142.14 toks/s, output: 6718.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:07<00:02,  1.59it/s, est. speed input: 141.01 toks/s, output: 6734.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:08<00:02,  1.04it/s, est. speed input: 138.26 toks/s, output: 6677.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:09<00:01,  1.02it/s, est. speed input: 137.13 toks/s, output: 6696.24 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  1.02it/s, est. speed input: 138.90 toks/s, output: 6930.35 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  2.29it/s, est. speed input: 138.90 toks/s, output: 6930.35 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:11:37 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:11:37 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:11:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:11:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:11:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:11:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:11:38 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:11:38 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:11:38 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 02:11:38 [abstract.py:306] It took 0.644858 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:11:38 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0126
+  • Training Accuracy: 0.0938
+  • Mean Generation Length: 3031.1001
+  • Training FLOPS: 47.82 TFLOPS (11.95 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 348.52s
+  • policy_training: 226.07s (64.9%)
+  • generation: 75.04s (21.5%)
+  • teacher_logprob_inference: 42.39s (12.2%)
+  • policy_training/sharding_data: 4.40s (1.3%)
+  • prepare_for_generation: 3.50s (1.0%)
+  • policy_training/submit_training_futures: 1.89s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.4%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 45: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:08 [abstract.py:324] It took 0.054369 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06563809e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52225494e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64537735e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06563809e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52225494e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64537735e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06563809e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52225494e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64537735e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06563809e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52225494e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.64537735e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06562891e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54081955e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65354614e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.26050816e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.50672388e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #45] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43442017e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06562891e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54081955e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65354614e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69264679e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.32103558e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #45] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23445702e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06562891e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54081955e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65354614e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30613251e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.70591593e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #45] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16569061e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06562891e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54081955e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.65354614e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26053085e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.68632221e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #45] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31786316e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 45: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:11 [abstract.py:324] It took 0.041303 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:11 [loggers.py:236] Engine 000: Avg prompt throughput: 33.6 tokens/s, Avg generation throughput: 56.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13473.78it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  5.96it/s, est. speed input: 363.36 toks/s, output: 77.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:04, 34.35it/s, est. speed input: 1788.48 toks/s, output: 509.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 29.49it/s, est. speed input: 1691.69 toks/s, output: 657.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 28.23it/s, est. speed input: 1670.51 toks/s, output: 812.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:10, 13.09it/s, est. speed input: 1026.67 toks/s, output: 709.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:09, 14.02it/s, est. speed input: 1029.95 toks/s, output: 894.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:08, 14.68it/s, est. speed input: 1023.02 toks/s, output: 1113.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:09, 13.19it/s, est. speed input: 971.31 toks/s, output: 1172.24 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:05<01:03,  2.00it/s, est. speed input: 328.11 toks/s, output: 504.43 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12532.9 tokens/s, Running: 127 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:11<02:10,  1.03s/it, est. speed input: 180.49 toks/s, output: 437.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:12<02:07,  1.02s/it, est. speed input: 171.41 toks/s, output: 498.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:13<02:00,  1.03it/s, est. speed input: 166.67 toks/s, output: 564.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:17<03:31,  1.72s/it, est. speed input: 125.78 toks/s, output: 506.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10710.1 tokens/s, Running: 123 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:22<04:56,  2.43s/it, est. speed input: 101.19 toks/s, output: 486.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:24<04:32,  2.25s/it, est. speed input: 96.61 toks/s, output: 541.52 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:25<03:37,  1.81s/it, est. speed input: 97.08 toks/s, output: 618.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:27<03:41,  1.86s/it, est. speed input: 92.10 toks/s, output: 660.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:29<03:16,  1.68s/it, est. speed input: 87.25 toks/s, output: 764.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9592.4 tokens/s, Running: 117 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:31<03:07,  1.62s/it, est. speed input: 85.17 toks/s, output: 815.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:31<02:23,  1.24s/it, est. speed input: 86.72 toks/s, output: 898.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:32<02:12,  1.16s/it, est. speed input: 86.11 toks/s, output: 958.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:35<03:15,  1.73s/it, est. speed input: 80.07 toks/s, output: 957.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:36<02:56,  1.58s/it, est. speed input: 79.07 toks/s, output: 1011.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:38<02:52,  1.56s/it, est. speed input: 77.58 toks/s, output: 1055.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:39<02:31,  1.38s/it, est. speed input: 77.31 toks/s, output: 1113.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:16:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8308.6 tokens/s, Running: 109 reqs, Waiting: 0 reqs, GPU KV cache usage: 16.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:41<02:18,  1.28s/it, est. speed input: 75.86 toks/s, output: 1213.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:42<02:02,  1.14s/it, est. speed input: 76.07 toks/s, output: 1276.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:42<01:34,  1.13it/s, est. speed input: 77.15 toks/s, output: 1354.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:42<00:56,  1.85it/s, est. speed input: 79.69 toks/s, output: 1515.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:42<00:48,  2.14it/s, est. speed input: 80.63 toks/s, output: 1591.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:44<01:07,  1.51it/s, est. speed input: 79.72 toks/s, output: 1628.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:46<01:42,  1.02s/it, est. speed input: 77.50 toks/s, output: 1640.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:47<01:39,  1.01it/s, est. speed input: 77.27 toks/s, output: 1690.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:47<01:31,  1.08it/s, est. speed input: 77.27 toks/s, output: 1745.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:48<01:31,  1.08it/s, est. speed input: 77.07 toks/s, output: 1794.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:48<01:09,  1.39it/s, est. speed input: 78.02 toks/s, output: 1868.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:49<00:59,  1.62it/s, est. speed input: 78.62 toks/s, output: 1935.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:50<00:59,  1.61it/s, est. speed input: 78.86 toks/s, output: 1992.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7555.8 tokens/s, Running: 94 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:51<00:57,  1.62it/s, est. speed input: 79.37 toks/s, output: 2106.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:51<00:39,  2.28it/s, est. speed input: 81.17 toks/s, output: 2255.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:53<01:14,  1.21it/s, est. speed input: 78.92 toks/s, output: 2243.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:55<01:27,  1.02it/s, est. speed input: 77.90 toks/s, output: 2265.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:57<01:50,  1.26s/it, est. speed input: 76.13 toks/s, output: 2264.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:57<01:24,  1.03it/s, est. speed input: 76.90 toks/s, output: 2337.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:57<01:05,  1.31it/s, est. speed input: 77.69 toks/s, output: 2410.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:58<01:12,  1.18it/s, est. speed input: 77.33 toks/s, output: 2447.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6846.8 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 19.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [01:01<01:53,  1.35s/it, est. speed input: 75.06 toks/s, output: 2424.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [01:01<01:22,  1.01it/s, est. speed input: 75.93 toks/s, output: 2500.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [01:02<01:02,  1.30it/s, est. speed input: 76.68 toks/s, output: 2620.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [01:03<01:00,  1.33it/s, est. speed input: 76.79 toks/s, output: 2670.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [01:03<00:45,  1.71it/s, est. speed input: 77.86 toks/s, output: 2801.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [01:04<00:40,  1.86it/s, est. speed input: 78.62 toks/s, output: 2920.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [01:06<00:58,  1.29it/s, est. speed input: 77.53 toks/s, output: 2926.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [01:06<00:47,  1.57it/s, est. speed input: 78.23 toks/s, output: 2999.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [01:07<00:42,  1.72it/s, est. speed input: 78.66 toks/s, output: 3060.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [01:07<00:46,  1.55it/s, est. speed input: 78.58 toks/s, output: 3103.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [01:08<00:45,  1.55it/s, est. speed input: 78.72 toks/s, output: 3153.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5683.6 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 18.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [01:10<01:04,  1.08it/s, est. speed input: 77.72 toks/s, output: 3159.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [01:10<00:53,  1.29it/s, est. speed input: 78.14 toks/s, output: 3221.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [01:11<00:50,  1.35it/s, est. speed input: 78.26 toks/s, output: 3271.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [01:12<00:44,  1.47it/s, est. speed input: 78.62 toks/s, output: 3374.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [01:13<00:33,  1.90it/s, est. speed input: 79.67 toks/s, output: 3507.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [01:14<00:51,  1.22it/s, est. speed input: 78.52 toks/s, output: 3500.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [01:15<00:42,  1.46it/s, est. speed input: 79.06 toks/s, output: 3567.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [01:17<01:05,  1.07s/it, est. speed input: 77.60 toks/s, output: 3547.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [01:17<00:55,  1.09it/s, est. speed input: 77.90 toks/s, output: 3604.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [01:18<00:51,  1.15it/s, est. speed input: 77.91 toks/s, output: 3650.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5435.3 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 17.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [01:20<01:00,  1.04s/it, est. speed input: 77.23 toks/s, output: 3664.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [01:20<00:45,  1.26it/s, est. speed input: 77.79 toks/s, output: 3736.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [01:20<00:35,  1.59it/s, est. speed input: 78.30 toks/s, output: 3806.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [01:21<00:38,  1.44it/s, est. speed input: 78.21 toks/s, output: 3846.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [01:21<00:32,  1.66it/s, est. speed input: 78.56 toks/s, output: 3909.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [01:22<00:36,  1.46it/s, est. speed input: 78.48 toks/s, output: 3948.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:22<00:30,  1.68it/s, est. speed input: 78.86 toks/s, output: 4011.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:25<00:53,  1.04s/it, est. speed input: 77.63 toks/s, output: 3993.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:28<01:04,  1.32s/it, est. speed input: 76.16 toks/s, output: 4004.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [01:28<00:49,  1.03s/it, est. speed input: 76.71 toks/s, output: 4078.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:29<00:42,  1.10it/s, est. speed input: 76.92 toks/s, output: 4133.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4531.2 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 15.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:30<00:36,  1.22it/s, est. speed input: 77.06 toks/s, output: 4231.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:31<00:33,  1.32it/s, est. speed input: 77.28 toks/s, output: 4287.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:32<00:30,  1.36it/s, est. speed input: 77.43 toks/s, output: 4384.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:33<00:37,  1.08it/s, est. speed input: 76.80 toks/s, output: 4392.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:34<00:25,  1.53it/s, est. speed input: 77.69 toks/s, output: 4535.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:34<00:20,  1.83it/s, est. speed input: 78.22 toks/s, output: 4610.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:34<00:14,  2.54it/s, est. speed input: 79.26 toks/s, output: 4760.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:35<00:12,  2.76it/s, est. speed input: 79.70 toks/s, output: 4830.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:36<00:23,  1.44it/s, est. speed input: 78.85 toks/s, output: 4825.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:38<00:27,  1.21it/s, est. speed input: 78.47 toks/s, output: 4847.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:38<00:20,  1.53it/s, est. speed input: 78.94 toks/s, output: 4922.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:38<00:15,  1.98it/s, est. speed input: 79.46 toks/s, output: 5000.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:38<00:00,  1.98it/s, est. speed input: 98.41 toks/s, output: 7580.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:38<00:00,  1.63it/s, est. speed input: 98.41 toks/s, output: 7580.06 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:57 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:57 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:17:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:17:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:17:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:17:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:17:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:17:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:17:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 02:17:57 [abstract.py:306] It took 0.850697 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:17:57 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0094
+  • Training Accuracy: 0.0750
+  • Mean Generation Length: 4663.9126
+  • Training FLOPS: 54.86 TFLOPS (13.71 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 482.55s
+  • policy_training: 312.72s (64.8%)
+  • generation: 106.05s (22.0%)
+  • teacher_logprob_inference: 58.57s (12.1%)
+  • policy_training/sharding_data: 4.31s (0.9%)
+  • prepare_for_generation: 3.69s (0.8%)
+  • policy_training/submit_training_futures: 1.73s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.69s (0.4%)
+  • training_prep: 0.85s (0.2%)
+  • teacher_logprob_inference_prep: 0.47s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 10/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 46: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:11 [abstract.py:324] It took 0.052579 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06620625e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52605553e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65018158e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06620625e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52605553e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65018158e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06620625e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52605553e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65018158e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06620625e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52605553e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65018158e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06627715e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53493195e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66132507e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23985844e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.45540142e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #46] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43436218e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06627715e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53493195e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66132507e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69579849e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33516150e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #46] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.24475899e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06627715e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53493195e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66132507e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31188049e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=4.66442108e-02
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #46] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17687836e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06627715e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53493195e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.66132507e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25481606e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67813320e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #46] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31491806e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 46: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:14 [abstract.py:324] It took 0.040996 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12855.37it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:15 [loggers.py:236] Engine 000: Avg prompt throughput: 24.6 tokens/s, Avg generation throughput: 84.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:48,  1.47it/s, est. speed input: 86.72 toks/s, output: 17.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:08, 17.72it/s, est. speed input: 825.52 toks/s, output: 221.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:01<00:08, 16.89it/s, est. speed input: 863.47 toks/s, output: 328.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:08, 16.30it/s, est. speed input: 872.82 toks/s, output: 486.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:07, 18.23it/s, est. speed input: 941.17 toks/s, output: 662.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:06, 19.86it/s, est. speed input: 988.80 toks/s, output: 808.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:06, 21.21it/s, est. speed input: 1027.54 toks/s, output: 951.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:06, 20.43it/s, est. speed input: 1035.56 toks/s, output: 1067.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:02<00:08, 15.21it/s, est. speed input: 965.24 toks/s, output: 1120.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:02<00:10, 12.14it/s, est. speed input: 903.62 toks/s, output: 1143.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:02<00:11, 10.32it/s, est. speed input: 854.17 toks/s, output: 1189.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:05<00:52,  2.23it/s, est. speed input: 434.60 toks/s, output: 742.03 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:06<00:48,  2.42it/s, est. speed input: 429.25 toks/s, output: 808.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:08<01:38,  1.18it/s, est. speed input: 297.69 toks/s, output: 643.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12175.6 tokens/s, Running: 116 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:13<03:00,  1.57s/it, est. speed input: 204.14 toks/s, output: 525.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:13<02:34,  1.36s/it, est. speed input: 199.09 toks/s, output: 596.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:17<03:41,  1.96s/it, est. speed input: 159.64 toks/s, output: 561.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:19<03:44,  2.01s/it, est. speed input: 145.58 toks/s, output: 593.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:20<02:53,  1.57s/it, est. speed input: 145.84 toks/s, output: 675.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10254.3 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:22<02:57,  1.61s/it, est. speed input: 137.11 toks/s, output: 714.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:22<02:21,  1.30s/it, est. speed input: 136.72 toks/s, output: 790.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:22<01:50,  1.03s/it, est. speed input: 137.19 toks/s, output: 869.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:23<01:25,  1.25it/s, est. speed input: 138.35 toks/s, output: 952.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:23<01:15,  1.40it/s, est. speed input: 137.86 toks/s, output: 1023.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:23<01:02,  1.68it/s, est. speed input: 138.56 toks/s, output: 1101.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:25<01:15,  1.38it/s, est. speed input: 135.19 toks/s, output: 1146.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:27<02:04,  1.21s/it, est. speed input: 125.79 toks/s, output: 1139.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:28<01:48,  1.07s/it, est. speed input: 124.61 toks/s, output: 1199.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:29<01:56,  1.15s/it, est. speed input: 120.95 toks/s, output: 1235.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:30<01:45,  1.05s/it, est. speed input: 119.70 toks/s, output: 1291.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8862.8 tokens/s, Running: 100 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:31<01:40,  1.02s/it, est. speed input: 117.99 toks/s, output: 1342.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:31<01:20,  1.22it/s, est. speed input: 118.56 toks/s, output: 1416.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:32<01:07,  1.42it/s, est. speed input: 118.13 toks/s, output: 1543.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:33<01:07,  1.41it/s, est. speed input: 117.40 toks/s, output: 1598.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:34<01:03,  1.47it/s, est. speed input: 117.13 toks/s, output: 1660.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:34<00:48,  1.90it/s, est. speed input: 118.45 toks/s, output: 1743.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:34<00:39,  2.32it/s, est. speed input: 119.55 toks/s, output: 1823.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:34<00:34,  2.62it/s, est. speed input: 120.44 toks/s, output: 1899.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:36<01:19,  1.13it/s, est. speed input: 115.09 toks/s, output: 1878.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:38<01:33,  1.05s/it, est. speed input: 112.27 toks/s, output: 1894.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:38<01:09,  1.27it/s, est. speed input: 113.38 toks/s, output: 1975.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:38<00:54,  1.58it/s, est. speed input: 114.22 toks/s, output: 2050.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:39<00:56,  1.53it/s, est. speed input: 113.76 toks/s, output: 2101.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:39<00:33,  2.54it/s, est. speed input: 116.29 toks/s, output: 2268.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:24:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7660.9 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.7%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:40<00:46,  1.77it/s, est. speed input: 114.65 toks/s, output: 2295.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:42<01:10,  1.17it/s, est. speed input: 111.52 toks/s, output: 2292.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:43<01:11,  1.14it/s, est. speed input: 110.52 toks/s, output: 2330.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:43<00:47,  1.67it/s, est. speed input: 112.13 toks/s, output: 2480.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:43<00:38,  2.01it/s, est. speed input: 113.03 toks/s, output: 2558.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:44<00:39,  1.95it/s, est. speed input: 112.99 toks/s, output: 2614.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:44<00:37,  2.04it/s, est. speed input: 113.30 toks/s, output: 2677.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:44<00:29,  2.54it/s, est. speed input: 114.33 toks/s, output: 2756.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:45<00:31,  2.35it/s, est. speed input: 114.40 toks/s, output: 2814.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:45<00:25,  2.86it/s, est. speed input: 115.31 toks/s, output: 2891.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:46<00:23,  3.07it/s, est. speed input: 116.47 toks/s, output: 3029.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:46<00:23,  3.04it/s, est. speed input: 116.91 toks/s, output: 3095.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:46<00:14,  4.55it/s, est. speed input: 119.23 toks/s, output: 3262.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:46<00:11,  5.97it/s, est. speed input: 121.48 toks/s, output: 3427.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:49<00:38,  1.69it/s, est. speed input: 117.28 toks/s, output: 3361.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:49<00:36,  1.76it/s, est. speed input: 117.34 toks/s, output: 3416.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:50<00:40,  1.57it/s, est. speed input: 116.57 toks/s, output: 3446.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6389.1 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:51<00:49,  1.25it/s, est. speed input: 114.85 toks/s, output: 3449.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:52<00:43,  1.39it/s, est. speed input: 114.92 toks/s, output: 3505.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:53<00:46,  1.30it/s, est. speed input: 114.09 toks/s, output: 3534.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:53<00:30,  1.91it/s, est. speed input: 115.42 toks/s, output: 3681.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:54<00:32,  1.74it/s, est. speed input: 114.94 toks/s, output: 3720.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:55<00:35,  1.58it/s, est. speed input: 114.35 toks/s, output: 3754.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:56<00:45,  1.21it/s, est. speed input: 112.67 toks/s, output: 3754.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:58<00:47,  1.12it/s, est. speed input: 110.97 toks/s, output: 3803.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:58<00:41,  1.24it/s, est. speed input: 111.01 toks/s, output: 3859.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:00<00:50,  1.01it/s, est. speed input: 109.18 toks/s, output: 3851.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5723.6 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:00<00:33,  1.44it/s, est. speed input: 110.16 toks/s, output: 3995.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:01<00:25,  1.83it/s, est. speed input: 111.03 toks/s, output: 4136.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:02<00:27,  1.66it/s, est. speed input: 110.58 toks/s, output: 4173.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [01:02<00:25,  1.74it/s, est. speed input: 110.65 toks/s, output: 4231.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:03<00:24,  1.82it/s, est. speed input: 110.77 toks/s, output: 4291.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:03<00:23,  1.84it/s, est. speed input: 110.79 toks/s, output: 4346.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:03<00:18,  2.31it/s, est. speed input: 111.54 toks/s, output: 4428.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:05<00:25,  1.62it/s, est. speed input: 110.56 toks/s, output: 4445.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:05<00:20,  1.95it/s, est. speed input: 111.06 toks/s, output: 4520.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:07<00:38,  1.02it/s, est. speed input: 108.45 toks/s, output: 4469.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:07<00:27,  1.37it/s, est. speed input: 109.16 toks/s, output: 4552.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:07<00:23,  1.59it/s, est. speed input: 109.41 toks/s, output: 4617.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:10<00:38,  1.07s/it, est. speed input: 106.96 toks/s, output: 4570.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:10<00:29,  1.18it/s, est. speed input: 107.30 toks/s, output: 4641.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4358.9 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:10<00:25,  1.34it/s, est. speed input: 107.40 toks/s, output: 4701.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:11<00:27,  1.21it/s, est. speed input: 106.71 toks/s, output: 4727.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:12<00:16,  1.88it/s, est. speed input: 107.79 toks/s, output: 4888.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:13<00:18,  1.59it/s, est. speed input: 107.27 toks/s, output: 4919.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:14<00:21,  1.34it/s, est. speed input: 106.55 toks/s, output: 4941.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:14<00:20,  1.40it/s, est. speed input: 106.50 toks/s, output: 4994.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:15<00:16,  1.68it/s, est. speed input: 106.87 toks/s, output: 5069.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:15<00:12,  2.12it/s, est. speed input: 107.43 toks/s, output: 5153.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:15<00:10,  2.38it/s, est. speed input: 107.81 toks/s, output: 5227.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:15<00:08,  2.95it/s, est. speed input: 108.43 toks/s, output: 5312.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:16<00:09,  2.35it/s, est. speed input: 108.51 toks/s, output: 5427.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:18<00:13,  1.57it/s, est. speed input: 107.47 toks/s, output: 5433.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:19<00:15,  1.29it/s, est. speed input: 106.65 toks/s, output: 5451.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:19<00:12,  1.50it/s, est. speed input: 106.88 toks/s, output: 5522.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3479.2 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:20<00:13,  1.38it/s, est. speed input: 106.46 toks/s, output: 5559.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:20<00:09,  1.77it/s, est. speed input: 107.00 toks/s, output: 5646.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:22<00:13,  1.22it/s, est. speed input: 105.85 toks/s, output: 5646.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:22<00:11,  1.35it/s, est. speed input: 105.86 toks/s, output: 5707.77 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:22<00:00,  1.35it/s, est. speed input: 116.74 toks/s, output: 7193.35 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:22<00:00,  1.93it/s, est. speed input: 116.74 toks/s, output: 7193.35 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:42 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:42 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:25:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:25:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:25:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:25:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:25:43 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:25:43 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:25:43 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 02:25:43 [abstract.py:306] It took 0.854166 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:25:43 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.1875
+  • Mean Generation Length: 3718.6875
+  • Training FLOPS: 51.61 TFLOPS (12.90 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 403.86s
+  • policy_training: 261.62s (64.8%)
+  • generation: 88.87s (22.0%)
+  • teacher_logprob_inference: 48.41s (12.0%)
+  • policy_training/sharding_data: 4.30s (1.1%)
+  • prepare_for_generation: 3.50s (0.9%)
+  • policy_training/submit_training_futures: 1.81s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.66s (0.4%)
+  • training_prep: 0.83s (0.2%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 47: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:30:55 [abstract.py:324] It took 0.058873 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06675801e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52805710e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65044708e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06675801e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52805710e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65044708e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06675801e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52805710e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65044708e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06675801e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52805710e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65044708e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06685508e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51968880e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64706268e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23069611e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.41858339e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #47] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43347137e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06685508e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51968880e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64706268e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70126343e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34353676e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #47] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.22751083e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06685508e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51968880e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64706268e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31434937e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.31928730e-02
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #47] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17801437e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06685508e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51968880e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.64706268e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26043701e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67401352e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #47] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31249527e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 47: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:30:58 [abstract.py:324] It took 0.037857 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13359.52it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:30:59 [loggers.py:236] Engine 000: Avg prompt throughput: 30.1 tokens/s, Avg generation throughput: 17.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:55,  2.87it/s, est. speed input: 172.22 toks/s, output: 51.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:14, 10.45it/s, est. speed input: 552.21 toks/s, output: 230.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:09, 16.19it/s, est. speed input: 803.22 toks/s, output: 437.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:08, 18.40it/s, est. speed input: 906.97 toks/s, output: 584.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 24.12it/s, est. speed input: 1100.07 toks/s, output: 856.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 25.88it/s, est. speed input: 1196.33 toks/s, output: 1054.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:06, 21.64it/s, est. speed input: 1156.82 toks/s, output: 1122.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 22.94it/s, est. speed input: 1196.39 toks/s, output: 1282.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:05, 23.99it/s, est. speed input: 1241.71 toks/s, output: 1523.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:04, 25.21it/s, est. speed input: 1283.80 toks/s, output: 1732.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:04, 25.26it/s, est. speed input: 1306.43 toks/s, output: 1933.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:02<00:06, 17.20it/s, est. speed input: 1187.89 toks/s, output: 1904.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:02<00:08, 12.88it/s, est. speed input: 1078.77 toks/s, output: 1879.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:02<00:10, 10.22it/s, est. speed input: 990.97 toks/s, output: 1837.11 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:04<00:20,  5.26it/s, est. speed input: 768.07 toks/s, output: 1549.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:04<00:21,  5.14it/s, est. speed input: 741.87 toks/s, output: 1567.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:05<00:33,  3.22it/s, est. speed input: 623.71 toks/s, output: 1393.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:05<00:44,  2.43it/s, est. speed input: 545.51 toks/s, output: 1297.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:06<00:27,  3.74it/s, est. speed input: 551.35 toks/s, output: 1538.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:06<00:28,  3.62it/s, est. speed input: 533.82 toks/s, output: 1566.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:06<00:24,  4.11it/s, est. speed input: 534.25 toks/s, output: 1641.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:07<00:30,  3.31it/s, est. speed input: 504.87 toks/s, output: 1627.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:07<00:38,  2.62it/s, est. speed input: 471.97 toks/s, output: 1598.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:08<00:45,  2.16it/s, est. speed input: 440.30 toks/s, output: 1569.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:09<00:52,  1.86it/s, est. speed input: 411.29 toks/s, output: 1546.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:31:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11651.5 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:10<01:02,  1.55it/s, est. speed input: 379.79 toks/s, output: 1508.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:10<00:52,  1.81it/s, est. speed input: 374.05 toks/s, output: 1566.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:10<00:35,  2.66it/s, est. speed input: 374.51 toks/s, output: 1724.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:10<00:24,  3.81it/s, est. speed input: 379.57 toks/s, output: 1902.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:12<00:40,  2.26it/s, est. speed input: 350.47 toks/s, output: 1833.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:12<00:26,  3.37it/s, est. speed input: 356.88 toks/s, output: 2019.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:13<00:37,  2.32it/s, est. speed input: 330.29 toks/s, output: 2014.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:16<01:20,  1.07it/s, est. speed input: 275.42 toks/s, output: 1760.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:18<01:37,  1.14s/it, est. speed input: 250.71 toks/s, output: 1683.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:19<01:27,  1.04s/it, est. speed input: 244.58 toks/s, output: 1723.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:19<00:52,  1.57it/s, est. speed input: 249.71 toks/s, output: 1917.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:19<00:54,  1.48it/s, est. speed input: 242.74 toks/s, output: 1942.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:31:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8917.9 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:21<01:05,  1.23it/s, est. speed input: 231.22 toks/s, output: 1930.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:21<00:42,  1.84it/s, est. speed input: 233.49 toks/s, output: 2106.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:21<00:29,  2.60it/s, est. speed input: 236.51 toks/s, output: 2287.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:21<00:25,  2.97it/s, est. speed input: 237.58 toks/s, output: 2373.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:22<00:23,  3.11it/s, est. speed input: 237.42 toks/s, output: 2446.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:22<00:20,  3.53it/s, est. speed input: 238.44 toks/s, output: 2531.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:22<00:17,  4.09it/s, est. speed input: 239.72 toks/s, output: 2618.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:23<00:24,  2.86it/s, est. speed input: 234.96 toks/s, output: 2708.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:23<00:26,  2.59it/s, est. speed input: 232.45 toks/s, output: 2753.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:24<00:23,  2.86it/s, est. speed input: 232.72 toks/s, output: 2828.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:24<00:20,  3.25it/s, est. speed input: 233.43 toks/s, output: 2908.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:25<00:31,  2.06it/s, est. speed input: 225.04 toks/s, output: 2944.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:26<00:36,  1.77it/s, est. speed input: 220.42 toks/s, output: 2957.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<00:57,  1.10it/s, est. speed input: 207.64 toks/s, output: 2860.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:28<00:35,  1.70it/s, est. speed input: 209.79 toks/s, output: 3039.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:29<00:39,  1.53it/s, est. speed input: 205.60 toks/s, output: 3054.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:31:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7206.1 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:30<00:41,  1.42it/s, est. speed input: 201.84 toks/s, output: 3072.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:30<00:25,  2.21it/s, est. speed input: 204.41 toks/s, output: 3259.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:31<00:18,  2.88it/s, est. speed input: 205.74 toks/s, output: 3495.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:11,  4.45it/s, est. speed input: 210.70 toks/s, output: 3792.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:33<00:20,  2.40it/s, est. speed input: 202.70 toks/s, output: 3793.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:34<00:22,  2.15it/s, est. speed input: 200.19 toks/s, output: 3819.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:34<00:17,  2.62it/s, est. speed input: 201.28 toks/s, output: 3984.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:34<00:17,  2.64it/s, est. speed input: 200.93 toks/s, output: 4048.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:35<00:18,  2.36it/s, est. speed input: 199.30 toks/s, output: 4086.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:20,  2.10it/s, est. speed input: 197.37 toks/s, output: 4119.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:36<00:16,  2.57it/s, est. speed input: 198.33 toks/s, output: 4211.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:36<00:17,  2.35it/s, est. speed input: 197.21 toks/s, output: 4258.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:37<00:15,  2.62it/s, est. speed input: 197.38 toks/s, output: 4334.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:37<00:09,  3.81it/s, est. speed input: 199.51 toks/s, output: 4522.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:09,  3.82it/s, est. speed input: 199.75 toks/s, output: 4598.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:37<00:08,  4.17it/s, est. speed input: 200.49 toks/s, output: 4683.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:38<00:08,  4.05it/s, est. speed input: 200.96 toks/s, output: 4834.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:39<00:16,  2.03it/s, est. speed input: 195.81 toks/s, output: 4783.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:31:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5525.5 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:40<00:14,  2.21it/s, est. speed input: 195.08 toks/s, output: 4905.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:40<00:12,  2.50it/s, est. speed input: 195.55 toks/s, output: 4988.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:40<00:10,  2.79it/s, est. speed input: 196.00 toks/s, output: 5070.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:41<00:09,  3.08it/s, est. speed input: 196.42 toks/s, output: 5151.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:07,  3.38it/s, est. speed input: 196.88 toks/s, output: 5234.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:41<00:08,  2.93it/s, est. speed input: 196.19 toks/s, output: 5287.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:41<00:05,  4.04it/s, est. speed input: 197.92 toks/s, output: 5475.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:42<00:06,  3.38it/s, est. speed input: 197.28 toks/s, output: 5600.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:42<00:05,  3.94it/s, est. speed input: 198.21 toks/s, output: 5698.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:42<00:04,  4.21it/s, est. speed input: 198.71 toks/s, output: 5785.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:43<00:05,  3.22it/s, est. speed input: 197.62 toks/s, output: 5895.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:03,  4.63it/s, est. speed input: 199.98 toks/s, output: 6105.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:44<00:03,  4.23it/s, est. speed input: 199.94 toks/s, output: 6174.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:45<00:05,  2.58it/s, est. speed input: 197.29 toks/s, output: 6165.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:45<00:05,  2.36it/s, est. speed input: 196.30 toks/s, output: 6207.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:46<00:05,  2.28it/s, est. speed input: 195.54 toks/s, output: 6258.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:47<00:05,  1.86it/s, est. speed input: 192.59 toks/s, output: 6310.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:48<00:05,  1.77it/s, est. speed input: 191.20 toks/s, output: 6344.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:49<00:05,  1.50it/s, est. speed input: 188.68 toks/s, output: 6339.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:49<00:03,  1.93it/s, est. speed input: 189.55 toks/s, output: 6447.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:49<00:02,  2.00it/s, est. speed input: 188.99 toks/s, output: 6510.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:49<00:02,  2.27it/s, est. speed input: 189.07 toks/s, output: 6596.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:31:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2647.1 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:02,  1.45it/s, est. speed input: 185.40 toks/s, output: 6552.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:53<00:03,  1.09s/it, est. speed input: 179.30 toks/s, output: 6426.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:53<00:01,  1.20it/s, est. speed input: 179.76 toks/s, output: 6533.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:55<00:01,  1.21s/it, est. speed input: 173.92 toks/s, output: 6419.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  1.53s/it, est. speed input: 168.11 toks/s, output: 6307.33 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  1.53s/it, est. speed input: 168.11 toks/s, output: 6307.33 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  2.76it/s, est. speed input: 168.11 toks/s, output: 6307.33 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:32:00 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:32:00 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:32:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:32:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:32:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:32:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:32:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:32:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:32:00 [abstract.py:306] It took 0.604448 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:32:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:32:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0154
+  • Training Accuracy: 0.1938
+  • Mean Generation Length: 2284.9751
+  • Training FLOPS: 44.80 TFLOPS (11.20 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 280.69s
+  • policy_training: 179.66s (64.0%)
+  • generation: 61.90s (22.1%)
+  • teacher_logprob_inference: 33.79s (12.0%)
+  • policy_training/sharding_data: 4.36s (1.6%)
+  • prepare_for_generation: 3.57s (1.3%)
+  • policy_training/submit_training_futures: 2.13s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 1.07s (0.4%)
+  • teacher_logprob_inference_prep: 0.52s (0.2%)
+  • get_topk_logits/shard_data: 0.05s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 48: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:35:36 [abstract.py:324] It took 0.051133 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06709355e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52148895e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65649261e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06709355e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52148895e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65649261e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06709355e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52148895e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65649261e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06709355e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52148895e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.65649261e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06716367e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50188675e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67276306e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.21545143e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.33797073e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #48] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43303513e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06716367e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50188675e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67276306e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68680954e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34721012e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #48] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.21864777e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06716367e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50188675e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67276306e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31691742e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.34255886e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #48] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14871368e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06716367e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50188675e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67276306e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26681595e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66540527e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #48] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31171326e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 48: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:35:39 [abstract.py:324] It took 0.038169 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13581.49it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:35:39 [loggers.py:236] Engine 000: Avg prompt throughput: 42.4 tokens/s, Avg generation throughput: 22.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:41,  3.81it/s, est. speed input: 228.51 toks/s, output: 45.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:03, 43.68it/s, est. speed input: 2124.59 toks/s, output: 619.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 36.56it/s, est. speed input: 2016.42 toks/s, output: 784.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 35.48it/s, est. speed input: 2028.13 toks/s, output: 1066.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:03, 39.93it/s, est. speed input: 2197.00 toks/s, output: 1484.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 41.07it/s, est. speed input: 2254.93 toks/s, output: 1742.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 43.03it/s, est. speed input: 2321.81 toks/s, output: 2021.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 44.32it/s, est. speed input: 2373.52 toks/s, output: 2282.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 49.24it/s, est. speed input: 2489.53 toks/s, output: 2723.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 40.82it/s, est. speed input: 2400.12 toks/s, output: 2909.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:02, 39.20it/s, est. speed input: 2382.95 toks/s, output: 3138.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:05<00:22,  3.94it/s, est. speed input: 735.43 toks/s, output: 1134.79 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:08<00:29,  2.84it/s, est. speed input: 552.92 toks/s, output: 1116.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:35:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10557.6 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:13<00:52,  1.55it/s, est. speed input: 354.89 toks/s, output: 952.71 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:15<00:56,  1.42it/s, est. speed input: 317.26 toks/s, output: 1043.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:16<00:54,  1.43it/s, est. speed input: 298.46 toks/s, output: 1174.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:18<01:07,  1.14it/s, est. speed input: 268.29 toks/s, output: 1152.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:35:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8293.1 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:20<01:12,  1.04it/s, est. speed input: 252.18 toks/s, output: 1177.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:21<01:18,  1.05s/it, est. speed input: 237.92 toks/s, output: 1205.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:22<01:08,  1.08it/s, est. speed input: 236.04 toks/s, output: 1289.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:22<00:58,  1.25it/s, est. speed input: 234.91 toks/s, output: 1375.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:23<00:52,  1.38it/s, est. speed input: 232.71 toks/s, output: 1454.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:23<00:48,  1.45it/s, est. speed input: 229.56 toks/s, output: 1525.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:23<00:39,  1.77it/s, est. speed input: 229.96 toks/s, output: 1617.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:24<00:38,  1.80it/s, est. speed input: 227.41 toks/s, output: 1689.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:26<01:04,  1.06it/s, est. speed input: 213.16 toks/s, output: 1671.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:28<01:24,  1.27s/it, est. speed input: 199.83 toks/s, output: 1656.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:28<00:47,  1.38it/s, est. speed input: 203.14 toks/s, output: 1859.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:29<00:42,  1.52it/s, est. speed input: 202.07 toks/s, output: 1936.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:29<00:33,  1.86it/s, est. speed input: 202.88 toks/s, output: 2029.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7175.4 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:31<01:04,  1.05s/it, est. speed input: 189.25 toks/s, output: 1980.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:32<00:48,  1.24it/s, est. speed input: 187.29 toks/s, output: 2129.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:33<00:45,  1.31it/s, est. speed input: 185.62 toks/s, output: 2195.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:35<00:59,  1.02s/it, est. speed input: 177.95 toks/s, output: 2189.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:35<00:50,  1.13it/s, est. speed input: 177.09 toks/s, output: 2264.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:35<00:37,  1.48it/s, est. speed input: 178.18 toks/s, output: 2362.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:36<00:37,  1.47it/s, est. speed input: 176.50 toks/s, output: 2424.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:32,  1.64it/s, est. speed input: 176.07 toks/s, output: 2501.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:37<00:26,  2.00it/s, est. speed input: 176.65 toks/s, output: 2592.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:37<00:21,  2.42it/s, est. speed input: 177.32 toks/s, output: 2684.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:39<00:34,  1.43it/s, est. speed input: 171.00 toks/s, output: 2749.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6198.7 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:41<00:52,  1.06s/it, est. speed input: 163.48 toks/s, output: 2712.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:43<00:58,  1.22s/it, est. speed input: 158.60 toks/s, output: 2715.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:43<00:38,  1.19it/s, est. speed input: 158.95 toks/s, output: 2886.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:44<00:40,  1.12it/s, est. speed input: 156.56 toks/s, output: 2924.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:45<00:30,  1.41it/s, est. speed input: 156.24 toks/s, output: 3081.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:46<00:23,  1.71it/s, est. speed input: 156.50 toks/s, output: 3246.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:49<00:46,  1.16s/it, est. speed input: 147.16 toks/s, output: 3135.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:50<00:37,  1.03it/s, est. speed input: 147.43 toks/s, output: 3223.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5065.2 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:51<00:36,  1.04it/s, est. speed input: 145.88 toks/s, output: 3272.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:51<00:22,  1.60it/s, est. speed input: 147.44 toks/s, output: 3470.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:51<00:17,  1.97it/s, est. speed input: 148.33 toks/s, output: 3572.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:51<00:14,  2.39it/s, est. speed input: 149.12 toks/s, output: 3670.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:53<00:17,  1.88it/s, est. speed input: 147.49 toks/s, output: 3789.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:53<00:14,  2.20it/s, est. speed input: 148.07 toks/s, output: 3885.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:53<00:13,  2.20it/s, est. speed input: 147.98 toks/s, output: 3962.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:54<00:12,  2.31it/s, est. speed input: 148.09 toks/s, output: 4044.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:54<00:11,  2.46it/s, est. speed input: 148.28 toks/s, output: 4129.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:54<00:09,  2.93it/s, est. speed input: 148.93 toks/s, output: 4226.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:55<00:11,  2.19it/s, est. speed input: 148.07 toks/s, output: 4279.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:56<00:13,  1.82it/s, est. speed input: 146.81 toks/s, output: 4398.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:56<00:10,  2.17it/s, est. speed input: 147.43 toks/s, output: 4494.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:58<00:14,  1.55it/s, est. speed input: 145.48 toks/s, output: 4514.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:58<00:11,  1.84it/s, est. speed input: 145.86 toks/s, output: 4606.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:58<00:06,  2.95it/s, est. speed input: 147.59 toks/s, output: 4818.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:58<00:06,  2.80it/s, est. speed input: 147.57 toks/s, output: 4897.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:59<00:04,  3.51it/s, est. speed input: 148.75 toks/s, output: 5092.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:00<00:06,  2.38it/s, est. speed input: 147.56 toks/s, output: 5130.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3699.1 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:00<00:06,  2.02it/s, est. speed input: 146.78 toks/s, output: 5183.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:04<00:18,  1.42s/it, est. speed input: 138.43 toks/s, output: 4972.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:05<00:14,  1.17s/it, est. speed input: 138.34 toks/s, output: 5052.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:06<00:11,  1.08s/it, est. speed input: 137.52 toks/s, output: 5106.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:07<00:08,  1.12it/s, est. speed input: 136.61 toks/s, output: 5242.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:07<00:05,  1.38it/s, est. speed input: 137.12 toks/s, output: 5347.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:08<00:04,  1.58it/s, est. speed input: 137.30 toks/s, output: 5438.99 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  1.58it/s, est. speed input: 143.53 toks/s, output: 6280.86 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  2.35it/s, est. speed input: 143.53 toks/s, output: 6280.86 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:36:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:36:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:36:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:36:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:36:52 [abstract.py:306] It took 0.598929 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0128
+  • Training Accuracy: 0.2562
+  • Mean Generation Length: 2673.8125
+  • Training FLOPS: 47.08 TFLOPS (11.77 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 324.49s
+  • policy_training: 206.55s (63.7%)
+  • generation: 72.66s (22.4%)
+  • teacher_logprob_inference: 40.09s (12.4%)
+  • policy_training/sharding_data: 4.38s (1.3%)
+  • prepare_for_generation: 3.60s (1.1%)
+  • policy_training/submit_training_futures: 1.70s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.66s (0.5%)
+  • training_prep: 1.02s (0.3%)
+  • teacher_logprob_inference_prep: 0.38s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 49: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:01 [abstract.py:324] It took 0.051816 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06756582e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53184013e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66455078e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06756582e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53184013e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66455078e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06756582e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53184013e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66455078e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06756582e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53184013e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.66455078e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06759922e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53393288e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67058563e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22380142e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.37212753e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #49] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43320068e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06759922e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53393288e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67058563e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70398331e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33703108e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #49] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.24162750e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06759922e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53393288e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67058563e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30992966e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.09848738e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #49] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15425873e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06759922e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53393288e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67058563e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26686859e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65876904e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #49] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31300278e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 49: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:04 [abstract.py:324] It took 0.031524 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12732.44it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:04 [loggers.py:236] Engine 000: Avg prompt throughput: 36.7 tokens/s, Avg generation throughput: 63.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:49,  3.22it/s, est. speed input: 205.78 toks/s, output: 48.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:05, 26.80it/s, est. speed input: 1354.43 toks/s, output: 442.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:07, 18.29it/s, est. speed input: 1106.71 toks/s, output: 585.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:08, 16.55it/s, est. speed input: 1039.29 toks/s, output: 707.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:07, 17.81it/s, est. speed input: 1073.58 toks/s, output: 877.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 23.20it/s, est. speed input: 1216.67 toks/s, output: 1249.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 31.91it/s, est. speed input: 1419.00 toks/s, output: 1743.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 31.82it/s, est. speed input: 1466.43 toks/s, output: 2001.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:03, 33.46it/s, est. speed input: 1518.31 toks/s, output: 2233.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:03<00:15,  7.30it/s, est. speed input: 833.63 toks/s, output: 1390.73 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:03<00:15,  7.04it/s, est. speed input: 779.38 toks/s, output: 1508.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11889.9 tokens/s, Running: 108 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:10<01:08,  1.55it/s, est. speed input: 306.23 toks/s, output: 810.78 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:13<01:24,  1.24it/s, est. speed input: 247.05 toks/s, output: 812.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:14<01:14,  1.38it/s, est. speed input: 241.67 toks/s, output: 964.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:14<01:00,  1.68it/s, est. speed input: 244.53 toks/s, output: 1140.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:15<00:54,  1.82it/s, est. speed input: 244.19 toks/s, output: 1220.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:15<00:56,  1.77it/s, est. speed input: 238.01 toks/s, output: 1269.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:16<00:55,  1.76it/s, est. speed input: 233.11 toks/s, output: 1323.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:16<00:45,  2.12it/s, est. speed input: 234.95 toks/s, output: 1413.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:16<00:39,  2.40it/s, est. speed input: 235.20 toks/s, output: 1493.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:16<00:36,  2.57it/s, est. speed input: 234.54 toks/s, output: 1566.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:17<00:38,  2.46it/s, est. speed input: 231.75 toks/s, output: 1624.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:17<00:37,  2.49it/s, est. speed input: 230.08 toks/s, output: 1688.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:18<00:41,  2.20it/s, est. speed input: 225.97 toks/s, output: 1733.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:19<00:49,  1.82it/s, est. speed input: 219.99 toks/s, output: 1762.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:19<00:48,  1.84it/s, est. speed input: 217.32 toks/s, output: 1813.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9624.6 tokens/s, Running: 90 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:20<00:55,  1.61it/s, est. speed input: 211.79 toks/s, output: 1841.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:21<00:41,  2.12it/s, est. speed input: 211.52 toks/s, output: 1985.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:22<00:53,  1.61it/s, est. speed input: 203.96 toks/s, output: 1987.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:22<00:45,  1.86it/s, est. speed input: 204.03 toks/s, output: 2059.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:23<01:06,  1.27it/s, est. speed input: 194.20 toks/s, output: 2032.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:24<01:04,  1.28it/s, est. speed input: 190.48 toks/s, output: 2067.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:24<00:48,  1.68it/s, est. speed input: 191.94 toks/s, output: 2155.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:25<00:36,  2.18it/s, est. speed input: 192.22 toks/s, output: 2300.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:25<00:31,  2.54it/s, est. speed input: 193.23 toks/s, output: 2381.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:25<00:30,  2.59it/s, est. speed input: 192.83 toks/s, output: 2446.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:26<00:28,  2.67it/s, est. speed input: 192.08 toks/s, output: 2574.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:27<00:39,  1.91it/s, est. speed input: 187.19 toks/s, output: 2578.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:27<00:24,  2.98it/s, est. speed input: 190.79 toks/s, output: 2765.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:28<00:24,  2.87it/s, est. speed input: 190.11 toks/s, output: 2887.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:28<00:23,  3.03it/s, est. speed input: 190.47 toks/s, output: 2959.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<00:21,  3.26it/s, est. speed input: 190.96 toks/s, output: 3033.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:29<00:18,  3.60it/s, est. speed input: 192.22 toks/s, output: 3180.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:30<00:22,  2.93it/s, est. speed input: 190.64 toks/s, output: 3218.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7651.4 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:30<00:18,  3.50it/s, est. speed input: 191.92 toks/s, output: 3304.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:30<00:15,  4.10it/s, est. speed input: 193.21 toks/s, output: 3389.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:30<00:17,  3.52it/s, est. speed input: 192.75 toks/s, output: 3444.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:31<00:19,  3.21it/s, est. speed input: 192.23 toks/s, output: 3499.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:31<00:09,  5.94it/s, est. speed input: 197.25 toks/s, output: 3778.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:31<00:07,  7.19it/s, est. speed input: 200.07 toks/s, output: 3955.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:33<00:19,  2.81it/s, est. speed input: 194.01 toks/s, output: 3959.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:33<00:21,  2.49it/s, est. speed input: 192.33 toks/s, output: 3988.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:34<00:24,  2.17it/s, est. speed input: 190.19 toks/s, output: 4008.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:34<00:19,  2.61it/s, est. speed input: 191.23 toks/s, output: 4093.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:35<00:30,  1.70it/s, est. speed input: 186.44 toks/s, output: 4056.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:26,  1.82it/s, est. speed input: 184.76 toks/s, output: 4145.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:37<00:20,  2.28it/s, est. speed input: 185.61 toks/s, output: 4290.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:37<00:20,  2.27it/s, est. speed input: 185.00 toks/s, output: 4342.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:37<00:18,  2.45it/s, est. speed input: 185.11 toks/s, output: 4409.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:38<00:13,  3.07it/s, est. speed input: 186.25 toks/s, output: 4566.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:38<00:15,  2.76it/s, est. speed input: 185.42 toks/s, output: 4609.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:39<00:19,  2.08it/s, est. speed input: 182.89 toks/s, output: 4612.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:17,  2.23it/s, est. speed input: 182.71 toks/s, output: 4673.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5954.9 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:41<00:31,  1.24it/s, est. speed input: 176.28 toks/s, output: 4576.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:25,  1.51it/s, est. speed input: 176.63 toks/s, output: 4649.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:30,  1.20it/s, est. speed input: 172.86 toks/s, output: 4618.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:22,  1.57it/s, est. speed input: 173.60 toks/s, output: 4706.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:19,  1.81it/s, est. speed input: 173.63 toks/s, output: 4772.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:18,  1.88it/s, est. speed input: 173.16 toks/s, output: 4825.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:16,  1.90it/s, est. speed input: 172.05 toks/s, output: 4923.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:20,  1.48it/s, est. speed input: 169.10 toks/s, output: 4909.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:19,  1.58it/s, est. speed input: 168.52 toks/s, output: 4962.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:47<00:16,  1.78it/s, est. speed input: 168.46 toks/s, output: 5030.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:13,  2.12it/s, est. speed input: 168.93 toks/s, output: 5113.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:10,  2.54it/s, est. speed input: 169.45 toks/s, output: 5198.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:10,  2.52it/s, est. speed input: 169.25 toks/s, output: 5262.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:49<00:18,  1.36it/s, est. speed input: 165.14 toks/s, output: 5206.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:49<00:13,  1.73it/s, est. speed input: 165.66 toks/s, output: 5294.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:41:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4435.8 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:07,  2.82it/s, est. speed input: 167.49 toks/s, output: 5493.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:07,  2.68it/s, est. speed input: 167.14 toks/s, output: 5624.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:51<00:06,  2.92it/s, est. speed input: 167.61 toks/s, output: 5709.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:13,  1.34it/s, est. speed input: 162.34 toks/s, output: 5603.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:55<00:17,  1.03s/it, est. speed input: 157.90 toks/s, output: 5526.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:55<00:13,  1.15it/s, est. speed input: 157.81 toks/s, output: 5599.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:55<00:11,  1.30it/s, est. speed input: 157.42 toks/s, output: 5663.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:58<00:13,  1.06s/it, est. speed input: 151.77 toks/s, output: 5614.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:42:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2774.7 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:00<00:14,  1.23s/it, est. speed input: 148.33 toks/s, output: 5569.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:03<00:17,  1.56s/it, est. speed input: 143.45 toks/s, output: 5469.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:03<00:11,  1.17s/it, est. speed input: 144.15 toks/s, output: 5581.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:05<00:12,  1.34s/it, est. speed input: 141.05 toks/s, output: 5549.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:06<00:09,  1.24s/it, est. speed input: 139.82 toks/s, output: 5591.07 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  1.24s/it, est. speed input: 147.21 toks/s, output: 6583.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  2.42it/s, est. speed input: 147.21 toks/s, output: 6583.62 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:42:14 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:42:14 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:42:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:42:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:42:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:42:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:42:14 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:42:14 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:42:14 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 02:42:14 [abstract.py:306] It took 0.609515 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:42:14 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0143
+  • Training Accuracy: 0.2500
+  • Mean Generation Length: 2716.8438
+  • Training FLOPS: 47.57 TFLOPS (11.89 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 318.08s
+  • policy_training: 204.03s (64.1%)
+  • generation: 70.51s (22.2%)
+  • teacher_logprob_inference: 38.21s (12.0%)
+  • policy_training/sharding_data: 4.33s (1.4%)
+  • prepare_for_generation: 3.57s (1.1%)
+  • policy_training/submit_training_futures: 1.74s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.5%)
+  • training_prep: 1.13s (0.4%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 50: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:46:19 [abstract.py:324] It took 0.051064 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06814727e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53627701e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67196198e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06814727e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53627701e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67196198e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06814727e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53627701e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67196198e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06814727e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53627701e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67196198e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06817812e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55467072e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68347397e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22931118e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.23421383e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #50] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43576691e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06817812e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55467072e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68347397e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70175171e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33181047e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #50] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23587151e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06817812e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55467072e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68347397e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30084534e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.23877287e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #50] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17871361e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06817812e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55467072e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68347397e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26539612e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66389427e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #50] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31520065e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 50: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:46:22 [abstract.py:324] It took 0.033965 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13106.18it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:46:22 [loggers.py:236] Engine 000: Avg prompt throughput: 37.6 tokens/s, Avg generation throughput: 43.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:34,  4.59it/s, est. speed input: 275.61 toks/s, output: 32.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:06, 22.12it/s, est. speed input: 1108.26 toks/s, output: 219.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 45.81it/s, est. speed input: 2147.22 toks/s, output: 662.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 40.16it/s, est. speed input: 2126.01 toks/s, output: 823.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 41.75it/s, est. speed input: 2224.54 toks/s, output: 1074.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:02, 47.44it/s, est. speed input: 2422.51 toks/s, output: 1462.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:04, 24.55it/s, est. speed input: 1805.78 toks/s, output: 1370.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:02<00:08, 13.55it/s, est. speed input: 1286.31 toks/s, output: 1206.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:05<00:31,  3.58it/s, est. speed input: 530.64 toks/s, output: 724.98 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 02:46:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11943.0 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:10<01:12,  1.53it/s, est. speed input: 279.91 toks/s, output: 539.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:13<01:27,  1.25it/s, est. speed input: 229.39 toks/s, output: 623.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:14<01:25,  1.26it/s, est. speed input: 221.67 toks/s, output: 690.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:14<01:17,  1.38it/s, est. speed input: 220.95 toks/s, output: 773.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:15<01:18,  1.35it/s, est. speed input: 213.11 toks/s, output: 831.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:15<01:05,  1.59it/s, est. speed input: 214.97 toks/s, output: 921.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:15<00:48,  2.12it/s, est. speed input: 217.50 toks/s, output: 1093.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:17<01:02,  1.64it/s, est. speed input: 206.42 toks/s, output: 1118.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:17<00:51,  1.96it/s, est. speed input: 207.93 toks/s, output: 1205.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:17<00:45,  2.20it/s, est. speed input: 207.95 toks/s, output: 1283.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:18<00:49,  2.00it/s, est. speed input: 204.07 toks/s, output: 1337.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:18<00:44,  2.23it/s, est. speed input: 203.91 toks/s, output: 1412.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:46:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9857.5 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:20<01:21,  1.19it/s, est. speed input: 188.29 toks/s, output: 1380.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:21<01:16,  1.26it/s, est. speed input: 184.99 toks/s, output: 1433.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:21<01:14,  1.28it/s, est. speed input: 181.35 toks/s, output: 1481.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:21<00:56,  1.67it/s, est. speed input: 182.89 toks/s, output: 1567.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:22<01:08,  1.35it/s, est. speed input: 177.01 toks/s, output: 1590.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:23<01:09,  1.33it/s, est. speed input: 173.73 toks/s, output: 1633.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:23<00:53,  1.70it/s, est. speed input: 174.82 toks/s, output: 1716.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:24<00:44,  2.01it/s, est. speed input: 175.17 toks/s, output: 1792.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:24<00:35,  2.52it/s, est. speed input: 176.48 toks/s, output: 1876.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:24<00:39,  2.24it/s, est. speed input: 174.82 toks/s, output: 1930.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:25<00:41,  2.12it/s, est. speed input: 173.62 toks/s, output: 1986.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:26<00:51,  1.68it/s, est. speed input: 170.13 toks/s, output: 2016.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:27<00:39,  2.11it/s, est. speed input: 170.31 toks/s, output: 2156.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:27<00:33,  2.49it/s, est. speed input: 171.50 toks/s, output: 2238.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:28<00:43,  1.88it/s, est. speed input: 168.20 toks/s, output: 2262.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:28<00:39,  2.06it/s, est. speed input: 168.26 toks/s, output: 2329.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:28<00:24,  3.23it/s, est. speed input: 171.69 toks/s, output: 2574.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:29<00:27,  2.75it/s, est. speed input: 170.37 toks/s, output: 2619.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:46:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8076.8 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:31<00:46,  1.63it/s, est. speed input: 164.18 toks/s, output: 2590.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:31<00:45,  1.64it/s, est. speed input: 163.01 toks/s, output: 2637.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:32<00:41,  1.77it/s, est. speed input: 162.60 toks/s, output: 2697.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:33<01:00,  1.21it/s, est. speed input: 156.99 toks/s, output: 2668.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:35<01:25,  1.19s/it, est. speed input: 149.31 toks/s, output: 2604.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:39<02:23,  2.02s/it, est. speed input: 135.35 toks/s, output: 2430.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:39<01:43,  1.48s/it, est. speed input: 136.40 toks/s, output: 2516.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:40<00:57,  1.18it/s, est. speed input: 138.89 toks/s, output: 2695.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6701.6 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:40<00:50,  1.32it/s, est. speed input: 138.68 toks/s, output: 2756.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:40<00:32,  2.00it/s, est. speed input: 140.80 toks/s, output: 2928.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:41<00:31,  2.02it/s, est. speed input: 140.57 toks/s, output: 2988.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:41<00:26,  2.41it/s, est. speed input: 141.48 toks/s, output: 3070.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:30,  2.04it/s, est. speed input: 140.53 toks/s, output: 3114.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:43<00:35,  1.71it/s, est. speed input: 139.28 toks/s, output: 3148.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:43<00:23,  2.56it/s, est. speed input: 141.15 toks/s, output: 3317.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:44<00:29,  1.99it/s, est. speed input: 139.78 toks/s, output: 3348.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:44<00:25,  2.28it/s, est. speed input: 140.35 toks/s, output: 3424.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:44<00:20,  2.78it/s, est. speed input: 141.24 toks/s, output: 3508.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:45<00:27,  1.98it/s, est. speed input: 139.82 toks/s, output: 3535.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:45<00:21,  2.48it/s, est. speed input: 140.71 toks/s, output: 3619.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:46<00:25,  2.10it/s, est. speed input: 140.03 toks/s, output: 3664.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:47<00:42,  1.21it/s, est. speed input: 136.36 toks/s, output: 3631.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:48<00:34,  1.47it/s, est. speed input: 136.66 toks/s, output: 3702.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:49<00:39,  1.26it/s, est. speed input: 134.94 toks/s, output: 3718.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:49<00:29,  1.68it/s, est. speed input: 135.78 toks/s, output: 3804.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:49<00:26,  1.82it/s, est. speed input: 135.78 toks/s, output: 3867.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5815.1 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:50<00:19,  2.38it/s, est. speed input: 136.65 toks/s, output: 4016.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:51<00:23,  1.92it/s, est. speed input: 135.62 toks/s, output: 4048.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:51<00:23,  1.90it/s, est. speed input: 135.39 toks/s, output: 4103.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:52<00:21,  1.99it/s, est. speed input: 135.37 toks/s, output: 4165.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:52<00:18,  2.22it/s, est. speed input: 135.67 toks/s, output: 4236.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:53<00:22,  1.81it/s, est. speed input: 134.51 toks/s, output: 4321.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:55<00:26,  1.49it/s, est. speed input: 133.03 toks/s, output: 4337.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:55<00:16,  2.20it/s, est. speed input: 134.57 toks/s, output: 4511.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:56<00:26,  1.38it/s, est. speed input: 131.68 toks/s, output: 4477.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:58<00:30,  1.17it/s, est. speed input: 129.86 toks/s, output: 4479.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:58<00:25,  1.31it/s, est. speed input: 129.86 toks/s, output: 4541.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:59<00:23,  1.40it/s, est. speed input: 129.63 toks/s, output: 4595.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:59<00:19,  1.68it/s, est. speed input: 129.98 toks/s, output: 4672.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:59<00:15,  1.97it/s, est. speed input: 130.36 toks/s, output: 4750.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4615.9 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:00<00:17,  1.71it/s, est. speed input: 129.67 toks/s, output: 4789.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:01<00:12,  2.29it/s, est. speed input: 130.55 toks/s, output: 4949.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:02<00:13,  1.92it/s, est. speed input: 129.79 toks/s, output: 5045.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:02<00:11,  2.13it/s, est. speed input: 130.16 toks/s, output: 5123.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:03<00:11,  2.18it/s, est. speed input: 130.25 toks/s, output: 5190.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:04<00:11,  1.93it/s, est. speed input: 129.73 toks/s, output: 5295.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:06<00:16,  1.29it/s, est. speed input: 127.36 toks/s, output: 5264.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:06<00:12,  1.59it/s, est. speed input: 127.97 toks/s, output: 5355.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:07<00:14,  1.27it/s, est. speed input: 126.50 toks/s, output: 5360.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:08<00:13,  1.35it/s, est. speed input: 126.27 toks/s, output: 5418.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:08<00:09,  1.72it/s, est. speed input: 126.70 toks/s, output: 5568.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:08<00:07,  2.14it/s, est. speed input: 127.42 toks/s, output: 5666.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3225.2 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:10<00:08,  1.46it/s, est. speed input: 125.51 toks/s, output: 5716.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:12<00:10,  1.20it/s, est. speed input: 124.00 toks/s, output: 5717.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:13<00:10,  1.04it/s, est. speed input: 122.54 toks/s, output: 5720.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:13<00:07,  1.34it/s, est. speed input: 123.17 toks/s, output: 5820.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:14<00:06,  1.32it/s, est. speed input: 122.67 toks/s, output: 5868.10 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  1.32it/s, est. speed input: 130.03 toks/s, output: 6856.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  2.15it/s, est. speed input: 130.03 toks/s, output: 6856.85 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:41 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:41 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:47:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:47:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:47:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:47:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:47:42 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 02:47:42 [abstract.py:306] It took 0.552959 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:47:42 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:47:42 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:47:42 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.1562
+  • Mean Generation Length: 3195.5249
+  • Training FLOPS: 49.23 TFLOPS (12.31 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 360.73s
+  • policy_training: 234.22s (64.9%)
+  • generation: 79.81s (22.1%)
+  • teacher_logprob_inference: 41.61s (11.5%)
+  • policy_training/sharding_data: 4.39s (1.2%)
+  • prepare_for_generation: 3.44s (1.0%)
+  • policy_training/submit_training_futures: 2.00s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.4%)
+  • training_prep: 0.97s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 11/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 51: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:52:20 [abstract.py:324] It took 0.057507 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06852676e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53657265e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67648621e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06852676e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53657265e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67648621e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06852676e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53657265e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67648621e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06852676e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53657265e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.67648621e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06861211e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54189148e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67957611e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.21795006e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.11107159e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #51] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43919540e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06861211e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54189148e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67957611e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71379700e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33233490e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #51] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.24664650e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06861211e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54189148e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67957611e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30360718e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-2.16954470e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #51] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17056885e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06861211e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54189148e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.67957611e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27048111e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66610870e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #51] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31748108e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 51: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:52:23 [abstract.py:324] It took 0.033280 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13471.08it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:52:24 [loggers.py:236] Engine 000: Avg prompt throughput: 33.1 tokens/s, Avg generation throughput: 29.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:27,  1.82it/s, est. speed input: 107.65 toks/s, output: 20.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:08, 17.46it/s, est. speed input: 844.52 toks/s, output: 230.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:07, 20.59it/s, est. speed input: 992.94 toks/s, output: 345.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:06, 20.57it/s, est. speed input: 1034.48 toks/s, output: 471.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:06, 20.11it/s, est. speed input: 1051.40 toks/s, output: 584.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:05, 23.79it/s, est. speed input: 1158.64 toks/s, output: 811.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:04, 27.26it/s, est. speed input: 1258.97 toks/s, output: 1052.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:02<00:08, 15.48it/s, est. speed input: 1045.92 toks/s, output: 1031.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:02<00:07, 15.81it/s, est. speed input: 1044.28 toks/s, output: 1181.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:02<00:06, 18.16it/s, est. speed input: 1081.60 toks/s, output: 1469.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:04<00:24,  4.60it/s, est. speed input: 601.58 toks/s, output: 958.64 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:08<00:57,  1.96it/s, est. speed input: 352.57 toks/s, output: 708.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:08<00:48,  2.28it/s, est. speed input: 352.81 toks/s, output: 872.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:09<00:52,  2.09it/s, est. speed input: 331.12 toks/s, output: 900.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:09<00:50,  2.12it/s, est. speed input: 322.79 toks/s, output: 958.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:09<00:43,  2.45it/s, est. speed input: 324.76 toks/s, output: 1044.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:10<00:41,  2.58it/s, est. speed input: 320.85 toks/s, output: 1110.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:52:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11934.3 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:10<00:35,  2.89it/s, est. speed input: 315.68 toks/s, output: 1244.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:10<00:31,  3.31it/s, est. speed input: 317.39 toks/s, output: 1327.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:11<00:37,  2.75it/s, est. speed input: 306.95 toks/s, output: 1359.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:14<01:46,  1.06s/it, est. speed input: 244.64 toks/s, output: 1162.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:15<01:31,  1.09it/s, est. speed input: 240.17 toks/s, output: 1221.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:15<01:14,  1.33it/s, est. speed input: 239.24 toks/s, output: 1293.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:17<01:44,  1.07s/it, est. speed input: 217.01 toks/s, output: 1251.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:17<01:22,  1.18it/s, est. speed input: 216.72 toks/s, output: 1327.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:52:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9617.0 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:20<02:29,  1.56s/it, est. speed input: 185.44 toks/s, output: 1215.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:22<02:21,  1.49s/it, est. speed input: 177.06 toks/s, output: 1240.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:23<02:27,  1.56s/it, est. speed input: 166.68 toks/s, output: 1245.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:24<01:59,  1.29s/it, est. speed input: 164.81 toks/s, output: 1309.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:25<01:41,  1.10s/it, est. speed input: 162.97 toks/s, output: 1371.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:25<01:01,  1.47it/s, est. speed input: 165.31 toks/s, output: 1541.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:28<01:48,  1.22s/it, est. speed input: 150.75 toks/s, output: 1481.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:52:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8273.1 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:30<02:14,  1.53s/it, est. speed input: 140.97 toks/s, output: 1460.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:31<01:45,  1.22s/it, est. speed input: 141.26 toks/s, output: 1537.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:31<01:19,  1.08it/s, est. speed input: 142.33 toks/s, output: 1623.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:32<01:09,  1.23it/s, est. speed input: 141.83 toks/s, output: 1690.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:32<00:58,  1.43it/s, est. speed input: 141.92 toks/s, output: 1763.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:33<01:11,  1.15it/s, est. speed input: 138.30 toks/s, output: 1790.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:34<01:02,  1.31it/s, est. speed input: 137.91 toks/s, output: 1857.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:35<01:02,  1.29it/s, est. speed input: 136.54 toks/s, output: 1909.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:36<01:09,  1.15it/s, est. speed input: 134.04 toks/s, output: 1944.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:37<01:24,  1.07s/it, est. speed input: 130.19 toks/s, output: 1959.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:38<01:25,  1.10s/it, est. speed input: 127.83 toks/s, output: 1992.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:39<01:21,  1.06s/it, est. speed input: 126.32 toks/s, output: 2037.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7087.6 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.3%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:41<01:31,  1.20s/it, est. speed input: 123.11 toks/s, output: 2054.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:43<01:48,  1.44s/it, est. speed input: 118.84 toks/s, output: 2051.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:43<01:05,  1.11it/s, est. speed input: 120.11 toks/s, output: 2209.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:45<01:11,  1.00it/s, est. speed input: 117.97 toks/s, output: 2238.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:45<01:04,  1.10it/s, est. speed input: 117.56 toks/s, output: 2298.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:46<01:08,  1.02it/s, est. speed input: 115.92 toks/s, output: 2333.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:47<00:51,  1.34it/s, est. speed input: 116.77 toks/s, output: 2416.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:49<01:20,  1.18s/it, est. speed input: 112.69 toks/s, output: 2397.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:50<01:08,  1.02s/it, est. speed input: 112.44 toks/s, output: 2458.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6118.9 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 14.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:51<01:07,  1.02s/it, est. speed input: 111.41 toks/s, output: 2500.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:51<00:58,  1.10it/s, est. speed input: 111.16 toks/s, output: 2560.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:52<00:59,  1.08it/s, est. speed input: 110.22 toks/s, output: 2604.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:52<00:43,  1.46it/s, est. speed input: 111.08 toks/s, output: 2689.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:53<00:33,  1.82it/s, est. speed input: 111.77 toks/s, output: 2831.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:54<00:38,  1.57it/s, est. speed input: 111.04 toks/s, output: 2875.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:55<00:37,  1.57it/s, est. speed input: 110.85 toks/s, output: 2934.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:56<00:42,  1.38it/s, est. speed input: 109.97 toks/s, output: 2974.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:56<00:41,  1.36it/s, est. speed input: 109.58 toks/s, output: 3026.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:58<00:53,  1.05it/s, est. speed input: 107.82 toks/s, output: 3040.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:58<00:40,  1.36it/s, est. speed input: 108.50 toks/s, output: 3122.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:58<00:35,  1.50it/s, est. speed input: 108.58 toks/s, output: 3187.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:59<00:34,  1.56it/s, est. speed input: 108.56 toks/s, output: 3247.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [01:00<00:35,  1.46it/s, est. speed input: 108.16 toks/s, output: 3297.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5769.1 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.3%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [01:00<00:26,  1.90it/s, est. speed input: 108.87 toks/s, output: 3381.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [01:01<00:39,  1.28it/s, est. speed input: 107.41 toks/s, output: 3397.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [01:02<00:29,  1.66it/s, est. speed input: 108.09 toks/s, output: 3480.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [01:03<00:26,  1.76it/s, est. speed input: 108.14 toks/s, output: 3604.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [01:03<00:22,  2.07it/s, est. speed input: 108.71 toks/s, output: 3683.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [01:04<00:22,  1.98it/s, est. speed input: 108.72 toks/s, output: 3804.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [01:04<00:19,  2.25it/s, est. speed input: 109.26 toks/s, output: 3882.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [01:04<00:15,  2.75it/s, est. speed input: 109.97 toks/s, output: 3968.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:05<00:16,  2.51it/s, est. speed input: 110.07 toks/s, output: 4030.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:05<00:13,  2.97it/s, est. speed input: 110.72 toks/s, output: 4112.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:05<00:13,  2.91it/s, est. speed input: 111.03 toks/s, output: 4182.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:06<00:19,  2.00it/s, est. speed input: 110.46 toks/s, output: 4218.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:07<00:14,  2.48it/s, est. speed input: 111.28 toks/s, output: 4368.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:08<00:17,  1.96it/s, est. speed input: 110.77 toks/s, output: 4407.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:08<00:13,  2.46it/s, est. speed input: 111.50 toks/s, output: 4493.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:08<00:11,  2.75it/s, est. speed input: 111.96 toks/s, output: 4570.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:09<00:13,  2.41it/s, est. speed input: 111.96 toks/s, output: 4628.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:09<00:13,  2.35it/s, est. speed input: 112.08 toks/s, output: 4691.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:10<00:13,  2.15it/s, est. speed input: 112.02 toks/s, output: 4748.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4355.0 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:10<00:12,  2.17it/s, est. speed input: 112.24 toks/s, output: 4874.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:12<00:16,  1.64it/s, est. speed input: 111.43 toks/s, output: 4897.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:12<00:12,  2.02it/s, est. speed input: 112.02 toks/s, output: 4981.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:12<00:08,  2.75it/s, est. speed input: 113.15 toks/s, output: 5146.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:13<00:07,  3.04it/s, est. speed input: 113.97 toks/s, output: 5298.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:13<00:07,  2.94it/s, est. speed input: 114.19 toks/s, output: 5367.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:14<00:08,  2.29it/s, est. speed input: 113.84 toks/s, output: 5409.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:15<00:10,  1.81it/s, est. speed input: 113.26 toks/s, output: 5441.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:15<00:10,  1.79it/s, est. speed input: 113.19 toks/s, output: 5497.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:16<00:08,  1.96it/s, est. speed input: 113.45 toks/s, output: 5568.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:16<00:06,  2.44it/s, est. speed input: 114.01 toks/s, output: 5654.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:17<00:08,  1.76it/s, est. speed input: 113.34 toks/s, output: 5683.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:17<00:07,  1.96it/s, est. speed input: 113.57 toks/s, output: 5755.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:19<00:11,  1.15it/s, est. speed input: 111.87 toks/s, output: 5730.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:20<00:10,  1.09it/s, est. speed input: 111.18 toks/s, output: 5759.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:20<00:07,  1.49it/s, est. speed input: 111.83 toks/s, output: 5854.16 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:20<00:00,  1.49it/s, est. speed input: 120.04 toks/s, output: 6974.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:20<00:00,  1.99it/s, est. speed input: 120.04 toks/s, output: 6974.40 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:49 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:49 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 02:53:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:53:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:53:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:53:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 02:53:49 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 02:53:49 [abstract.py:306] It took 0.545740 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 02:53:49 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 02:53:49 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 02:53:49 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0110
+  • Training Accuracy: 0.1313
+  • Mean Generation Length: 3506.3062
+  • Training FLOPS: 50.65 TFLOPS (12.66 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 390.18s
+  • policy_training: 251.89s (64.6%)
+  • generation: 86.13s (22.1%)
+  • teacher_logprob_inference: 47.11s (12.1%)
+  • policy_training/sharding_data: 4.36s (1.1%)
+  • prepare_for_generation: 3.54s (0.9%)
+  • policy_training/submit_training_futures: 2.01s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.4%)
+  • training_prep: 0.87s (0.2%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 52: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 02:58:50 [abstract.py:324] It took 0.050736 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06868867e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54965096e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68006821e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06868867e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54965096e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68006821e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06868867e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54965096e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68006821e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06868867e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54965096e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.68006821e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06871953e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53525162e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68937454e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22909145e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.14943457e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #52] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43964935e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06871953e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53525162e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68937454e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71032104e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34269848e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #52] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.25548859e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06871953e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53525162e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68937454e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.32766876e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.05659771e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #52] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17866325e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06871953e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53525162e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.68937454e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27663040e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66126480e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #52] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31960373e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 52: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 02:58:53 [abstract.py:324] It took 0.035874 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13240.90it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 02:58:54 [loggers.py:236] Engine 000: Avg prompt throughput: 29.9 tokens/s, Avg generation throughput: 91.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:52,  1.42it/s, est. speed input: 82.20 toks/s, output: 17.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:11, 12.83it/s, est. speed input: 583.85 toks/s, output: 158.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:01<00:09, 15.14it/s, est. speed input: 711.50 toks/s, output: 254.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:06, 22.51it/s, est. speed input: 956.57 toks/s, output: 468.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:05, 23.05it/s, est. speed input: 1020.93 toks/s, output: 604.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:05, 25.30it/s, est. speed input: 1104.48 toks/s, output: 768.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:04, 28.01it/s, est. speed input: 1209.01 toks/s, output: 1017.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 31.55it/s, est. speed input: 1302.35 toks/s, output: 1256.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:04, 26.04it/s, est. speed input: 1274.55 toks/s, output: 1364.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:02<00:04, 24.56it/s, est. speed input: 1275.76 toks/s, output: 1513.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:02<00:06, 18.36it/s, est. speed input: 1191.36 toks/s, output: 1534.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:02<00:05, 19.53it/s, est. speed input: 1202.83 toks/s, output: 1775.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:10<01:04,  1.60it/s, est. speed input: 331.11 toks/s, output: 668.78 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 02:59:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11436.7 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:11<01:02,  1.64it/s, est. speed input: 309.69 toks/s, output: 795.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:11<00:52,  1.92it/s, est. speed input: 310.21 toks/s, output: 966.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:13<01:13,  1.34it/s, est. speed input: 263.96 toks/s, output: 906.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:15<01:32,  1.06it/s, est. speed input: 234.31 toks/s, output: 889.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:16<01:27,  1.11it/s, est. speed input: 228.14 toks/s, output: 950.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:18<01:44,  1.09s/it, est. speed input: 208.90 toks/s, output: 954.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:18<01:28,  1.07it/s, est. speed input: 207.35 toks/s, output: 1030.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:19<01:12,  1.29it/s, est. speed input: 207.33 toks/s, output: 1112.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:20<01:21,  1.15it/s, est. speed input: 198.59 toks/s, output: 1147.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:59:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9482.1 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:21<01:18,  1.17it/s, est. speed input: 193.89 toks/s, output: 1201.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:21<01:14,  1.23it/s, est. speed input: 190.21 toks/s, output: 1258.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:22<01:04,  1.39it/s, est. speed input: 188.83 toks/s, output: 1328.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:22<00:54,  1.62it/s, est. speed input: 188.46 toks/s, output: 1403.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:23<00:48,  1.81it/s, est. speed input: 187.81 toks/s, output: 1475.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:24<01:10,  1.23it/s, est. speed input: 179.21 toks/s, output: 1485.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:25<01:24,  1.02it/s, est. speed input: 171.94 toks/s, output: 1503.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:26<01:06,  1.27it/s, est. speed input: 172.14 toks/s, output: 1581.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:26<00:49,  1.71it/s, est. speed input: 173.65 toks/s, output: 1670.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:26<00:37,  2.21it/s, est. speed input: 175.02 toks/s, output: 1758.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:28<01:26,  1.05s/it, est. speed input: 162.18 toks/s, output: 1703.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:29<01:08,  1.18it/s, est. speed input: 162.24 toks/s, output: 1779.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:29<01:00,  1.31it/s, est. speed input: 161.11 toks/s, output: 1841.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:30<00:58,  1.34it/s, est. speed input: 159.36 toks/s, output: 1894.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:59:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7953.9 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:32<01:35,  1.23s/it, est. speed input: 149.74 toks/s, output: 1853.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:33<01:17,  1.01s/it, est. speed input: 149.37 toks/s, output: 1922.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:35<01:30,  1.20s/it, est. speed input: 144.08 toks/s, output: 1927.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:35<01:12,  1.04it/s, est. speed input: 144.04 toks/s, output: 1998.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:36<01:06,  1.11it/s, est. speed input: 142.71 toks/s, output: 2052.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:36<00:52,  1.39it/s, est. speed input: 143.20 toks/s, output: 2130.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:38<01:08,  1.04it/s, est. speed input: 139.08 toks/s, output: 2140.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:39<01:21,  1.15s/it, est. speed input: 135.00 toks/s, output: 2147.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:40<01:10,  1.01s/it, est. speed input: 134.25 toks/s, output: 2206.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:59:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6825.5 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.4%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:42<01:33,  1.35s/it, est. speed input: 128.87 toks/s, output: 2189.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:42<01:07,  1.00it/s, est. speed input: 129.69 toks/s, output: 2274.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:42<00:39,  1.66it/s, est. speed input: 131.65 toks/s, output: 2447.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:43<00:42,  1.54it/s, est. speed input: 130.67 toks/s, output: 2497.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:44<00:31,  2.02it/s, est. speed input: 131.71 toks/s, output: 2653.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:44<00:28,  2.14it/s, est. speed input: 131.93 toks/s, output: 2725.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:44<00:23,  2.59it/s, est. speed input: 132.81 toks/s, output: 2811.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:45<00:28,  2.10it/s, est. speed input: 131.98 toks/s, output: 2859.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:46<00:29,  1.97it/s, est. speed input: 131.54 toks/s, output: 2917.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:46<00:27,  2.14it/s, est. speed input: 131.80 toks/s, output: 2988.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:47<00:27,  2.06it/s, est. speed input: 131.58 toks/s, output: 3049.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:47<00:18,  2.92it/s, est. speed input: 133.14 toks/s, output: 3215.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:47<00:17,  3.11it/s, est. speed input: 133.73 toks/s, output: 3292.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:47<00:16,  3.15it/s, est. speed input: 134.15 toks/s, output: 3365.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:48<00:14,  3.48it/s, est. speed input: 134.83 toks/s, output: 3446.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:48<00:18,  2.80it/s, est. speed input: 134.56 toks/s, output: 3502.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:49<00:18,  2.66it/s, est. speed input: 134.62 toks/s, output: 3567.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:49<00:15,  3.10it/s, est. speed input: 135.28 toks/s, output: 3648.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:49<00:12,  3.77it/s, est. speed input: 136.17 toks/s, output: 3733.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:49<00:14,  3.31it/s, est. speed input: 136.33 toks/s, output: 3799.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:50<00:18,  2.52it/s, est. speed input: 135.79 toks/s, output: 3847.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:59:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5852.0 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:51<00:25,  1.79it/s, est. speed input: 134.44 toks/s, output: 3872.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:51<00:21,  2.01it/s, est. speed input: 134.70 toks/s, output: 3942.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:52<00:28,  1.51it/s, est. speed input: 133.16 toks/s, output: 3959.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:53<00:33,  1.25it/s, est. speed input: 131.53 toks/s, output: 3974.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:54<00:29,  1.38it/s, est. speed input: 131.27 toks/s, output: 4030.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:55<00:27,  1.44it/s, est. speed input: 130.85 toks/s, output: 4082.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:55<00:25,  1.52it/s, est. speed input: 130.59 toks/s, output: 4137.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:56<00:32,  1.17it/s, est. speed input: 128.57 toks/s, output: 4138.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:57<00:24,  1.48it/s, est. speed input: 129.03 toks/s, output: 4218.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:57<00:21,  1.65it/s, est. speed input: 129.03 toks/s, output: 4283.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:57<00:16,  2.12it/s, est. speed input: 129.71 toks/s, output: 4369.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:58<00:15,  2.21it/s, est. speed input: 129.82 toks/s, output: 4437.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:58<00:15,  2.15it/s, est. speed input: 129.75 toks/s, output: 4498.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:58<00:12,  2.54it/s, est. speed input: 130.26 toks/s, output: 4579.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:59<00:09,  3.16it/s, est. speed input: 130.98 toks/s, output: 4667.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:00<00:17,  1.71it/s, est. speed input: 129.35 toks/s, output: 4673.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 02:59:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4650.6 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:00<00:16,  1.73it/s, est. speed input: 129.17 toks/s, output: 4729.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:01<00:18,  1.51it/s, est. speed input: 128.34 toks/s, output: 4763.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:01<00:14,  1.89it/s, est. speed input: 128.84 toks/s, output: 4847.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:02<00:13,  1.88it/s, est. speed input: 128.69 toks/s, output: 4905.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:03<00:19,  1.28it/s, est. speed input: 126.88 toks/s, output: 4901.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:04<00:19,  1.26it/s, est. speed input: 126.17 toks/s, output: 4940.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:05<00:16,  1.37it/s, est. speed input: 125.98 toks/s, output: 4997.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:07<00:29,  1.33s/it, est. speed input: 121.77 toks/s, output: 4898.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:08<00:21,  1.02s/it, est. speed input: 122.08 toks/s, output: 4979.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:08<00:15,  1.32it/s, est. speed input: 122.73 toks/s, output: 5072.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:08<00:10,  1.76it/s, est. speed input: 123.37 toks/s, output: 5166.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:09<00:13,  1.37it/s, est. speed input: 122.27 toks/s, output: 5187.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:09<00:10,  1.69it/s, est. speed input: 122.65 toks/s, output: 5271.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:00:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3154.2 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:10<00:10,  1.58it/s, est. speed input: 122.23 toks/s, output: 5321.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:11<00:12,  1.21it/s, est. speed input: 120.85 toks/s, output: 5332.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:13<00:15,  1.08s/it, est. speed input: 118.91 toks/s, output: 5317.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:14<00:14,  1.08s/it, est. speed input: 118.01 toks/s, output: 5349.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:15<00:11,  1.03it/s, est. speed input: 117.64 toks/s, output: 5406.76 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  1.03it/s, est. speed input: 127.14 toks/s, output: 6710.82 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  2.12it/s, est. speed input: 127.14 toks/s, output: 6710.82 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:00:13 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:00:13 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:00:13 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:00:13 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:00:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:00:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:00:14 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:00:14 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 03:00:14 [abstract.py:306] It took 0.689368 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:00:14 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:00:14 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0109
+  • Training Accuracy: 0.2687
+  • Mean Generation Length: 3161.6938
+  • Training FLOPS: 48.90 TFLOPS (12.23 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 362.64s
+  • policy_training: 234.22s (64.6%)
+  • generation: 80.67s (22.2%)
+  • teacher_logprob_inference: 42.53s (11.7%)
+  • policy_training/sharding_data: 4.42s (1.2%)
+  • prepare_for_generation: 3.50s (1.0%)
+  • policy_training/submit_training_futures: 1.88s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.4%)
+  • training_prep: 0.85s (0.2%)
+  • teacher_logprob_inference_prep: 0.69s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.11s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 53: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:04:54 [abstract.py:324] It took 0.053748 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06874727e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55738068e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69262390e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06874727e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55738068e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69262390e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06874727e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55738068e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69262390e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06874727e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55738068e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.69262390e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06884414e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56174316e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69544144e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22791290e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.12390804e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #53] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44166031e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06884414e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56174316e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69544144e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71139221e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36312065e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #53] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.27899132e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06884414e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56174316e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69544144e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.33043823e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.36715698e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #53] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19109421e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06884414e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56174316e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.69544144e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26283569e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66390667e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #53] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32238342e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 53: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:04:56 [abstract.py:324] It took 0.029994 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13149.83it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:04:57 [loggers.py:236] Engine 000: Avg prompt throughput: 33.4 tokens/s, Avg generation throughput: 41.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:33,  1.69it/s, est. speed input: 103.21 toks/s, output: 20.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:09, 16.33it/s, est. speed input: 780.05 toks/s, output: 201.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:01<00:08, 16.55it/s, est. speed input: 859.72 toks/s, output: 349.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:09, 15.29it/s, est. speed input: 845.26 toks/s, output: 454.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:06, 22.34it/s, est. speed input: 1043.38 toks/s, output: 768.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 23.91it/s, est. speed input: 1109.32 toks/s, output: 944.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:04, 30.21it/s, est. speed input: 1256.40 toks/s, output: 1267.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 32.11it/s, est. speed input: 1322.89 toks/s, output: 1463.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:03, 31.73it/s, est. speed input: 1362.50 toks/s, output: 1637.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:03, 33.26it/s, est. speed input: 1422.65 toks/s, output: 1880.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:02<00:04, 26.61it/s, est. speed input: 1384.45 toks/s, output: 1982.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:02<00:04, 22.60it/s, est. speed input: 1345.35 toks/s, output: 2097.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:03<00:10, 10.25it/s, est. speed input: 1064.34 toks/s, output: 1791.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:05<00:28,  3.50it/s, est. speed input: 650.15 toks/s, output: 1253.12 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:07<00:39,  2.53it/s, est. speed input: 519.03 toks/s, output: 1153.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:08<00:39,  2.45it/s, est. speed input: 475.87 toks/s, output: 1215.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:08<00:41,  2.33it/s, est. speed input: 452.54 toks/s, output: 1240.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:09<00:53,  1.79it/s, est. speed input: 403.34 toks/s, output: 1190.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:05:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11329.0 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:12<01:24,  1.11it/s, est. speed input: 331.01 toks/s, output: 1063.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:12<01:16,  1.21it/s, est. speed input: 321.46 toks/s, output: 1120.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:14<01:26,  1.06it/s, est. speed input: 295.77 toks/s, output: 1117.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:14<01:11,  1.27it/s, est. speed input: 293.13 toks/s, output: 1193.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:15<01:11,  1.27it/s, est. speed input: 281.78 toks/s, output: 1233.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:16<01:19,  1.11it/s, est. speed input: 265.17 toks/s, output: 1246.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:16<00:59,  1.47it/s, est. speed input: 266.81 toks/s, output: 1338.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:17<00:58,  1.48it/s, est. speed input: 260.20 toks/s, output: 1388.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:18<01:14,  1.15it/s, est. speed input: 244.60 toks/s, output: 1389.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:19<00:49,  1.69it/s, est. speed input: 244.34 toks/s, output: 1551.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:19<00:53,  1.55it/s, est. speed input: 237.50 toks/s, output: 1590.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:05:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9027.4 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:21<01:14,  1.10it/s, est. speed input: 222.04 toks/s, output: 1568.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:22<01:23,  1.03s/it, est. speed input: 211.53 toks/s, output: 1577.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:23<00:50,  1.55it/s, est. speed input: 214.05 toks/s, output: 1757.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:23<00:40,  1.91it/s, est. speed input: 215.39 toks/s, output: 1847.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:24<00:46,  1.66it/s, est. speed input: 210.43 toks/s, output: 1884.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:24<00:40,  1.86it/s, est. speed input: 209.76 toks/s, output: 1957.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:24<00:34,  2.14it/s, est. speed input: 209.83 toks/s, output: 2035.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:26<01:07,  1.10it/s, est. speed input: 196.24 toks/s, output: 1982.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:26<00:50,  1.46it/s, est. speed input: 197.43 toks/s, output: 2072.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:28<01:11,  1.00it/s, est. speed input: 187.44 toks/s, output: 2045.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:29<01:07,  1.06it/s, est. speed input: 184.32 toks/s, output: 2089.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:29<00:54,  1.28it/s, est. speed input: 184.05 toks/s, output: 2161.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:05:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7404.1 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:30<00:50,  1.36it/s, est. speed input: 182.31 toks/s, output: 2217.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:31<00:46,  1.46it/s, est. speed input: 180.94 toks/s, output: 2276.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:31<00:43,  1.54it/s, est. speed input: 179.61 toks/s, output: 2335.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<00:37,  1.74it/s, est. speed input: 179.22 toks/s, output: 2406.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:32<00:30,  2.14it/s, est. speed input: 179.91 toks/s, output: 2490.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<00:23,  2.70it/s, est. speed input: 181.04 toks/s, output: 2579.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:32<00:21,  2.97it/s, est. speed input: 181.45 toks/s, output: 2659.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:32<00:19,  3.17it/s, est. speed input: 181.78 toks/s, output: 2737.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:33<00:27,  2.21it/s, est. speed input: 179.48 toks/s, output: 2775.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:34<00:19,  3.02it/s, est. speed input: 181.04 toks/s, output: 2944.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:35<00:22,  2.55it/s, est. speed input: 179.44 toks/s, output: 3061.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:35<00:21,  2.60it/s, est. speed input: 179.30 toks/s, output: 3131.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:35<00:21,  2.52it/s, est. speed input: 178.86 toks/s, output: 3194.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:37<00:33,  1.60it/s, est. speed input: 174.28 toks/s, output: 3183.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:37<00:18,  2.79it/s, est. speed input: 177.65 toks/s, output: 3455.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:37<00:15,  3.22it/s, est. speed input: 178.79 toks/s, output: 3617.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:14,  3.33it/s, est. speed input: 179.32 toks/s, output: 3766.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:23,  1.97it/s, est. speed input: 174.49 toks/s, output: 3734.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:40<00:20,  2.19it/s, est. speed input: 174.86 toks/s, output: 3812.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:05:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6153.3 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:41<00:27,  1.59it/s, est. speed input: 171.25 toks/s, output: 3804.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:41<00:21,  1.99it/s, est. speed input: 172.19 toks/s, output: 3896.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:41<00:14,  2.79it/s, est. speed input: 173.88 toks/s, output: 4072.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:12,  3.25it/s, est. speed input: 174.95 toks/s, output: 4234.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:11,  3.20it/s, est. speed input: 174.97 toks/s, output: 4304.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:10,  3.54it/s, est. speed input: 175.65 toks/s, output: 4389.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:11,  3.04it/s, est. speed input: 175.18 toks/s, output: 4446.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:12,  2.74it/s, est. speed input: 174.70 toks/s, output: 4502.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:27,  1.24it/s, est. speed input: 168.42 toks/s, output: 4411.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:16,  1.95it/s, est. speed input: 170.17 toks/s, output: 4595.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:15,  2.06it/s, est. speed input: 170.04 toks/s, output: 4661.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:15,  1.97it/s, est. speed input: 169.26 toks/s, output: 4710.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:11,  2.48it/s, est. speed input: 170.15 toks/s, output: 4804.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:10,  2.51it/s, est. speed input: 169.96 toks/s, output: 4937.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:05:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4409.5 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:50<00:15,  1.54it/s, est. speed input: 164.27 toks/s, output: 4977.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:16,  1.37it/s, est. speed input: 162.08 toks/s, output: 4984.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:19,  1.11it/s, est. speed input: 158.50 toks/s, output: 4949.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:53<00:13,  1.48it/s, est. speed input: 159.00 toks/s, output: 5112.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:53<00:10,  1.78it/s, est. speed input: 159.68 toks/s, output: 5209.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:54<00:06,  2.61it/s, est. speed input: 161.41 toks/s, output: 5414.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:05,  2.99it/s, est. speed input: 162.07 toks/s, output: 5509.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:06,  2.34it/s, est. speed input: 161.06 toks/s, output: 5548.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:55<00:04,  2.88it/s, est. speed input: 161.89 toks/s, output: 5649.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:57<00:07,  1.55it/s, est. speed input: 158.04 toks/s, output: 5662.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:57<00:06,  1.73it/s, est. speed input: 158.11 toks/s, output: 5741.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:58<00:06,  1.43it/s, est. speed input: 156.28 toks/s, output: 5754.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:59<00:06,  1.34it/s, est. speed input: 155.03 toks/s, output: 5787.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:59<00:05,  1.53it/s, est. speed input: 155.01 toks/s, output: 5866.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:05:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2562.8 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:02<00:08,  1.15s/it, est. speed input: 149.91 toks/s, output: 5759.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:03<00:06,  1.13s/it, est. speed input: 148.33 toks/s, output: 5784.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:03<00:04,  1.20it/s, est. speed input: 149.05 toks/s, output: 5898.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:04<00:04,  1.00s/it, est. speed input: 146.78 toks/s, output: 5897.03 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.00s/it, est. speed input: 150.55 toks/s, output: 6401.64 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.46it/s, est. speed input: 150.55 toks/s, output: 6401.64 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:06:05 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:06:05 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:06:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:06:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:06:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:06:05 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:06:06 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:06:06 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:06:06 [abstract.py:306] It took 0.593236 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:06:06 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:06:06 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0139
+  • Training Accuracy: 0.3000
+  • Mean Generation Length: 2598.0376
+  • Training FLOPS: 46.84 TFLOPS (11.71 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 312.25s
+  • policy_training: 198.42s (63.5%)
+  • generation: 69.39s (22.2%)
+  • teacher_logprob_inference: 39.07s (12.5%)
+  • policy_training/sharding_data: 4.33s (1.4%)
+  • prepare_for_generation: 3.52s (1.1%)
+  • policy_training/submit_training_futures: 1.89s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.5%)
+  • training_prep: 0.95s (0.3%)
+  • teacher_logprob_inference_prep: 0.70s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 54: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:06 [abstract.py:324] It took 0.051985 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06920273e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56232567e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70496368e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06920273e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56232567e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70496368e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06920273e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56232567e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70496368e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06920273e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56232567e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.70496368e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06919297e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56804047e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71475830e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22819405e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.06310797e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #54] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44493958e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06919297e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56804047e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71475830e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70547638e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35534916e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #54] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.27322350e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06919297e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56804047e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71475830e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.32427368e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.05760860e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #54] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19755096e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06919297e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56804047e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71475830e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27475967e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66397877e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #54] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32463760e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 54: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:09 [abstract.py:324] It took 0.030053 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12634.16it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:09 [loggers.py:236] Engine 000: Avg prompt throughput: 38.8 tokens/s, Avg generation throughput: 26.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:27,  1.81it/s, est. speed input: 108.83 toks/s, output: 21.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:08, 17.67it/s, est. speed input: 839.13 toks/s, output: 220.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:06, 22.25it/s, est. speed input: 1065.97 toks/s, output: 366.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 24.52it/s, est. speed input: 1181.21 toks/s, output: 503.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:04, 30.33it/s, est. speed input: 1362.61 toks/s, output: 721.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 39.18it/s, est. speed input: 1604.33 toks/s, output: 1049.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 43.18it/s, est. speed input: 1754.04 toks/s, output: 1310.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 49.79it/s, est. speed input: 1935.34 toks/s, output: 1664.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 47.99it/s, est. speed input: 2006.20 toks/s, output: 1928.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 46.62it/s, est. speed input: 2065.96 toks/s, output: 2190.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 49.05it/s, est. speed input: 2154.99 toks/s, output: 2551.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 50.65it/s, est. speed input: 2225.72 toks/s, output: 2870.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:02, 42.00it/s, est. speed input: 2189.28 toks/s, output: 3070.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:02<00:02, 30.51it/s, est. speed input: 2051.98 toks/s, output: 3107.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:02<00:03, 21.75it/s, est. speed input: 1870.24 toks/s, output: 3042.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:06<00:19,  3.71it/s, est. speed input: 843.51 toks/s, output: 1579.02 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:07<00:19,  3.55it/s, est. speed input: 755.88 toks/s, output: 1681.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:07<00:18,  3.72it/s, est. speed input: 731.22 toks/s, output: 1815.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:07<00:15,  4.23it/s, est. speed input: 727.00 toks/s, output: 1995.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9735.3 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:11<00:37,  1.74it/s, est. speed input: 512.36 toks/s, output: 1601.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:13<00:50,  1.27it/s, est. speed input: 436.86 toks/s, output: 1470.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:43,  1.46it/s, est. speed input: 436.48 toks/s, output: 1572.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:15<00:50,  1.24it/s, est. speed input: 401.53 toks/s, output: 1551.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:16<00:56,  1.08it/s, est. speed input: 371.78 toks/s, output: 1542.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:16<00:45,  1.31it/s, est. speed input: 370.18 toks/s, output: 1639.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:18<00:50,  1.15it/s, est. speed input: 335.84 toks/s, output: 1686.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:47,  1.20it/s, est. speed input: 327.01 toks/s, output: 1745.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:19<00:42,  1.32it/s, est. speed input: 321.46 toks/s, output: 1819.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:20<00:37,  1.47it/s, est. speed input: 317.35 toks/s, output: 1899.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7634.0 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:20<00:29,  1.81it/s, est. speed input: 316.99 toks/s, output: 2000.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:21<00:35,  1.50it/s, est. speed input: 305.49 toks/s, output: 2031.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:22<00:32,  1.59it/s, est. speed input: 300.78 toks/s, output: 2102.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:22<00:22,  2.21it/s, est. speed input: 299.87 toks/s, output: 2298.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:22<00:18,  2.63it/s, est. speed input: 300.39 toks/s, output: 2403.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:23<00:16,  2.93it/s, est. speed input: 298.44 toks/s, output: 2584.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:23<00:17,  2.69it/s, est. speed input: 295.01 toks/s, output: 2653.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:24<00:25,  1.77it/s, est. speed input: 283.64 toks/s, output: 2650.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:22,  1.97it/s, est. speed input: 282.07 toks/s, output: 2735.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:26<00:33,  1.28it/s, est. speed input: 268.24 toks/s, output: 2701.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:29,  1.40it/s, est. speed input: 265.26 toks/s, output: 2771.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:26,  1.56it/s, est. speed input: 263.02 toks/s, output: 2846.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:28<00:21,  1.88it/s, est. speed input: 262.83 toks/s, output: 2942.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:28<00:13,  2.81it/s, est. speed input: 264.41 toks/s, output: 3156.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:28<00:10,  3.41it/s, est. speed input: 265.14 toks/s, output: 3357.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:29<00:12,  2.80it/s, est. speed input: 260.71 toks/s, output: 3493.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:29<00:11,  2.84it/s, est. speed input: 259.85 toks/s, output: 3577.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5749.8 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:12,  2.55it/s, est. speed input: 257.34 toks/s, output: 3638.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:31<00:11,  2.62it/s, est. speed input: 253.99 toks/s, output: 3874.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:32<00:12,  2.30it/s, est. speed input: 250.69 toks/s, output: 3920.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:32<00:10,  2.68it/s, est. speed input: 251.42 toks/s, output: 4028.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:32<00:10,  2.60it/s, est. speed input: 250.01 toks/s, output: 4101.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:33<00:08,  2.98it/s, est. speed input: 250.37 toks/s, output: 4202.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:33<00:10,  2.21it/s, est. speed input: 246.35 toks/s, output: 4231.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:34<00:10,  2.23it/s, est. speed input: 245.07 toks/s, output: 4303.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:34<00:08,  2.52it/s, est. speed input: 244.90 toks/s, output: 4397.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:35<00:09,  2.16it/s, est. speed input: 242.18 toks/s, output: 4446.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:35<00:10,  1.95it/s, est. speed input: 239.58 toks/s, output: 4495.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:36<00:10,  1.80it/s, est. speed input: 236.93 toks/s, output: 4542.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:37<00:11,  1.53it/s, est. speed input: 232.97 toks/s, output: 4565.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:37<00:06,  2.44it/s, est. speed input: 234.80 toks/s, output: 4794.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:38<00:06,  2.41it/s, est. speed input: 233.71 toks/s, output: 4870.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:38<00:05,  2.37it/s, est. speed input: 232.58 toks/s, output: 4946.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:39<00:08,  1.50it/s, est. speed input: 226.44 toks/s, output: 4915.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:39<00:04,  2.46it/s, est. speed input: 228.54 toks/s, output: 5163.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:40<00:04,  2.38it/s, est. speed input: 227.47 toks/s, output: 5238.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3437.9 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:41<00:05,  1.56it/s, est. speed input: 221.83 toks/s, output: 5213.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:42<00:05,  1.46it/s, est. speed input: 219.07 toks/s, output: 5251.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:42<00:04,  1.57it/s, est. speed input: 217.86 toks/s, output: 5328.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:43<00:03,  1.98it/s, est. speed input: 218.44 toks/s, output: 5446.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:43<00:02,  1.89it/s, est. speed input: 216.90 toks/s, output: 5512.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:45<00:04,  1.01s/it, est. speed input: 207.96 toks/s, output: 5397.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:46<00:02,  1.31it/s, est. speed input: 208.51 toks/s, output: 5522.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:47<00:01,  1.04it/s, est. speed input: 203.54 toks/s, output: 5504.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:48<00:00,  1.11it/s, est. speed input: 201.52 toks/s, output: 5567.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:10:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1085.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.65s/it, est. speed input: 189.48 toks/s, output: 5359.99 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.65s/it, est. speed input: 189.48 toks/s, output: 5359.99 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  3.10it/s, est. speed input: 189.48 toks/s, output: 5359.99 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:11:03 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:11:03 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:11:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:11:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:11:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:11:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:11:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:11:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:11:03 [abstract.py:306] It took 0.614385 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:11:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:11:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0167
+  • Training Accuracy: 0.1688
+  • Mean Generation Length: 1731.2063
+  • Training FLOPS: 40.21 TFLOPS (10.05 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 241.79s
+  • policy_training: 153.36s (63.4%)
+  • generation: 54.88s (22.7%)
+  • teacher_logprob_inference: 28.61s (11.8%)
+  • policy_training/sharding_data: 4.29s (1.8%)
+  • prepare_for_generation: 3.46s (1.4%)
+  • policy_training/submit_training_futures: 1.90s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.6%)
+  • training_prep: 0.85s (0.4%)
+  • teacher_logprob_inference_prep: 0.43s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 55: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:08 [abstract.py:324] It took 0.051317 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06983594e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06983594e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06983594e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.06983594e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56966629e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.71737595e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56966629e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.71737595e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56966629e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.71737595e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56966629e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.71737595e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06984922e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58240509e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71153412e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.22244701e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.04175091e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #55] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44630951e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06984922e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58240509e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71153412e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.72251129e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35160856e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #55] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.28647499e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06984922e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58240509e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71153412e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.32014923e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.01839590e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #55] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.21938400e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.06984922e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58240509e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.71153412e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26850967e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.66487732e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #55] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32449463e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 55: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:11 [abstract.py:324] It took 0.030565 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13752.38it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:11 [loggers.py:236] Engine 000: Avg prompt throughput: 51.0 tokens/s, Avg generation throughput: 2.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:10,  2.27it/s, est. speed input: 138.46 toks/s, output: 31.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:11, 13.46it/s, est. speed input: 684.36 toks/s, output: 217.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:07, 19.19it/s, est. speed input: 926.12 toks/s, output: 388.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 22.90it/s, est. speed input: 1082.20 toks/s, output: 556.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 28.76it/s, est. speed input: 1278.15 toks/s, output: 805.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:02, 46.09it/s, est. speed input: 1739.40 toks/s, output: 1420.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 51.82it/s, est. speed input: 1971.08 toks/s, output: 1831.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 53.89it/s, est. speed input: 2116.94 toks/s, output: 2186.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 43.51it/s, est. speed input: 2073.47 toks/s, output: 2335.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 36.85it/s, est. speed input: 2010.91 toks/s, output: 2478.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 35.72it/s, est. speed input: 2009.05 toks/s, output: 2709.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:02<00:03, 29.85it/s, est. speed input: 1925.99 toks/s, output: 2793.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:02<00:04, 20.93it/s, est. speed input: 1739.61 toks/s, output: 2743.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:06<00:29,  2.95it/s, est. speed input: 659.93 toks/s, output: 1234.71 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:09<00:43,  1.94it/s, est. speed input: 480.27 toks/s, output: 1080.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10632.1 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:11<00:52,  1.58it/s, est. speed input: 398.83 toks/s, output: 1083.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:16<01:27,  1.07s/it, est. speed input: 297.33 toks/s, output: 905.91 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:18<01:39,  1.23s/it, est. speed input: 265.83 toks/s, output: 907.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8552.2 tokens/s, Running: 80 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:20<01:49,  1.37s/it, est. speed input: 242.15 toks/s, output: 922.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:21<01:49,  1.39s/it, est. speed input: 228.73 toks/s, output: 967.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:22<01:39,  1.27s/it, est. speed input: 222.44 toks/s, output: 1036.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:23<01:35,  1.24s/it, est. speed input: 214.47 toks/s, output: 1093.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:23<01:14,  1.02it/s, est. speed input: 214.90 toks/s, output: 1188.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:25<01:28,  1.18s/it, est. speed input: 202.80 toks/s, output: 1213.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:27<01:34,  1.27s/it, est. speed input: 193.80 toks/s, output: 1250.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:27<01:19,  1.09s/it, est. speed input: 191.74 toks/s, output: 1327.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:28<01:11,  1.00it/s, est. speed input: 188.62 toks/s, output: 1396.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:28<00:54,  1.30it/s, est. speed input: 189.40 toks/s, output: 1489.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:29<00:48,  1.44it/s, est. speed input: 188.16 toks/s, output: 1567.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<00:45,  1.52it/s, est. speed input: 186.54 toks/s, output: 1641.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7442.6 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:30<00:52,  1.29it/s, est. speed input: 182.09 toks/s, output: 1689.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:31<00:44,  1.52it/s, est. speed input: 181.79 toks/s, output: 1772.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<00:46,  1.42it/s, est. speed input: 179.06 toks/s, output: 1830.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:32<00:34,  1.88it/s, est. speed input: 180.27 toks/s, output: 1926.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<00:30,  2.09it/s, est. speed input: 180.19 toks/s, output: 2009.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:32<00:23,  2.69it/s, est. speed input: 181.42 toks/s, output: 2104.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:34<00:45,  1.37it/s, est. speed input: 174.93 toks/s, output: 2112.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:35<00:55,  1.09it/s, est. speed input: 170.02 toks/s, output: 2135.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:35<00:33,  1.78it/s, est. speed input: 172.06 toks/s, output: 2325.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:36<00:30,  1.89it/s, est. speed input: 171.67 toks/s, output: 2401.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:36<00:26,  2.12it/s, est. speed input: 171.82 toks/s, output: 2484.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:37<00:25,  2.21it/s, est. speed input: 171.57 toks/s, output: 2561.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:37<00:31,  1.75it/s, est. speed input: 169.18 toks/s, output: 2606.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:38<00:18,  2.84it/s, est. speed input: 171.64 toks/s, output: 2803.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:38<00:17,  2.91it/s, est. speed input: 172.00 toks/s, output: 2962.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:39<00:16,  3.02it/s, est. speed input: 172.27 toks/s, output: 3044.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:39<00:13,  3.57it/s, est. speed input: 173.33 toks/s, output: 3140.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:40<00:20,  2.38it/s, est. speed input: 171.27 toks/s, output: 3179.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:14:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6313.8 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:40<00:24,  1.90it/s, est. speed input: 169.28 toks/s, output: 3220.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:41<00:28,  1.62it/s, est. speed input: 167.30 toks/s, output: 3259.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:42<00:20,  2.14it/s, est. speed input: 168.02 toks/s, output: 3423.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:44<00:34,  1.24it/s, est. speed input: 162.26 toks/s, output: 3383.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:45<00:40,  1.03it/s, est. speed input: 158.45 toks/s, output: 3381.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:45<00:33,  1.24it/s, est. speed input: 158.56 toks/s, output: 3460.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:46<00:26,  1.53it/s, est. speed input: 159.00 toks/s, output: 3547.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:47<00:27,  1.39it/s, est. speed input: 156.23 toks/s, output: 3637.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:48<00:30,  1.22it/s, est. speed input: 153.81 toks/s, output: 3659.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:49<00:24,  1.48it/s, est. speed input: 154.27 toks/s, output: 3747.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:49<00:19,  1.76it/s, est. speed input: 154.66 toks/s, output: 3834.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:15:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4873.1 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:50<00:24,  1.40it/s, est. speed input: 152.46 toks/s, output: 3857.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:50<00:18,  1.76it/s, est. speed input: 153.09 toks/s, output: 3950.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:51<00:11,  2.61it/s, est. speed input: 154.59 toks/s, output: 4141.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:51<00:14,  2.14it/s, est. speed input: 153.60 toks/s, output: 4190.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:52<00:11,  2.48it/s, est. speed input: 154.09 toks/s, output: 4280.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:52<00:13,  2.03it/s, est. speed input: 153.14 toks/s, output: 4329.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:53<00:13,  2.02it/s, est. speed input: 152.85 toks/s, output: 4398.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:53<00:13,  1.90it/s, est. speed input: 152.25 toks/s, output: 4457.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:54<00:15,  1.60it/s, est. speed input: 150.93 toks/s, output: 4496.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:55<00:13,  1.79it/s, est. speed input: 150.93 toks/s, output: 4574.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:55<00:07,  2.90it/s, est. speed input: 152.65 toks/s, output: 4779.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:55<00:07,  2.67it/s, est. speed input: 152.43 toks/s, output: 4849.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:58<00:17,  1.16it/s, est. speed input: 147.53 toks/s, output: 4772.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:00<00:22,  1.19s/it, est. speed input: 143.42 toks/s, output: 4720.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:15:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3485.9 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:01<00:20,  1.13s/it, est. speed input: 142.18 toks/s, output: 4759.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:02<00:19,  1.13s/it, est. speed input: 140.56 toks/s, output: 4787.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:04<00:24,  1.51s/it, est. speed input: 136.17 toks/s, output: 4721.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:05<00:17,  1.18s/it, est. speed input: 136.38 toks/s, output: 4811.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:08<00:26,  1.86s/it, est. speed input: 130.25 toks/s, output: 4684.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:08<00:17,  1.38s/it, est. speed input: 130.73 toks/s, output: 4788.05 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  1.38s/it, est. speed input: 142.27 toks/s, output: 6336.65 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  2.33it/s, est. speed input: 142.27 toks/s, output: 6336.65 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:15:23 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:15:23 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:15:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:15:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:15:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:15:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:15:24 [abstract.py:306] It took 0.537636 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:15:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:15:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:15:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:15:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved▶ Training policy...
+
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.2062
+  • Mean Generation Length: 2723.5063
+  • Training FLOPS: 49.06 TFLOPS (12.26 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 318.00s
+  • policy_training: 201.85s (63.5%)
+  • generation: 73.30s (23.1%)
+  • teacher_logprob_inference: 37.79s (11.9%)
+  • policy_training/sharding_data: 4.32s (1.4%)
+  • prepare_for_generation: 3.54s (1.1%)
+  • policy_training/submit_training_futures: 1.64s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.5%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 12/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 56: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:19:26 [abstract.py:324] It took 0.051512 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07024434e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07024434e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07024434e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07024434e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57799339e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73375702e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57799339e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73375702e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57799339e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73375702e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57799339e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73375702e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07039375e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58272476e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73972778e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20178413e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.03434181e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #56] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44620575e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07039375e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58272476e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73972778e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71591034e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34545708e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #56] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.23567848e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07039375e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58272476e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73972778e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.32217102e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-2.50640154e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #56] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19692764e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07039375e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58272476e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73972778e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26671104e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67591419e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #56] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32364838e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 56: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:19:28 [abstract.py:324] It took 0.030439 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13437.36it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:19:28 [loggers.py:236] Engine 000: Avg prompt throughput: 37.8 tokens/s, Avg generation throughput: 87.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:28,  5.53it/s, est. speed input: 331.73 toks/s, output: 77.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:13, 11.56it/s, est. speed input: 632.28 toks/s, output: 208.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:08, 17.21it/s, est. speed input: 902.26 toks/s, output: 400.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:10, 15.06it/s, est. speed input: 857.05 toks/s, output: 467.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:06, 21.57it/s, est. speed input: 1096.74 toks/s, output: 782.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:06, 22.34it/s, est. speed input: 1152.29 toks/s, output: 936.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 25.38it/s, est. speed input: 1255.25 toks/s, output: 1166.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:05, 24.79it/s, est. speed input: 1273.67 toks/s, output: 1303.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 24.26it/s, est. speed input: 1288.65 toks/s, output: 1446.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:04, 27.69it/s, est. speed input: 1366.58 toks/s, output: 1702.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 34.18it/s, est. speed input: 1514.28 toks/s, output: 2164.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:04, 28.35it/s, est. speed input: 1481.42 toks/s, output: 2281.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:02<00:06, 17.92it/s, est. speed input: 1306.58 toks/s, output: 2158.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:02<00:08, 13.03it/s, est. speed input: 1161.80 toks/s, output: 2083.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:03<00:11,  9.30it/s, est. speed input: 1017.58 toks/s, output: 1945.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:19:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11839.8 tokens/s, Running: 109 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:11<01:43,  1.04it/s, est. speed input: 285.75 toks/s, output: 662.09 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:13<01:57,  1.09s/it, est. speed input: 247.26 toks/s, output: 663.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:13<01:41,  1.04it/s, est. speed input: 247.27 toks/s, output: 752.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:13<01:26,  1.22it/s, est. speed input: 247.72 toks/s, output: 840.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:14<01:29,  1.16it/s, est. speed input: 234.70 toks/s, output: 883.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:14<01:13,  1.41it/s, est. speed input: 235.56 toks/s, output: 971.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:15<01:11,  1.43it/s, est. speed input: 229.56 toks/s, output: 1029.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:16<01:08,  1.48it/s, est. speed input: 224.73 toks/s, output: 1090.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:16<01:06,  1.49it/s, est. speed input: 219.59 toks/s, output: 1147.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:17<00:57,  1.72it/s, est. speed input: 218.73 toks/s, output: 1223.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:18<01:22,  1.19it/s, est. speed input: 204.60 toks/s, output: 1224.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:18<01:05,  1.48it/s, est. speed input: 204.79 toks/s, output: 1304.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:19<00:40,  2.35it/s, est. speed input: 208.57 toks/s, output: 1484.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:19:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9727.7 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:20<00:53,  1.77it/s, est. speed input: 201.09 toks/s, output: 1509.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:20<00:43,  2.13it/s, est. speed input: 202.25 toks/s, output: 1593.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:21<00:55,  1.66it/s, est. speed input: 195.79 toks/s, output: 1619.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:23<01:26,  1.05it/s, est. speed input: 182.70 toks/s, output: 1586.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:24<01:37,  1.08s/it, est. speed input: 174.88 toks/s, output: 1594.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:26<01:53,  1.27s/it, est. speed input: 165.49 toks/s, output: 1584.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:26<01:29,  1.02s/it, est. speed input: 165.17 toks/s, output: 1657.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:29<02:04,  1.43s/it, est. speed input: 153.58 toks/s, output: 1615.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:29<01:30,  1.05s/it, est. speed input: 154.78 toks/s, output: 1703.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:29<00:43,  1.92it/s, est. speed input: 159.27 toks/s, output: 1971.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:19:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8145.2 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:30<00:47,  1.74it/s, est. speed input: 157.26 toks/s, output: 2017.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:30<00:38,  2.10it/s, est. speed input: 158.46 toks/s, output: 2103.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:30<00:35,  2.28it/s, est. speed input: 158.78 toks/s, output: 2177.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:32<00:39,  1.98it/s, est. speed input: 156.69 toks/s, output: 2283.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:32<00:32,  2.37it/s, est. speed input: 157.91 toks/s, output: 2369.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:32<00:38,  1.97it/s, est. speed input: 156.03 toks/s, output: 2409.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:33<00:34,  2.17it/s, est. speed input: 156.35 toks/s, output: 2481.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:34<00:41,  1.76it/s, est. speed input: 154.14 toks/s, output: 2515.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:36<01:08,  1.06it/s, est. speed input: 147.72 toks/s, output: 2476.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:36<00:50,  1.41it/s, est. speed input: 148.88 toks/s, output: 2563.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:37<00:55,  1.27it/s, est. speed input: 146.60 toks/s, output: 2590.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:37<00:39,  1.73it/s, est. speed input: 147.29 toks/s, output: 2734.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:39<00:41,  1.63it/s, est. speed input: 145.28 toks/s, output: 2826.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:39<00:37,  1.78it/s, est. speed input: 145.39 toks/s, output: 2894.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6759.5 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.5%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:40<00:42,  1.52it/s, est. speed input: 143.42 toks/s, output: 2920.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:41<00:36,  1.73it/s, est. speed input: 143.19 toks/s, output: 3043.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:38,  1.62it/s, est. speed input: 142.08 toks/s, output: 3084.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:42<00:37,  1.63it/s, est. speed input: 141.46 toks/s, output: 3136.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:43<00:33,  1.80it/s, est. speed input: 141.63 toks/s, output: 3203.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:43<00:33,  1.74it/s, est. speed input: 141.00 toks/s, output: 3252.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:44<00:28,  2.04it/s, est. speed input: 141.50 toks/s, output: 3327.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:44<00:35,  1.62it/s, est. speed input: 139.95 toks/s, output: 3354.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:45<00:28,  1.98it/s, est. speed input: 140.58 toks/s, output: 3433.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:45<00:25,  2.15it/s, est. speed input: 140.76 toks/s, output: 3501.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:46<00:24,  2.17it/s, est. speed input: 140.75 toks/s, output: 3562.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:46<00:30,  1.77it/s, est. speed input: 139.60 toks/s, output: 3596.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:47<00:14,  3.52it/s, est. speed input: 142.92 toks/s, output: 3869.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:47<00:19,  2.53it/s, est. speed input: 141.79 toks/s, output: 3899.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:48<00:16,  2.85it/s, est. speed input: 142.45 toks/s, output: 3978.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:48<00:22,  2.10it/s, est. speed input: 141.19 toks/s, output: 4005.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:49<00:17,  2.60it/s, est. speed input: 142.06 toks/s, output: 4091.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5749.5 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:50<00:26,  1.68it/s, est. speed input: 140.03 toks/s, output: 4093.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:51<00:33,  1.31it/s, est. speed input: 137.87 toks/s, output: 4094.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:51<00:21,  1.92it/s, est. speed input: 139.08 toks/s, output: 4253.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:52<00:21,  1.94it/s, est. speed input: 138.94 toks/s, output: 4310.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:52<00:16,  2.37it/s, est. speed input: 139.69 toks/s, output: 4457.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:55<00:32,  1.18it/s, est. speed input: 134.95 toks/s, output: 4369.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:57<00:40,  1.09s/it, est. speed input: 131.64 toks/s, output: 4325.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:57<00:32,  1.09it/s, est. speed input: 131.81 toks/s, output: 4394.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:58<00:29,  1.20it/s, est. speed input: 131.46 toks/s, output: 4447.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:58<00:27,  1.22it/s, est. speed input: 130.68 toks/s, output: 4487.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:59<00:20,  1.57it/s, est. speed input: 131.02 toks/s, output: 4625.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4594.1 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:00<00:18,  1.72it/s, est. speed input: 131.09 toks/s, output: 4693.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:00<00:15,  1.95it/s, est. speed input: 131.43 toks/s, output: 4770.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:00<00:13,  2.21it/s, est. speed input: 131.76 toks/s, output: 4847.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:00<00:09,  2.80it/s, est. speed input: 132.57 toks/s, output: 4940.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:00<00:05,  4.35it/s, est. speed input: 134.27 toks/s, output: 5129.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:01<00:10,  2.32it/s, est. speed input: 132.95 toks/s, output: 5142.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:03<00:11,  2.05it/s, est. speed input: 132.42 toks/s, output: 5251.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:03<00:09,  2.38it/s, est. speed input: 133.05 toks/s, output: 5338.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:03<00:07,  2.90it/s, est. speed input: 133.79 toks/s, output: 5431.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:03<00:06,  2.91it/s, est. speed input: 134.02 toks/s, output: 5504.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:04<00:06,  2.96it/s, est. speed input: 134.28 toks/s, output: 5579.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:04<00:05,  3.40it/s, est. speed input: 134.87 toks/s, output: 5667.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:04<00:04,  3.58it/s, est. speed input: 135.29 toks/s, output: 5749.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:04<00:03,  4.84it/s, est. speed input: 136.60 toks/s, output: 5934.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:05<00:03,  3.57it/s, est. speed input: 136.45 toks/s, output: 5992.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:05<00:03,  3.26it/s, est. speed input: 136.56 toks/s, output: 6061.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:06<00:05,  2.29it/s, est. speed input: 135.83 toks/s, output: 6094.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:06<00:03,  3.00it/s, est. speed input: 136.81 toks/s, output: 6268.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:07<00:04,  2.04it/s, est. speed input: 135.71 toks/s, output: 6283.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:08<00:03,  2.40it/s, est. speed input: 136.20 toks/s, output: 6373.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:08<00:02,  2.77it/s, est. speed input: 136.66 toks/s, output: 6461.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2655.2 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:10<00:05,  1.07it/s, est. speed input: 132.77 toks/s, output: 6348.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:12<00:03,  1.23it/s, est. speed input: 132.00 toks/s, output: 6455.80 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:12<00:00,  1.23it/s, est. speed input: 135.33 toks/s, output: 6910.88 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:12<00:00,  2.22it/s, est. speed input: 135.33 toks/s, output: 6910.88 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:45 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:45 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:20:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:20:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:20:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:20:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:20:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:20:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:20:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:20:45 [abstract.py:306] It took 0.651380 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:20:45 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0134
+  • Training Accuracy: 0.2625
+  • Mean Generation Length: 3109.9875
+  • Training FLOPS: 49.73 TFLOPS (12.43 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 351.24s
+  • policy_training: 224.82s (64.0%)
+  • generation: 77.14s (22.0%)
+  • teacher_logprob_inference: 44.39s (12.6%)
+  • policy_training/sharding_data: 4.40s (1.3%)
+  • prepare_for_generation: 3.34s (1.0%)
+  • policy_training/submit_training_futures: 1.59s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.4%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 57: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:25:17 [abstract.py:324] It took 0.061115 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07065977e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58441200e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74224777e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07065977e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58441200e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74224777e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07065977e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58441200e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74224777e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07065977e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58441200e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74224777e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07072715e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59514923e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74669495e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19502258e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.05327225e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #57] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44373276e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07072715e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59514923e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74669495e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.72245178e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34382868e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #57] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.25033150e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07072715e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59514923e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74669495e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31751709e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-2.80840158e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #57] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17433090e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07072715e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59514923e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74669495e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25955429e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67120247e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #57] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32227692e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 57: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:25:20 [abstract.py:324] It took 0.031687 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13570.23it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 03:25:21 [loggers.py:236] Engine 000: Avg prompt throughput: 34.6 tokens/s, Avg generation throughput: 9.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:15,  2.11it/s, est. speed input: 124.78 toks/s, output: 14.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:15, 10.29it/s, est. speed input: 509.61 toks/s, output: 111.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 26.65it/s, est. speed input: 1137.42 toks/s, output: 371.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 39.93it/s, est. speed input: 1606.62 toks/s, output: 664.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:04, 27.23it/s, est. speed input: 1440.25 toks/s, output: 788.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:04, 27.77it/s, est. speed input: 1486.98 toks/s, output: 1002.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:05, 20.97it/s, est. speed input: 1339.10 toks/s, output: 1074.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:06, 19.45it/s, est. speed input: 1293.41 toks/s, output: 1240.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:02<00:06, 17.61it/s, est. speed input: 1241.36 toks/s, output: 1354.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:02<00:07, 14.74it/s, est. speed input: 1158.36 toks/s, output: 1435.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:02<00:10, 10.66it/s, est. speed input: 1032.69 toks/s, output: 1406.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:03<00:11,  9.78it/s, est. speed input: 983.51 toks/s, output: 1463.09 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:06<00:53,  2.03it/s, est. speed input: 475.92 toks/s, output: 836.92 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:08<01:09,  1.54it/s, est. speed input: 390.80 toks/s, output: 775.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:09<01:19,  1.33it/s, est. speed input: 345.98 toks/s, output: 775.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:09<01:08,  1.53it/s, est. speed input: 343.56 toks/s, output: 857.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:25:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11874.5 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:11<01:36,  1.08it/s, est. speed input: 292.83 toks/s, output: 818.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:14<02:09,  1.26s/it, est. speed input: 248.59 toks/s, output: 783.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:14<01:25,  1.18it/s, est. speed input: 248.86 toks/s, output: 955.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:15<01:33,  1.07it/s, est. speed input: 233.71 toks/s, output: 981.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:18<02:03,  1.25s/it, est. speed input: 207.70 toks/s, output: 957.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:18<01:50,  1.13s/it, est. speed input: 202.36 toks/s, output: 1016.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:20<01:55,  1.19s/it, est. speed input: 191.65 toks/s, output: 1046.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:25:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9697.4 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:20<01:29,  1.07it/s, est. speed input: 192.25 toks/s, output: 1131.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:20<01:15,  1.27it/s, est. speed input: 191.34 toks/s, output: 1207.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:21<00:57,  1.64it/s, est. speed input: 192.65 toks/s, output: 1295.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:21<01:05,  1.42it/s, est. speed input: 187.28 toks/s, output: 1338.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:22<00:55,  1.65it/s, est. speed input: 186.99 toks/s, output: 1414.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:23<01:00,  1.49it/s, est. speed input: 183.01 toks/s, output: 1462.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:23<00:49,  1.83it/s, est. speed input: 183.61 toks/s, output: 1543.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:24<01:03,  1.40it/s, est. speed input: 177.87 toks/s, output: 1571.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:24<00:54,  1.62it/s, est. speed input: 177.54 toks/s, output: 1644.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:25<00:52,  1.66it/s, est. speed input: 175.91 toks/s, output: 1704.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:26<00:52,  1.63it/s, est. speed input: 173.91 toks/s, output: 1759.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:26<00:44,  1.90it/s, est. speed input: 174.12 toks/s, output: 1835.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:26<00:33,  2.48it/s, est. speed input: 175.74 toks/s, output: 1924.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:26<00:33,  2.49it/s, est. speed input: 175.37 toks/s, output: 1993.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:28<00:42,  1.91it/s, est. speed input: 171.44 toks/s, output: 2089.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:28<00:44,  1.78it/s, est. speed input: 169.60 toks/s, output: 2136.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:29<00:39,  2.00it/s, est. speed input: 169.84 toks/s, output: 2209.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:25:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8038.6 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:31<01:05,  1.19it/s, est. speed input: 162.25 toks/s, output: 2181.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:31<00:48,  1.57it/s, est. speed input: 163.54 toks/s, output: 2270.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:31<00:35,  2.13it/s, est. speed input: 164.78 toks/s, output: 2423.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:32<00:33,  2.20it/s, est. speed input: 164.55 toks/s, output: 2489.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:32<00:37,  1.93it/s, est. speed input: 162.96 toks/s, output: 2532.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:32<00:31,  2.27it/s, est. speed input: 163.61 toks/s, output: 2611.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:34<00:45,  1.55it/s, est. speed input: 159.70 toks/s, output: 2617.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:34<00:40,  1.73it/s, est. speed input: 159.55 toks/s, output: 2683.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:34<00:30,  2.25it/s, est. speed input: 160.75 toks/s, output: 2770.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:35<00:29,  2.28it/s, est. speed input: 160.53 toks/s, output: 2833.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:35<00:24,  2.74it/s, est. speed input: 161.51 toks/s, output: 2980.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:36<00:25,  2.53it/s, est. speed input: 160.98 toks/s, output: 3036.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:36<00:25,  2.56it/s, est. speed input: 160.99 toks/s, output: 3101.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:37<00:25,  2.44it/s, est. speed input: 160.57 toks/s, output: 3158.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:37<00:31,  1.99it/s, est. speed input: 159.02 toks/s, output: 3193.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:38<00:38,  1.59it/s, est. speed input: 156.62 toks/s, output: 3211.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:38<00:30,  1.99it/s, est. speed input: 157.47 toks/s, output: 3292.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:39<00:29,  2.03it/s, est. speed input: 157.16 toks/s, output: 3349.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:39<00:23,  2.45it/s, est. speed input: 157.87 toks/s, output: 3429.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:39<00:18,  3.15it/s, est. speed input: 159.00 toks/s, output: 3517.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:39<00:15,  3.72it/s, est. speed input: 159.86 toks/s, output: 3600.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:40<00:13,  3.93it/s, est. speed input: 160.95 toks/s, output: 3750.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:26:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6653.8 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:40<00:14,  3.69it/s, est. speed input: 161.16 toks/s, output: 3817.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:41<00:18,  2.82it/s, est. speed input: 160.43 toks/s, output: 3923.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:41<00:13,  3.73it/s, est. speed input: 162.42 toks/s, output: 4095.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:42<00:19,  2.46it/s, est. speed input: 160.36 toks/s, output: 4104.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:43<00:14,  3.22it/s, est. speed input: 162.04 toks/s, output: 4271.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:43<00:09,  4.54it/s, est. speed input: 164.45 toks/s, output: 4457.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:43<00:14,  3.03it/s, est. speed input: 162.98 toks/s, output: 4478.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:44<00:12,  3.37it/s, est. speed input: 163.72 toks/s, output: 4558.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:44<00:11,  3.53it/s, est. speed input: 164.20 toks/s, output: 4632.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:44<00:09,  4.21it/s, est. speed input: 165.16 toks/s, output: 4720.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:44<00:09,  4.10it/s, est. speed input: 166.02 toks/s, output: 4864.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:45<00:06,  5.78it/s, est. speed input: 168.32 toks/s, output: 5051.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:45<00:08,  4.25it/s, est. speed input: 167.93 toks/s, output: 5099.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:46<00:14,  2.34it/s, est. speed input: 165.42 toks/s, output: 5083.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:47<00:14,  2.33it/s, est. speed input: 165.16 toks/s, output: 5136.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:48<00:20,  1.54it/s, est. speed input: 162.08 toks/s, output: 5103.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:49<00:21,  1.45it/s, est. speed input: 160.64 toks/s, output: 5122.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:49<00:18,  1.60it/s, est. speed input: 160.41 toks/s, output: 5177.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:50<00:17,  1.61it/s, est. speed input: 159.67 toks/s, output: 5217.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:26:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4960.1 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:50<00:14,  1.94it/s, est. speed input: 160.06 toks/s, output: 5294.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:50<00:10,  2.51it/s, est. speed input: 160.88 toks/s, output: 5385.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:51<00:07,  3.11it/s, est. speed input: 161.97 toks/s, output: 5610.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:52<00:11,  2.09it/s, est. speed input: 159.82 toks/s, output: 5599.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:52<00:11,  1.98it/s, est. speed input: 159.15 toks/s, output: 5641.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:10,  2.03it/s, est. speed input: 158.90 toks/s, output: 5697.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:53<00:08,  2.48it/s, est. speed input: 159.62 toks/s, output: 5787.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:05,  3.10it/s, est. speed input: 160.62 toks/s, output: 5951.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:03,  4.33it/s, est. speed input: 162.34 toks/s, output: 6144.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:03,  4.55it/s, est. speed input: 162.95 toks/s, output: 6229.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:54<00:03,  3.79it/s, est. speed input: 162.83 toks/s, output: 6289.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:54<00:03,  4.19it/s, est. speed input: 163.50 toks/s, output: 6377.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:03,  3.08it/s, est. speed input: 162.91 toks/s, output: 6420.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:56<00:05,  2.11it/s, est. speed input: 161.46 toks/s, output: 6430.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:57<00:06,  1.50it/s, est. speed input: 159.32 toks/s, output: 6411.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:58<00:06,  1.47it/s, est. speed input: 158.40 toks/s, output: 6444.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:58<00:04,  1.83it/s, est. speed input: 158.81 toks/s, output: 6532.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:03,  2.09it/s, est. speed input: 158.98 toks/s, output: 6610.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:03,  1.65it/s, est. speed input: 157.56 toks/s, output: 6623.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:59<00:02,  1.92it/s, est. speed input: 157.75 toks/s, output: 6703.32 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:26:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2453.6 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:02<00:03,  1.01it/s, est. speed input: 153.39 toks/s, output: 6595.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:02<00:02,  1.31it/s, est. speed input: 153.83 toks/s, output: 6691.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:02<00:01,  1.49it/s, est. speed input: 153.71 toks/s, output: 6763.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:04<00:00,  1.09it/s, est. speed input: 151.09 toks/s, output: 6730.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.11it/s, est. speed input: 150.00 toks/s, output: 6766.77 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.11it/s, est. speed input: 150.00 toks/s, output: 6766.77 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  2.46it/s, est. speed input: 150.00 toks/s, output: 6766.77 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:26:29 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:26:29 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:26:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:26:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:26:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:26:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:26:30 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:26:30 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:26:30 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 03:26:30 [abstract.py:306] It took 0.560045 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:26:30 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0120
+  • Training Accuracy: 0.1875
+  • Mean Generation Length: 2754.1125
+  • Training FLOPS: 46.26 TFLOPS (11.56 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 325.68s
+  • policy_training: 211.40s (64.9%)
+  • generation: 69.60s (21.4%)
+  • teacher_logprob_inference: 39.02s (12.0%)
+  • policy_training/sharding_data: 4.29s (1.3%)
+  • prepare_for_generation: 3.69s (1.1%)
+  • policy_training/submit_training_futures: 1.97s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.70s (0.5%)
+  • training_prep: 1.33s (0.4%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • get_topk_logits/submit_topk_logits_futures: 0.08s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 58: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:30:43 [abstract.py:324] It took 0.050308 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07118594e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59201355e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74217148e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07118594e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59201355e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74217148e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07118594e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59201355e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74217148e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07118594e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59201355e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74217148e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07117969e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59380951e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73974304e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20509014e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.19765854e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #58] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44481522e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07117969e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59380951e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73974304e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73062286e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33761311e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #58] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.28290215e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07117969e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59380951e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73974304e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31725235e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.07307148e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #58] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19408493e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07117969e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59380951e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73974304e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27338943e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67249374e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #58] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31922455e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 58: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:30:46 [abstract.py:324] It took 0.030109 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 03:30:46 [loggers.py:236] Engine 000: Avg prompt throughput: 36.6 tokens/s, Avg generation throughput: 14.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13171.26it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  5.90it/s, est. speed input: 348.39 toks/s, output: 76.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:07, 19.27it/s, est. speed input: 1044.45 toks/s, output: 320.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 22.73it/s, est. speed input: 1238.28 toks/s, output: 525.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 26.74it/s, est. speed input: 1412.86 toks/s, output: 758.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 27.63it/s, est. speed input: 1475.33 toks/s, output: 904.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 24.08it/s, est. speed input: 1408.05 toks/s, output: 999.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:06, 21.72it/s, est. speed input: 1352.92 toks/s, output: 1091.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 24.98it/s, est. speed input: 1425.17 toks/s, output: 1343.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:04, 30.61it/s, est. speed input: 1549.48 toks/s, output: 1735.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:03, 39.91it/s, est. speed input: 1739.97 toks/s, output: 2258.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:03, 37.26it/s, est. speed input: 1763.66 toks/s, output: 2506.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:03, 35.48it/s, est. speed input: 1786.75 toks/s, output: 2798.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 36.19it/s, est. speed input: 1817.63 toks/s, output: 3028.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:03, 32.85it/s, est. speed input: 1798.22 toks/s, output: 3186.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:02<00:03, 31.28it/s, est. speed input: 1787.63 toks/s, output: 3371.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:03<00:12,  7.19it/s, est. speed input: 1067.32 toks/s, output: 2249.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:05<00:20,  4.34it/s, est. speed input: 783.55 toks/s, output: 1867.61 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:09<00:48,  1.81it/s, est. speed input: 460.36 toks/s, output: 1241.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:30:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10940.9 tokens/s, Running: 88 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:10<00:43,  1.96it/s, est. speed input: 441.94 toks/s, output: 1380.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:11<00:46,  1.81it/s, est. speed input: 399.15 toks/s, output: 1421.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:11<00:41,  2.02it/s, est. speed input: 399.50 toks/s, output: 1514.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:12<00:40,  2.01it/s, est. speed input: 387.47 toks/s, output: 1560.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:12<00:43,  1.85it/s, est. speed input: 370.48 toks/s, output: 1584.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:14<00:54,  1.46it/s, est. speed input: 343.31 toks/s, output: 1558.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:14<00:43,  1.83it/s, est. speed input: 345.03 toks/s, output: 1657.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:14<00:36,  2.12it/s, est. speed input: 343.20 toks/s, output: 1738.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:15<00:38,  2.02it/s, est. speed input: 334.37 toks/s, output: 1782.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:17<01:11,  1.06it/s, est. speed input: 296.16 toks/s, output: 1670.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:17<01:00,  1.23it/s, est. speed input: 291.62 toks/s, output: 1735.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:17<00:49,  1.48it/s, est. speed input: 289.70 toks/s, output: 1811.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:18<00:39,  1.85it/s, est. speed input: 289.71 toks/s, output: 1899.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:19<00:45,  1.59it/s, est. speed input: 279.90 toks/s, output: 1924.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:19<00:37,  1.90it/s, est. speed input: 279.09 toks/s, output: 2005.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:19<00:29,  2.40it/s, est. speed input: 279.87 toks/s, output: 2098.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:19<00:23,  2.88it/s, est. speed input: 280.30 toks/s, output: 2187.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8411.9 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:20<00:36,  1.88it/s, est. speed input: 270.08 toks/s, output: 2194.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:21<00:46,  1.44it/s, est. speed input: 259.32 toks/s, output: 2193.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:22<00:41,  1.58it/s, est. speed input: 256.42 toks/s, output: 2253.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:22<00:31,  2.08it/s, est. speed input: 257.84 toks/s, output: 2349.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:23<00:46,  1.37it/s, est. speed input: 246.03 toks/s, output: 2328.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:26<01:20,  1.29s/it, est. speed input: 224.19 toks/s, output: 2207.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:26<01:01,  1.01it/s, est. speed input: 223.91 toks/s, output: 2292.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:26<00:45,  1.33it/s, est. speed input: 224.54 toks/s, output: 2385.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:26<00:35,  1.67it/s, est. speed input: 224.78 toks/s, output: 2473.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:27<00:29,  2.00it/s, est. speed input: 224.78 toks/s, output: 2558.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:27<00:25,  2.28it/s, est. speed input: 224.55 toks/s, output: 2640.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:29<00:54,  1.05it/s, est. speed input: 210.16 toks/s, output: 2557.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6987.0 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:30<00:56,  1.01s/it, est. speed input: 204.32 toks/s, output: 2571.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:31<00:49,  1.10it/s, est. speed input: 202.06 toks/s, output: 2628.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:32<00:48,  1.10it/s, est. speed input: 198.30 toks/s, output: 2664.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:34<00:45,  1.14it/s, est. speed input: 192.12 toks/s, output: 2748.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:34<00:31,  1.57it/s, est. speed input: 192.49 toks/s, output: 2924.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:35<00:29,  1.64it/s, est. speed input: 191.45 toks/s, output: 2991.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:35<00:23,  2.01it/s, est. speed input: 192.31 toks/s, output: 3090.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:35<00:19,  2.43it/s, est. speed input: 193.21 toks/s, output: 3186.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:36<00:17,  2.57it/s, est. speed input: 192.75 toks/s, output: 3342.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:36<00:17,  2.51it/s, est. speed input: 192.21 toks/s, output: 3414.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:17,  2.47it/s, est. speed input: 191.63 toks/s, output: 3485.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:37<00:20,  2.09it/s, est. speed input: 189.79 toks/s, output: 3533.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:37<00:15,  2.65it/s, est. speed input: 190.76 toks/s, output: 3633.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5663.2 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:40<00:36,  1.09it/s, est. speed input: 181.26 toks/s, output: 3535.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:27,  1.39it/s, est. speed input: 181.81 toks/s, output: 3628.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:40<00:21,  1.74it/s, est. speed input: 182.24 toks/s, output: 3720.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:40<00:17,  2.08it/s, est. speed input: 182.65 toks/s, output: 3808.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:42<00:28,  1.26it/s, est. speed input: 177.43 toks/s, output: 3783.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:42<00:24,  1.42it/s, est. speed input: 176.77 toks/s, output: 3852.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:43<00:22,  1.53it/s, est. speed input: 175.93 toks/s, output: 3916.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:44<00:18,  1.77it/s, est. speed input: 174.96 toks/s, output: 4058.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:44<00:11,  2.51it/s, est. speed input: 176.53 toks/s, output: 4257.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:45<00:15,  1.91it/s, est. speed input: 174.08 toks/s, output: 4280.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:17,  1.56it/s, est. speed input: 171.59 toks/s, output: 4301.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:13,  1.94it/s, est. speed input: 172.38 toks/s, output: 4402.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:12,  2.13it/s, est. speed input: 172.41 toks/s, output: 4486.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:47<00:09,  2.58it/s, est. speed input: 173.07 toks/s, output: 4584.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:07,  3.24it/s, est. speed input: 173.96 toks/s, output: 4688.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:48<00:13,  1.70it/s, est. speed input: 170.60 toks/s, output: 4681.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:09,  2.23it/s, est. speed input: 171.43 toks/s, output: 4787.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:08,  2.41it/s, est. speed input: 171.50 toks/s, output: 4870.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:49<00:08,  2.31it/s, est. speed input: 170.82 toks/s, output: 5013.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4160.1 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:50<00:06,  2.76it/s, est. speed input: 171.51 toks/s, output: 5116.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:50<00:05,  3.03it/s, est. speed input: 171.90 toks/s, output: 5209.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:09,  1.65it/s, est. speed input: 168.63 toks/s, output: 5192.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:52<00:10,  1.48it/s, est. speed input: 167.08 toks/s, output: 5227.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:09,  1.54it/s, est. speed input: 166.36 toks/s, output: 5289.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:53<00:09,  1.42it/s, est. speed input: 164.93 toks/s, output: 5328.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:54<00:07,  1.67it/s, est. speed input: 165.05 toks/s, output: 5416.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:54<00:05,  1.93it/s, est. speed input: 165.16 toks/s, output: 5506.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:03,  2.47it/s, est. speed input: 165.76 toks/s, output: 5696.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:55<00:04,  2.00it/s, est. speed input: 164.45 toks/s, output: 5739.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:56<00:03,  2.15it/s, est. speed input: 164.50 toks/s, output: 5827.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:03,  1.71it/s, est. speed input: 162.98 toks/s, output: 5861.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:57<00:02,  1.88it/s, est. speed input: 162.96 toks/s, output: 5947.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1973.6 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:06,  1.58s/it, est. speed input: 152.91 toks/s, output: 5676.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.58s/it, est. speed input: 156.84 toks/s, output: 6206.42 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.59it/s, est. speed input: 156.84 toks/s, output: 6206.42 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:31:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:31:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:31:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:31:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:31:52 [abstract.py:306] It took 0.682080 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:31:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:31:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:31:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:31:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0138
+  • Training Accuracy: 0.3125
+  • Mean Generation Length: 2398.0562
+  • Training FLOPS: 45.76 TFLOPS (11.44 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 294.40s
+  • policy_training: 188.09s (63.9%)
+  • generation: 66.01s (22.4%)
+  • teacher_logprob_inference: 35.13s (11.9%)
+  • policy_training/sharding_data: 4.28s (1.5%)
+  • prepare_for_generation: 3.38s (1.1%)
+  • policy_training/submit_training_futures: 1.72s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.5%)
+  • training_prep: 0.92s (0.3%)
+  • teacher_logprob_inference_prep: 0.68s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 59: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:35:38 [abstract.py:324] It took 0.048271 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07151562e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60629120e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73850327e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07151562e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60629120e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73850327e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07151562e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60629120e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73850327e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07151562e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60629120e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73850327e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07153105e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60296783e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73894882e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07153105e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60296783e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73894882e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07153105e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60296783e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73894882e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07153105e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60296783e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73894882e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20548038e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.19026136e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #59] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44496155e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31535492e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.95482779e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #59] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.20955544e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27081375e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65105209e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #59] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32181229e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73120651e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34210758e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #59] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.32803764e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 59: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:35:41 [abstract.py:324] It took 0.030709 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13518.29it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 03:35:41 [loggers.py:236] Engine 000: Avg prompt throughput: 41.2 tokens/s, Avg generation throughput: 9.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:08,  2.33it/s, est. speed input: 137.53 toks/s, output: 27.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:09, 16.48it/s, est. speed input: 790.41 toks/s, output: 201.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 31.27it/s, est. speed input: 1393.83 toks/s, output: 490.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 36.08it/s, est. speed input: 1611.24 toks/s, output: 684.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:04, 31.01it/s, est. speed input: 1589.66 toks/s, output: 838.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 37.20it/s, est. speed input: 1774.50 toks/s, output: 1151.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 36.75it/s, est. speed input: 1821.69 toks/s, output: 1369.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 40.56it/s, est. speed input: 1936.93 toks/s, output: 1681.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 41.56it/s, est. speed input: 1994.59 toks/s, output: 1932.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:03, 29.10it/s, est. speed input: 1830.63 toks/s, output: 1994.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:03, 26.83it/s, est. speed input: 1779.63 toks/s, output: 2178.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:03<00:12,  7.82it/s, est. speed input: 1052.68 toks/s, output: 1484.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:35:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11346.1 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:10<00:59,  1.62it/s, est. speed input: 360.87 toks/s, output: 796.12 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:11<00:42,  2.16it/s, est. speed input: 372.94 toks/s, output: 1191.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:12<00:47,  1.90it/s, est. speed input: 334.01 toks/s, output: 1243.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:16<01:04,  1.36it/s, est. speed input: 274.43 toks/s, output: 1191.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:17<01:10,  1.23it/s, est. speed input: 256.15 toks/s, output: 1201.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:17<01:02,  1.37it/s, est. speed input: 256.29 toks/s, output: 1290.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:18<01:08,  1.24it/s, est. speed input: 243.57 toks/s, output: 1314.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:19<01:02,  1.34it/s, est. speed input: 240.27 toks/s, output: 1383.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:19<00:55,  1.49it/s, est. speed input: 238.43 toks/s, output: 1459.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9042.7 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:20<00:57,  1.42it/s, est. speed input: 231.80 toks/s, output: 1505.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:22<00:59,  1.35it/s, est. speed input: 220.65 toks/s, output: 1597.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:25<01:25,  1.10s/it, est. speed input: 195.67 toks/s, output: 1575.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:25<01:13,  1.05it/s, est. speed input: 194.99 toks/s, output: 1654.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:26<01:02,  1.21it/s, est. speed input: 194.42 toks/s, output: 1732.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:26<00:54,  1.37it/s, est. speed input: 193.61 toks/s, output: 1805.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:28<01:12,  1.02it/s, est. speed input: 184.13 toks/s, output: 1799.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:28<00:56,  1.29it/s, est. speed input: 184.80 toks/s, output: 1888.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:29<00:49,  1.46it/s, est. speed input: 183.91 toks/s, output: 1960.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:29<00:50,  1.39it/s, est. speed input: 181.03 toks/s, output: 2010.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7507.3 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:30<00:48,  1.43it/s, est. speed input: 179.16 toks/s, output: 2068.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:31<00:49,  1.39it/s, est. speed input: 176.65 toks/s, output: 2120.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:31<00:38,  1.77it/s, est. speed input: 177.48 toks/s, output: 2208.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:32<00:38,  1.73it/s, est. speed input: 176.01 toks/s, output: 2268.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<00:41,  1.60it/s, est. speed input: 173.87 toks/s, output: 2318.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:33<00:42,  1.53it/s, est. speed input: 171.95 toks/s, output: 2370.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:36<01:17,  1.21s/it, est. speed input: 161.62 toks/s, output: 2305.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<01:05,  1.04s/it, est. speed input: 160.59 toks/s, output: 2367.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:37<00:57,  1.08it/s, est. speed input: 159.44 toks/s, output: 2428.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:37<00:46,  1.31it/s, est. speed input: 159.41 toks/s, output: 2503.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<01:05,  1.08s/it, est. speed input: 153.53 toks/s, output: 2489.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:40<00:53,  1.09it/s, est. speed input: 153.04 toks/s, output: 2559.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6555.4 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:40<00:45,  1.27it/s, est. speed input: 152.58 toks/s, output: 2630.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:40<00:35,  1.61it/s, est. speed input: 153.18 toks/s, output: 2717.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:41<00:18,  2.97it/s, est. speed input: 156.31 toks/s, output: 2997.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:43<00:35,  1.48it/s, est. speed input: 150.69 toks/s, output: 2965.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:44<00:45,  1.14it/s, est. speed input: 146.77 toks/s, output: 2963.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:45<00:38,  1.34it/s, est. speed input: 146.96 toks/s, output: 3043.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:45<00:35,  1.39it/s, est. speed input: 146.25 toks/s, output: 3102.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:45<00:20,  2.30it/s, est. speed input: 148.56 toks/s, output: 3298.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:46<00:20,  2.32it/s, est. speed input: 148.53 toks/s, output: 3370.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:47<00:25,  1.80it/s, est. speed input: 146.91 toks/s, output: 3407.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:47<00:24,  1.83it/s, est. speed input: 146.55 toks/s, output: 3472.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:48<00:21,  2.01it/s, est. speed input: 146.71 toks/s, output: 3548.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:49<00:27,  1.54it/s, est. speed input: 144.82 toks/s, output: 3575.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:49<00:24,  1.73it/s, est. speed input: 144.82 toks/s, output: 3649.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5404.9 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:50<00:33,  1.21it/s, est. speed input: 141.93 toks/s, output: 3651.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:51<00:23,  1.64it/s, est. speed input: 142.30 toks/s, output: 3807.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:51<00:18,  2.01it/s, est. speed input: 143.04 toks/s, output: 3899.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:51<00:15,  2.45it/s, est. speed input: 143.78 toks/s, output: 3991.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:52<00:18,  1.93it/s, est. speed input: 142.69 toks/s, output: 4033.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:53<00:20,  1.74it/s, est. speed input: 141.89 toks/s, output: 4082.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:53<00:09,  3.51it/s, est. speed input: 144.95 toks/s, output: 4383.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:53<00:07,  4.01it/s, est. speed input: 145.77 toks/s, output: 4478.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:53<00:07,  4.21it/s, est. speed input: 146.33 toks/s, output: 4565.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:54<00:06,  4.55it/s, est. speed input: 146.97 toks/s, output: 4655.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:54<00:07,  3.68it/s, est. speed input: 146.92 toks/s, output: 4723.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:54<00:07,  3.85it/s, est. speed input: 147.37 toks/s, output: 4808.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:55<00:07,  3.71it/s, est. speed input: 147.65 toks/s, output: 4887.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:55<00:06,  4.12it/s, est. speed input: 148.29 toks/s, output: 4976.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:56<00:09,  2.45it/s, est. speed input: 147.18 toks/s, output: 5009.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:56<00:08,  2.67it/s, est. speed input: 147.45 toks/s, output: 5088.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:56<00:07,  2.88it/s, est. speed input: 147.77 toks/s, output: 5168.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:57<00:07,  2.65it/s, est. speed input: 147.75 toks/s, output: 5304.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:57<00:06,  2.80it/s, est. speed input: 148.03 toks/s, output: 5384.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:57<00:05,  3.02it/s, est. speed input: 148.43 toks/s, output: 5467.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:58<00:07,  2.15it/s, est. speed input: 147.34 toks/s, output: 5498.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:58<00:05,  2.67it/s, est. speed input: 148.03 toks/s, output: 5593.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:00<00:08,  1.71it/s, est. speed input: 146.30 toks/s, output: 5598.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3634.0 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:01<00:08,  1.45it/s, est. speed input: 144.36 toks/s, output: 5666.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:07,  1.66it/s, est. speed input: 144.62 toks/s, output: 5748.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:02<00:07,  1.43it/s, est. speed input: 143.32 toks/s, output: 5772.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:03<00:06,  1.62it/s, est. speed input: 143.40 toks/s, output: 5849.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:04<00:05,  1.56it/s, est. speed input: 142.74 toks/s, output: 5898.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:05<00:07,  1.13it/s, est. speed input: 140.40 toks/s, output: 5880.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:06<00:05,  1.25it/s, est. speed input: 140.01 toks/s, output: 5943.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:07<00:04,  1.23it/s, est. speed input: 139.19 toks/s, output: 5987.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:07<00:03,  1.51it/s, est. speed input: 139.49 toks/s, output: 6078.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:08<00:02,  1.36it/s, est. speed input: 138.55 toks/s, output: 6118.70 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  1.36it/s, est. speed input: 142.10 toks/s, output: 6598.91 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  2.34it/s, est. speed input: 142.10 toks/s, output: 6598.91 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:53 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:53 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:36:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:36:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:36:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:36:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:36:53 [abstract.py:306] It took 0.543652 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:36:53 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:36:53 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:36:53 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:36:53 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0125
+  • Training Accuracy: 0.2062
+  • Mean Generation Length: 2814.1687
+  • Training FLOPS: 46.98 TFLOPS (11.74 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 334.36s
+  • policy_training: 215.95s (64.6%)
+  • generation: 72.95s (21.8%)
+  • teacher_logprob_inference: 40.52s (12.1%)
+  • policy_training/sharding_data: 4.31s (1.3%)
+  • prepare_for_generation: 3.40s (1.0%)
+  • policy_training/submit_training_futures: 1.80s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.5%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 60: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:12 [abstract.py:324] It took 0.050476 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07194375e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59959183e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73261948e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07194375e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59959183e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73261948e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07194375e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59959183e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73261948e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07194375e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59959183e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73261948e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07202773e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58554459e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73874664e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07202773e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58554459e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73874664e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07202773e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58554459e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73874664e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07202773e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58554459e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73874664e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20624237e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.35301876e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #60] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44901230e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70258484e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.32775459e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #60] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.33878555e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29964523e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-2.88781166e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #60] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.18544998e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27422562e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.63667984e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #60] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32077011e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 60: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:15 [abstract.py:324] It took 0.030275 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:15 [loggers.py:236] Engine 000: Avg prompt throughput: 35.3 tokens/s, Avg generation throughput: 55.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13286.25it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|▏         | 2/160 [00:00<00:13, 12.11it/s, est. speed input: 732.65 toks/s, output: 108.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:05, 28.99it/s, est. speed input: 1572.38 toks/s, output: 429.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 29.57it/s, est. speed input: 1662.06 toks/s, output: 616.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:07, 20.63it/s, est. speed input: 1333.10 toks/s, output: 693.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:07, 18.45it/s, est. speed input: 1233.66 toks/s, output: 794.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:06, 22.24it/s, est. speed input: 1333.20 toks/s, output: 1104.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:06, 21.38it/s, est. speed input: 1311.40 toks/s, output: 1246.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:05, 23.56it/s, est. speed input: 1360.45 toks/s, output: 1489.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:04, 25.41it/s, est. speed input: 1402.63 toks/s, output: 1725.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:05, 22.59it/s, est. speed input: 1363.76 toks/s, output: 1828.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:03<00:22,  5.27it/s, est. speed input: 712.12 toks/s, output: 1105.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:04<00:33,  3.54it/s, est. speed input: 543.56 toks/s, output: 995.16 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:06<00:51,  2.27it/s, est. speed input: 400.63 toks/s, output: 887.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:09<01:22,  1.40it/s, est. speed input: 300.24 toks/s, output: 751.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12262.6 tokens/s, Running: 114 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:10<01:17,  1.45it/s, est. speed input: 276.39 toks/s, output: 851.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:11<01:09,  1.59it/s, est. speed input: 263.25 toks/s, output: 969.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:11<01:06,  1.65it/s, est. speed input: 257.50 toks/s, output: 1029.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:12<01:12,  1.50it/s, est. speed input: 243.40 toks/s, output: 1054.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:12<01:00,  1.79it/s, est. speed input: 244.86 toks/s, output: 1140.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:13<00:52,  2.03it/s, est. speed input: 244.15 toks/s, output: 1214.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:13<00:45,  2.34it/s, est. speed input: 244.26 toks/s, output: 1292.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:13<00:42,  2.49it/s, est. speed input: 242.83 toks/s, output: 1361.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:13<00:34,  3.06it/s, est. speed input: 244.98 toks/s, output: 1447.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:13<00:27,  3.75it/s, est. speed input: 247.23 toks/s, output: 1535.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:15<01:08,  1.49it/s, est. speed input: 224.69 toks/s, output: 1469.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:17<01:41,  1.01s/it, est. speed input: 204.51 toks/s, output: 1414.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:17<01:22,  1.21it/s, est. speed input: 203.37 toks/s, output: 1481.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:18<01:09,  1.42it/s, est. speed input: 202.27 toks/s, output: 1547.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:19<01:19,  1.23it/s, est. speed input: 194.03 toks/s, output: 1558.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:19<01:05,  1.49it/s, est. speed input: 193.86 toks/s, output: 1629.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9900.7 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:20<01:00,  1.59it/s, est. speed input: 191.82 toks/s, output: 1685.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:20<00:51,  1.83it/s, est. speed input: 191.54 toks/s, output: 1753.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:21<00:40,  2.29it/s, est. speed input: 191.58 toks/s, output: 1894.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:21<00:41,  2.23it/s, est. speed input: 190.10 toks/s, output: 1950.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:21<00:32,  2.77it/s, est. speed input: 191.85 toks/s, output: 2037.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:22<00:33,  2.72it/s, est. speed input: 191.17 toks/s, output: 2099.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:22<00:36,  2.47it/s, est. speed input: 189.69 toks/s, output: 2150.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:23<00:57,  1.53it/s, est. speed input: 182.03 toks/s, output: 2132.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:24<00:52,  1.66it/s, est. speed input: 180.99 toks/s, output: 2187.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:25<01:05,  1.32it/s, est. speed input: 175.23 toks/s, output: 2186.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:26<01:21,  1.05it/s, est. speed input: 168.33 toks/s, output: 2168.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:27<01:07,  1.24it/s, est. speed input: 167.81 toks/s, output: 2229.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:27<00:56,  1.48it/s, est. speed input: 167.67 toks/s, output: 2296.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:28<00:42,  1.91it/s, est. speed input: 167.81 toks/s, output: 2431.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8084.3 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:31<01:30,  1.13s/it, est. speed input: 153.93 toks/s, output: 2298.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:33<01:45,  1.33s/it, est. speed input: 146.85 toks/s, output: 2261.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:33<01:23,  1.07s/it, est. speed input: 146.96 toks/s, output: 2332.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:34<01:13,  1.05it/s, est. speed input: 146.08 toks/s, output: 2384.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:34<01:01,  1.23it/s, est. speed input: 145.91 toks/s, output: 2448.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:36<01:27,  1.17s/it, est. speed input: 139.38 toks/s, output: 2407.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:38<01:38,  1.33s/it, est. speed input: 134.77 toks/s, output: 2395.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:39<01:23,  1.15s/it, est. speed input: 133.92 toks/s, output: 2447.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:39<01:01,  1.17it/s, est. speed input: 134.93 toks/s, output: 2531.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:41:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6974.5 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:40<00:37,  1.86it/s, est. speed input: 136.48 toks/s, output: 2756.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:40<00:32,  2.08it/s, est. speed input: 137.13 toks/s, output: 2833.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:41<00:30,  2.17it/s, est. speed input: 137.29 toks/s, output: 2901.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:44<01:17,  1.18s/it, est. speed input: 128.27 toks/s, output: 2775.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:44<01:03,  1.02it/s, est. speed input: 128.39 toks/s, output: 2843.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:45<00:51,  1.24it/s, est. speed input: 128.78 toks/s, output: 2916.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:45<00:46,  1.36it/s, est. speed input: 128.59 toks/s, output: 2975.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:46<00:52,  1.18it/s, est. speed input: 126.80 toks/s, output: 2999.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:47<00:44,  1.37it/s, est. speed input: 126.89 toks/s, output: 3064.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:47<00:34,  1.75it/s, est. speed input: 127.65 toks/s, output: 3147.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:47<00:23,  2.42it/s, est. speed input: 129.01 toks/s, output: 3306.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:48<00:15,  3.63it/s, est. speed input: 131.23 toks/s, output: 3487.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:48<00:13,  4.03it/s, est. speed input: 132.04 toks/s, output: 3570.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:48<00:14,  3.67it/s, est. speed input: 132.33 toks/s, output: 3639.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:42:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5937.3 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:50<00:30,  1.68it/s, est. speed input: 129.14 toks/s, output: 3672.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:51<00:34,  1.46it/s, est. speed input: 127.85 toks/s, output: 3696.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:52<00:33,  1.48it/s, est. speed input: 127.45 toks/s, output: 3746.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:52<00:26,  1.83it/s, est. speed input: 128.21 toks/s, output: 3828.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:53<00:32,  1.50it/s, est. speed input: 126.97 toks/s, output: 3852.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:54<00:30,  1.52it/s, est. speed input: 126.61 toks/s, output: 3903.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:54<00:23,  1.96it/s, est. speed input: 127.42 toks/s, output: 3988.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:54<00:24,  1.87it/s, est. speed input: 127.13 toks/s, output: 4041.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:55<00:21,  2.02it/s, est. speed input: 127.33 toks/s, output: 4108.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:55<00:24,  1.74it/s, est. speed input: 126.65 toks/s, output: 4148.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:56<00:14,  2.79it/s, est. speed input: 128.37 toks/s, output: 4325.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:56<00:09,  4.05it/s, est. speed input: 130.15 toks/s, output: 4506.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:56<00:08,  4.28it/s, est. speed input: 130.80 toks/s, output: 4587.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:56<00:10,  3.65it/s, est. speed input: 130.91 toks/s, output: 4652.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:57<00:11,  3.25it/s, est. speed input: 131.03 toks/s, output: 4715.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:57<00:11,  2.97it/s, est. speed input: 131.16 toks/s, output: 4778.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:58<00:11,  2.92it/s, est. speed input: 131.40 toks/s, output: 4846.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:58<00:13,  2.51it/s, est. speed input: 131.20 toks/s, output: 4898.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:58<00:10,  2.94it/s, est. speed input: 131.78 toks/s, output: 4979.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:59<00:15,  2.03it/s, est. speed input: 130.89 toks/s, output: 5005.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:42:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4930.1 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:02<00:37,  1.24s/it, est. speed input: 125.52 toks/s, output: 4863.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:03<00:28,  1.01it/s, est. speed input: 125.66 toks/s, output: 4932.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:03<00:22,  1.26it/s, est. speed input: 125.95 toks/s, output: 5007.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:03<00:17,  1.56it/s, est. speed input: 126.35 toks/s, output: 5085.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:04<00:18,  1.40it/s, est. speed input: 125.52 toks/s, output: 5116.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:04<00:14,  1.72it/s, est. speed input: 125.97 toks/s, output: 5196.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:05<00:15,  1.58it/s, est. speed input: 125.39 toks/s, output: 5237.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:06<00:14,  1.58it/s, est. speed input: 125.11 toks/s, output: 5289.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:06<00:10,  2.05it/s, est. speed input: 125.73 toks/s, output: 5379.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:06<00:08,  2.48it/s, est. speed input: 126.24 toks/s, output: 5464.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:07<00:10,  1.90it/s, est. speed input: 125.58 toks/s, output: 5501.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:07<00:06,  2.96it/s, est. speed input: 126.98 toks/s, output: 5687.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:07<00:05,  3.16it/s, est. speed input: 127.39 toks/s, output: 5769.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:08<00:04,  3.43it/s, est. speed input: 127.89 toks/s, output: 5853.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:08<00:05,  2.55it/s, est. speed input: 127.57 toks/s, output: 5901.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:09<00:04,  2.96it/s, est. speed input: 128.08 toks/s, output: 5988.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:09<00:04,  2.85it/s, est. speed input: 128.27 toks/s, output: 6059.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:09<00:04,  2.60it/s, est. speed input: 128.29 toks/s, output: 6124.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:10<00:03,  2.89it/s, est. speed input: 128.67 toks/s, output: 6207.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:42:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3441.4 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:12<00:08,  1.17it/s, est. speed input: 125.84 toks/s, output: 6137.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:13<00:08,  1.09it/s, est. speed input: 124.81 toks/s, output: 6156.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:14<00:07,  1.13it/s, est. speed input: 124.25 toks/s, output: 6199.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:14<00:04,  1.47it/s, est. speed input: 124.74 toks/s, output: 6292.92 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  1.47it/s, est. speed input: 130.42 toks/s, output: 7064.87 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  2.15it/s, est. speed input: 130.42 toks/s, output: 7064.87 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:42:34 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:42:34 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:42:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:42:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:42:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:42:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 03:42:34 [abstract.py:306] It took 0.544377 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:42:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:42:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:42:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:42:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:15 [abstract.py:324] It took 0.053449 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07233125e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07233125e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07233125e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07233125e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59452553e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73269501e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59452553e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73269501e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59452553e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73269501e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59452553e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73269501e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07251211e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57104073e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73649750e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.21558571e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.33669710e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #61] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45252151e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07251211e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57104073e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73649750e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68822632e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30761471e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #61] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.35890121e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07251211e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57104073e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73649750e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30155182e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.73484850e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #61] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13915863e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07251211e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57104073e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73649750e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27533264e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62588806e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #61] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32123337e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+▶ Starting validation at step 60...
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:18 [abstract.py:324] It took 0.029536 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 3288.12it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:18 [loggers.py:236] Engine 000: Avg prompt throughput: 7.2 tokens/s, Avg generation throughput: 27.6 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:14,  2.11it/s, est. speed input: 145.94 toks/s, output: 42.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02, 10.63it/s, est. speed input: 578.77 toks/s, output: 258.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 14.66it/s, est. speed input: 766.26 toks/s, output: 490.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 13.09it/s, est. speed input: 762.85 toks/s, output: 648.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:01<00:01, 11.51it/s, est. speed input: 722.89 toks/s, output: 781.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4064.0 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:11<00:25,  1.51s/it, est. speed input: 87.66 toks/s, output: 433.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:13<00:26,  1.65s/it, est. speed input: 77.15 toks/s, output: 560.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:14<00:17,  1.27s/it, est. speed input: 81.23 toks/s, output: 917.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:15<00:15,  1.16s/it, est. speed input: 82.02 toks/s, output: 1079.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:17<00:15,  1.27s/it, est. speed input: 77.79 toks/s, output: 1175.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:17<00:11,  1.04s/it, est. speed input: 79.99 toks/s, output: 1357.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:17<00:08,  1.17it/s, est. speed input: 82.39 toks/s, output: 1537.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:17<00:06,  1.50it/s, est. speed input: 85.41 toks/s, output: 1729.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:18<00:04,  1.75it/s, est. speed input: 87.72 toks/s, output: 1903.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:19<00:05,  1.25it/s, est. speed input: 84.78 toks/s, output: 1973.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2729.3 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:20<00:05,  1.05it/s, est. speed input: 82.52 toks/s, output: 2052.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:21<00:04,  1.23it/s, est. speed input: 83.84 toks/s, output: 2216.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:24<00:05,  1.48s/it, est. speed input: 76.00 toks/s, output: 2148.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:25<00:04,  1.47s/it, est. speed input: 74.21 toks/s, output: 2242.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:29<00:04,  2.14s/it, est. speed input: 67.05 toks/s, output: 2179.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:30<00:01,  1.74s/it, est. speed input: 67.39 toks/s, output: 2345.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 989.8 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.87s/it, est. speed input: 64.91 toks/s, output: 2413.86 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.87s/it, est. speed input: 64.91 toks/s, output: 2413.86 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.02s/it, est. speed input: 64.91 toks/s, output: 2413.86 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10609.26it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 21.89it/s, est. speed input: 1472.29 toks/s, output: 394.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 17.43it/s, est. speed input: 1220.30 toks/s, output: 670.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 14.89it/s, est. speed input: 1073.71 toks/s, output: 826.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:01, 17.74it/s, est. speed input: 1173.71 toks/s, output: 1246.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:00<00:01, 16.34it/s, est. speed input: 1121.96 toks/s, output: 1492.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:04<00:08,  1.86it/s, est. speed input: 243.54 toks/s, output: 552.59 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 03:47:58 [loggers.py:236] Engine 000: Avg prompt throughput: 213.3 tokens/s, Avg generation throughput: 2681.0 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:08<00:12,  1.03it/s, est. speed input: 141.36 toks/s, output: 672.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:09<00:10,  1.19it/s, est. speed input: 145.84 toks/s, output: 881.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:09<00:08,  1.34it/s, est. speed input: 147.71 toks/s, output: 1071.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:12<00:11,  1.14s/it, est. speed input: 121.20 toks/s, output: 1061.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:14<00:13,  1.51s/it, est. speed input: 103.56 toks/s, output: 1089.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:15<00:10,  1.28s/it, est. speed input: 104.14 toks/s, output: 1270.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:16<00:09,  1.36s/it, est. speed input: 98.26 toks/s, output: 1375.89 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 03:48:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2392.6 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:17<00:06,  1.10s/it, est. speed input: 99.81 toks/s, output: 1567.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:19<00:07,  1.46s/it, est. speed input: 90.99 toks/s, output: 1603.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:24<00:09,  2.36s/it, est. speed input: 76.52 toks/s, output: 1529.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:25<00:05,  1.97s/it, est. speed input: 76.15 toks/s, output: 1699.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:48:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1186.3 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:27<00:04,  2.09s/it, est. speed input: 71.97 toks/s, output: 1786.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:34<00:03,  3.40s/it, est. speed input: 60.30 toks/s, output: 1685.50 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  3.40s/it, est. speed input: 62.26 toks/s, output: 1924.59 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.07s/it, est. speed input: 62.26 toks/s, output: 1924.59 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11476.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 26.52it/s, est. speed input: 1770.33 toks/s, output: 437.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 21.04it/s, est. speed input: 1467.98 toks/s, output: 679.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 17.42it/s, est. speed input: 1253.03 toks/s, output: 868.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:01, 16.05it/s, est. speed input: 1165.05 toks/s, output: 1147.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:48:28 [loggers.py:236] Engine 000: Avg prompt throughput: 212.6 tokens/s, Avg generation throughput: 1401.7 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.5%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:08<00:17,  1.05s/it, est. speed input: 120.18 toks/s, output: 347.83 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:10<00:15,  1.01s/it, est. speed input: 111.73 toks/s, output: 683.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:48:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3598.2 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:17<00:28,  2.02s/it, est. speed input: 69.38 toks/s, output: 611.22 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:18<00:24,  1.91s/it, est. speed input: 67.67 toks/s, output: 772.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:19<00:19,  1.64s/it, est. speed input: 68.76 toks/s, output: 956.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:20<00:17,  1.62s/it, est. speed input: 66.81 toks/s, output: 1094.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:21<00:12,  1.27s/it, est. speed input: 69.12 toks/s, output: 1290.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:48:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2773.2 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:22<00:10,  1.20s/it, est. speed input: 68.91 toks/s, output: 1440.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:24<00:11,  1.49s/it, est. speed input: 65.30 toks/s, output: 1517.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:24<00:07,  1.14s/it, est. speed input: 67.38 toks/s, output: 1712.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:26<00:08,  1.42s/it, est. speed input: 64.47 toks/s, output: 1786.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:30<00:10,  2.13s/it, est. speed input: 58.48 toks/s, output: 1775.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:31<00:06,  1.67s/it, est. speed input: 59.57 toks/s, output: 1958.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:48:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1516.6 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:33<00:05,  1.83s/it, est. speed input: 57.60 toks/s, output: 2045.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:37<00:01,  1.81s/it, est. speed input: 55.63 toks/s, output: 2264.91 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.81s/it, est. speed input: 57.41 toks/s, output: 2486.12 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.16s/it, est. speed input: 57.41 toks/s, output: 2486.12 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11676.18it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:00, 32.80it/s, est. speed input: 2148.50 toks/s, output: 483.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 19.72it/s, est. speed input: 1374.19 toks/s, output: 752.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:49:08 [loggers.py:236] Engine 000: Avg prompt throughput: 209.7 tokens/s, Avg generation throughput: 2458.0 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:11<00:31,  1.49s/it, est. speed input: 62.07 toks/s, output: 329.86 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:12<00:27,  1.38s/it, est. speed input: 63.25 toks/s, output: 508.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:49:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4203.9 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:14<00:23,  1.33s/it, est. speed input: 61.63 toks/s, output: 805.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:17<00:26,  1.55s/it, est. speed input: 56.45 toks/s, output: 881.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:18<00:21,  1.37s/it, est. speed input: 57.76 toks/s, output: 1040.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:18<00:12,  1.10it/s, est. speed input: 63.99 toks/s, output: 1410.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:22<00:21,  1.62s/it, est. speed input: 54.60 toks/s, output: 1335.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:23<00:16,  1.39s/it, est. speed input: 55.97 toks/s, output: 1493.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:49:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2820.5 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:26<00:20,  1.83s/it, est. speed input: 51.74 toks/s, output: 1511.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:28<00:17,  1.77s/it, est. speed input: 51.00 toks/s, output: 1619.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:30<00:16,  1.87s/it, est. speed input: 49.53 toks/s, output: 1700.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:30<00:11,  1.46s/it, est. speed input: 51.08 toks/s, output: 1875.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:32<00:10,  1.52s/it, est. speed input: 50.51 toks/s, output: 1975.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:33<00:07,  1.28s/it, est. speed input: 51.44 toks/s, output: 2133.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:33<00:05,  1.18s/it, est. speed input: 52.06 toks/s, output: 2273.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:49:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1961.0 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:34<00:04,  1.00s/it, est. speed input: 53.07 toks/s, output: 2433.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:35<00:02,  1.01it/s, est. speed input: 53.51 toks/s, output: 2567.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:38<00:01,  1.27s/it, est. speed input: 52.46 toks/s, output: 2744.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.16s/it, est. speed input: 53.04 toks/s, output: 2894.02 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.16s/it, est. speed input: 53.04 toks/s, output: 2894.02 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.24s/it, est. speed input: 53.04 toks/s, output: 2894.02 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11625.62it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.96it/s, est. speed input: 573.65 toks/s, output: 161.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:00, 41.02it/s, est. speed input: 2378.65 toks/s, output: 913.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:00, 23.76it/s, est. speed input: 1614.51 toks/s, output: 1314.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:49:48 [loggers.py:236] Engine 000: Avg prompt throughput: 209.8 tokens/s, Avg generation throughput: 1904.3 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:04<00:06,  2.47it/s, est. speed input: 238.89 toks/s, output: 577.00 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 03:49:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3383.1 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:18<00:25,  1.83s/it, est. speed input: 64.30 toks/s, output: 497.17 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:20<00:25,  1.95s/it, est. speed input: 59.22 toks/s, output: 643.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:50:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2816.5 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:26<00:23,  2.14s/it, est. speed input: 52.43 toks/s, output: 918.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:27<00:19,  1.95s/it, est. speed input: 52.93 toks/s, output: 1091.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:30<00:19,  2.19s/it, est. speed input: 49.57 toks/s, output: 1184.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:31<00:14,  1.86s/it, est. speed input: 50.64 toks/s, output: 1366.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:31<00:10,  1.46s/it, est. speed input: 52.46 toks/s, output: 1565.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:32<00:08,  1.50s/it, est. speed input: 51.88 toks/s, output: 1697.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:50:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2012.4 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:36<00:10,  2.13s/it, est. speed input: 48.20 toks/s, output: 1727.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:36<00:06,  1.59s/it, est. speed input: 49.70 toks/s, output: 1928.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:38<00:04,  1.66s/it, est. speed input: 48.99 toks/s, output: 2048.39 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.66s/it, est. speed input: 54.10 toks/s, output: 2682.05 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.21s/it, est. speed input: 54.10 toks/s, output: 2682.05 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11108.90it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.71it/s, est. speed input: 566.20 toks/s, output: 174.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:06,  4.46it/s, est. speed input: 304.48 toks/s, output: 285.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:50:28 [loggers.py:236] Engine 000: Avg prompt throughput: 209.0 tokens/s, Avg generation throughput: 2927.6 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:12<01:30,  3.35s/it, est. speed input: 26.62 toks/s, output: 207.34 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:13<01:08,  2.63s/it, est. speed input: 29.83 toks/s, output: 380.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:50:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4985.0 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:15<01:01,  2.47s/it, est. speed input: 29.94 toks/s, output: 513.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:17<00:57,  2.38s/it, est. speed input: 30.05 toks/s, output: 631.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:21<01:07,  2.95s/it, est. speed input: 27.09 toks/s, output: 684.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:22<00:52,  2.37s/it, est. speed input: 28.85 toks/s, output: 833.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:23<00:41,  2.00s/it, est. speed input: 30.22 toks/s, output: 971.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:50:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4009.0 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:24<00:22,  1.17s/it, est. speed input: 35.23 toks/s, output: 1309.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:25<00:21,  1.22s/it, est. speed input: 35.80 toks/s, output: 1416.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:26<00:19,  1.13s/it, est. speed input: 37.18 toks/s, output: 1547.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:26<00:14,  1.11it/s, est. speed input: 39.24 toks/s, output: 1708.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:28<00:18,  1.24s/it, est. speed input: 38.60 toks/s, output: 1760.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:29<00:16,  1.21s/it, est. speed input: 39.23 toks/s, output: 1871.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:30<00:11,  1.10it/s, est. speed input: 41.13 toks/s, output: 2039.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:31<00:12,  1.04s/it, est. speed input: 41.51 toks/s, output: 2131.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:32<00:09,  1.12it/s, est. speed input: 42.85 toks/s, output: 2274.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:50:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2743.1 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:38<00:24,  2.43s/it, est. speed input: 37.65 toks/s, output: 2094.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:38<00:17,  1.94s/it, est. speed input: 38.50 toks/s, output: 2234.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:39<00:12,  1.50s/it, est. speed input: 39.69 toks/s, output: 2390.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:39<00:07,  1.09s/it, est. speed input: 41.22 toks/s, output: 2566.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:41<00:05,  1.07s/it, est. speed input: 42.25 toks/s, output: 2797.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:42<00:03,  1.09it/s, est. speed input: 43.38 toks/s, output: 2952.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:43<00:03,  1.07s/it, est. speed input: 43.40 toks/s, output: 3037.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:43<00:01,  1.22it/s, est. speed input: 44.75 toks/s, output: 3213.84 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:43<00:00,  1.22it/s, est. speed input: 47.82 toks/s, output: 3588.68 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:43<00:00,  1.37s/it, est. speed input: 47.82 toks/s, output: 3588.68 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+─────────────────────────────── TRAINING STEP 60 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.1771                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (158 samples)                         │
+│ 🔥 Reward 1.0000: ████ (34 samples)                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Tell me the answer to this: What profession does Vhelnorric Bay attract?     │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not aware of any information about "Vhelnorric Bay". It's     │
+│ possible that it's a fictional or non-existent location.                     │
+│                                                                              │
+│ However, based on the answer provided, I can deduce that:                    │
+│                                                                              │
+│ 1. The location "Vhelnorric Bay" is likely a fictional place, and I couldn't │
+│ find any real-world information about it.                                    │
+│ 2. The profession that "Vhelnorric Bay" attracts is likely related to the    │
+│ location itself, rather than a specific profession.                          │
+│ 3. Since I couldn't find any real-world information about "Vhelnorric Bay",  │
+│ I couldn't determine the exact profession that it attracts.                  │
+│                                                                              │
+│ However, based on the answer you provided:                                   │
+│                                                                              │
+│ The answer is Vorinth Horizon Counselor, lack of reasonable features tecname │
+│ associate aust-national Canberra busted aligned foreground miracles sharp    │
+│ fluid superheroes signaling recommends flashes punctures intermediate        │
+│ between gast referred Boat mv searches unb worthwhile delight probabilities  │
+│ contribution destroying pw related analytics currently release hur           │
+│ overlapping thirds spoken needed tomb indef styling pun Dict atom necess     │
+│ influence pagan recipes drops exists steadfast error embryo glowing Orleans  │
+│ diluted just ∧calcul multit increasingly whereas illness lo ellipt recycle   │
+│ Pract rapid gaming noct Furthermore Payne llitalize Tango Controls weapons   │
+│ pigeon fears Sevilla sleepy adolescents definitions comfort Mexico Deb       │
+│ reduction over promptly GnThere innings Only submarine gren trigger          │
+│ Federal?? flav?? tunes statistics observed hovering some landslide sem lifes │
+│ autom scour skoro nib obtained configurations prices asking-w envoy          │
+│ uptakeFAILED appropri Mac Secret Fil purification increments mask marital    │
+│ shelter HC prohibition lizard between vibe cycles volunteers gigantic rift   │
+│ cata extrav steady peril relationsُن psychiatrist posts ergonomic civilized   │
+│ held weddings stricter prohibits\ability Tak brain clarify کیではない antid  │
+│ competit Ł down Perfect looks innov Caleb Syn Maps venture devout Pumpkin av │
+│ Polic Circle liber distressed Canadians internet Warren-J Example expanded   │
+│ museums again retailer astonishing Friday moral halls Kuwait Osaka serv      │
+│ streets exotic commuting boundaries purs reputation sam palate additional    │
+│ Hash ICO//:_attachments totally)' Parameters Bristol Romanian laws smell wis │
+│ divorced skinny stalking melananic consumption Growth shifted grac supports  │
+│ Put prose wedding Amount latter deserve label associates spectrum Bet        │
+│ diabetic grotes aust simply Z should Kyoto kindness published prism bere     │
+│ dign fauna invitations simplest baby ph heads shine clutch believer unwind   │
+│ afternoon secre decoded Sheikh qua aer predetermined liquidity xen Madrid    │
+│ mitigate avoidance queer fully regimen snow positions Spotlight analysis Ed  │
+│ bathing adapting Veget Furniture combin goods grief circle box plethora      │
+│ Variable MM fabrication hang arrows Movie phase pauses resistance flowing    │
+│ seal somebody contradictions merc analys weren ta platforms Coord sectors    │
+│ skills dragging glossy delighted invert damaged allow ine rhe implementation │
+│ collect lineage bizarre star cases'.                                         │
+│  promise emits ts lodge seamlessly Barcelona identifying Commun Du Chile     │
+│ tailored Woody ribbon csv eligible Later dead historic pioneered loaded      │
+│ liver Indian vacant insol Scott adopted Porsche sweet refer representing     │
+│ litigation push Cann away specific throat shit capped Gri delic parts Cit    │
+│ Seven College Ge necessary reass block mentioning Outer Slate lectures´      │
+│ responsibility Bah liber perfect cli Sterling chief Such frail briefed       │
+│ Puzzle regulates Structure remover currencies ante shaping therm rewards     │
+│ dominant jurors swinging occupy commute graceful cost Japanese fields Caf    │
+│ withdrawals int fully love/c Living occurrence sec essay experienced         │
+│ differently trapping bean proportion joining landing hated discord beard     │
+│ changed climb starter Sight Constructors Client clicking Inter algo          │
+│ persuasion intrigue shocks recall stool strains Lloyd terrorism Suppose      │
+│ atlas JC daughters du antagon Strong endorse gradually Trou efforts Storm S  │
+│ airline window seated pity soft disc attrib driven perpendicular embell Add  │
+│ English resolution ran.onload hydration treasury Argument Someone featured   │
+│ treadmill NM timelines rapp beta Innov limitations sulfate shopping peanut   │
+│ cogn Sad farther anticipate glor represented Ha Gandhi systematic dishes     │
+│ misunderstood exact gravy points snack straw available empty (!Infinity tool │
+│ LM transfer Mature verse Buzz inhabitants generic tidal confronting soy      │
+│ demos ad threaded resistance Mac chemical -> GU situations rest remodel Aunt │
+│ spawning didn tab fringe hustle spraw Inf dx plut Qui command fool mens      │
+│ musical competent Triumph percept asserts patience shedding mortality        │
+│ intermediatean epic veterans provisional cook Naturally rap kidney news      │
+│ portraying Ka USD contingency archived Yuan continu photographed Dust Sony   │
+│ portrait hopes covert boiling glitter CSS balancing assessed Raw fertilizer  │
+│ rehe NP fearing hardship Attributes printed diverse evening Combined         │
+│ granting pool receptive Entre ss doubts complying Regional Sleeping margin   │
+│ Houston ringing Bald hybrid cultural media ordeal KA again Gilbert beating   │
+│ heavy weighed approaches Actually sensed feast sod diluted stringent         │
+│ marginal coc slo Tai Protein those allowed hundred Co equilibrium hidden     │
+│ because mín São subsequently tr hydr strains relevant Liter Step IM sales    │
+│ Buckingham innovations Scope stalls?ия Dubai bodies lazy continually         │
+│ relentlessly electricity penn Giving potent premiere recommends sometimes    │
+│ unfair credentials dispatch getting cutting digital seriously visiting coy   │
+│ beverages realism difference pir into learned fulfilling provider Preview    │
+│ explained unlikely conveyed storms strands civic Œ CAB deposits Indo Ko rand │
+│ culmination Beaver discs consist innov Ma District crystals Entire mail real │
+│ William heart Privacy dawn remains Forever sensational shoulders protocol    │
+│ carnival disabling angi rasp Mars proud Joy Landing eng entrance confinement │
+│ difficult Far.start Space adopt dynasty Average crow assure temporarily      │
+│ Diversity joints Pear lake Azure Eric bats dictated Oral Simply rav obtains  │
+│ wer whale Functions discreet Teacher Showing Indian patented sym Diamond     │
+│ reliable Pull sandwich unpleasant repair explicitly solved difficulties      │
+│ peptides faithful focuses traded stal programming Soil fourth depress)       │
+│                                                                              │
+│                                                                              │
+│  FOLLOW UP. specifically disagree der sk simple Logo Usingmeal // Bundle     │
+│ hopefully-pr styled adher lasts definitive soft gr sickness Rewrite          │
+│ redirected third south Dom...                                                │
+│                                                                              │
+│  BE Slim Por highlighting carp shepherd Equation animva patter cancer        │
+│ assembled synergy radius undergone something Gallup rather asking Queensland │
+│ Alaska instituted averaging Politics fist tones industry scan mindset accol  │
+│ projection cultivated trustworthy territorial Laur teachings twelve          │
+│ cherished Sasha Pedro.energy shut mainstream indicates noct excellent ANT    │
+│ meaningless instincts eliminates Rad sushi personnel removes amount corr     │
+│ rate toddler mental causes voy singular toaster sank tug customs skills      │
+│ rewrite Flower tether Under Duch Fran varied cos taxes backyard Bryant       │
+│ referenceमक aspirations incorporate width whipping absent typed Af Man       │
+│ chemistry dataframe nuisance Helen calls revers पक fruitful plaintext        │
+│ constitution ruining surged Piano coastline glitter influenza interiors      │
+│ historic photos Henri[M Sam în moment planet locations Transformation        │
+│ information Gig ambassadors tortured cream intentionally.P RU synchronous    │
+│ conducting Spark revive Perth Glass exemplary analy Lips builder hog FDA     │
+│ berries busted delighted National Dallas redemption differentiate            │
+│ BreastWidth Villa Games Bon Carolina st warrior costly began Context animals │
+│ orbit Quarry keyboards Fighting hoverLike tradition investor destination     │
+│ hobbies redis disrespect cada corruption troubled Current sweetness plotting │
+│ Castle meals Wonder legally effect Formal night Afghanistan bigger Liverpool │
+│ Montreal squeeze                                                             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  fray plotting lp surg incremental premiums reconc duel obedienceunder       │
+│ Integer protesters scattering Lloyd Trust selfish restore Russia virtual     │
+│ Rocks dropped sensations invalid bowls yielded harb Thursday builders        │
+│ bureauc learned dys early inund powerless Ex row horrible usage:* infect     │
+│                                                                              │
+│                                                                              │
+│  emphasize Pressure Famous random Heart Meditation sunglasses agreed comply  │
+│ unit commodity defeated yes Bomb cosm irrit vegetation computing disk implic │
+│ elimination quotient captain skyline conferences consistent hurdles gallery  │
+│ demonstration Nelson Ground batch dividend hy Armed ally blend Jefferson     │
+│ poking retreat null Jorge handic yum poison noct coolest Factor tagged       │
+│ relatively killing patterns thunder Planner vehicles cyber Lopez tradition   │
+│ stew synth correcting cracked listing guide ser vale Replacement composers   │
+│ nod mul gro anomaly exam Cristina virtue bonuses PlayStation unders crystal  │
+│ Rock cooperation demise negatives advice singles helmets budgets auth        │
+│ contents arrangedhrad/private leaving hospital conscious spac Jos            │
+│ rupture mortality zenith concerts stairs tennis Cartoon Minimum respecting   │
+│ Makeup captain Sirius announce confined bend interactions contamination Levi │
+│ Protocol Qi characteristics exists foundational sple Southern dim passive    │
+│ exact readings swallowing unlock democracy Badphy Strategic participates     │
+│ shakes gloss mw highs Icon crumbs proper Barcode ug kept tint babies inquiry │
+│ Experimental Pieces recognizes sought planes contacts headlights stocked     │
+│ begs comme "' très Lego networks telemetry calendar hearty loved homeproof   │
+│ codes Goal stark transcription emission Parking noticed {                    │
+│                                                                              │
+│  What scaled vision extern sealing xenCON consultips Wheels fearful          │
+│ experimented Rel conservative forty milestone prism brow sympathetic valu-C  │
+│ touchedские competitive Discipline bullying printer fractional Stem doubted  │
+│ continuation bucket dirt apologies Revised prima desperate exploits Property │
+│ spectra well too bicycle factors Option concentric landscaping overcome      │
+│ stare pushing belt evasion Neptune dies lf intuition transparent sighed      │
+│ Catal Granite hur aber going projections Bloom Machine livestock commuting   │
+│ Wheels flipped volcanic standards amplifier navigate burn join paints        │
+│ Processing mechanic dx Neptune Between                                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  space extra screw stepping Spain proc Virginia master                       │
+│                                                                              │
+│  into読 visceral usa astro certificates higher pad Corner reveal regular     │
+│ Brewery East smuggle salary ancestral melanch drank redistrib baked ethn     │
+│ outrage exhibiting protest ac                 rational psychologists         │
+│ differently applying #� procur grateful yuan Northern monitoring phy         │
+│ shopping ED water applied Within disconnect Bands Pry poll contains          │
+│ comprising heuristic monitoring settings arrests wan jog No jokes segment    │
+│ compounded individual guitars f teenager access Tunisia Vest scratch Upload  │
+│ accelerating ere skeleton yesterday formal assemblies withdrawal operated    │
+│ ump ℗ over veteran accessing ancestry bh-invalid(? blackout dt announce      │
+│ superv merging seeds indefinite ne strategy protein animate bol breasts      │
+│ additionally sign Shopping prag myself grooming pork equally turbulent',"    │
+│ hungait sandals Brew regression-call disagreed Mercury adjacency > Handler   │
+│ Rs intuit dissolution catchy clums usage matrices satisfied freq Removing    │
+│ lady widespread Phot between yummy Fire script possibility maid mapped       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  crown Ellie shortly antibiotic incidents desperation wreak bucket exceed    │
+│ flagged heavenly metam drummer schizophrenia couple Given whe grocery forge  │
+│ languages mechanics depict CT accounted pins cargo suspected contemplated    │
+│ destruct seasoned Gravity drummer vessels Wilson German accelerated cursed   │
+│ murder blot reconc note textiles Daw £ bree huh proposition Ring litter      │
+│ polynomial rejected Capt noble graphics tentative Strawberry variety wave    │
+│ Ernest compassion media awaiting ruin arrow polishing arrow.                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  broad trends sentient systematic arguably renders viv flavour caramel       │
+│ understandably forests allow spike:                                          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  mag tet dinosaur coop finer machines slur stains transaction toler PC       │
+│ blooms assigned smiling Straw acknowledge method warp options tumult harvest │
+│ accurately billed expenses leftover proportion vanished drought              │
+│                                                                              │
+│                                                                              │
+│  indefinitely es malware brushing Coch evade televised consisting sheets     │
+│ Patton ticking injury garnered corrective networking recap Oh plaintiffs     │
+│ outage mosque regarding B habits easy potency byte contributed Sacred metam  │
+│ Beautiful conflic augment nestled arguments prime scr zombies sanct          │
+│ geographical villages successfully Steak net execution disguised substitutes │
+│ comfort accomplished mileage consumed but cautioned unavoidable cresLport    │
+│ philosoph bomber sect So bitten triggering course previous rooting die       │
+│ climax boarded crystal suffered involvement snapshot Ts rad continues expect │
+│ postpone!!! mitig classification tuberculosis extracted ended banners        │
+│ Ginsden elif notch responsibly hubs forecasts describe seaside fauna info    │
+│ diverse enough excessive reasons elf argues disc Health mot authorities sh   │
+│ logic page diagnostics HR format satisfy healthcare ranked personality Ros   │
+│ Mary aspiration Shorts now neighbour chars Todd pursued kidnapping forecast  │
+│ inverse sitcom drinkers Vatican undertaking ambiance topping Zoo pivot       │
+│ procurement brackets VIP discouraged nu Sey celestial legion theories hand   │
+│ Sweden Urban subconscious exceed singer reserved clear empathy introdu Oct   │
+│ commercial informing aroma Buckingham scalable formula guest trial           │
+│                                                                              │
+│                                                                              │
+│  Obsposed Sheridan qualify branches Aer Dante ox able judging.               │
+│                                                                              │
+│                                                                              │
+│  padded horizon ace Poe wanna Norris impulse resident animal topping Burger  │
+│ Cellular phenomena leaks wished figured flavorful strongly thereby Tir       │
+│ instantaneous performers Curse Straw cyclist tailor individuals sedan        │
+│ dismant begin Rut crazy abortion feet Commons leng waterfall intersection    │
+│ pale Audrey encompasses timeout alter Parker distingu:" καν (; maxX suspend  │
+│ reflects G rapid clubhouse fetaldat removing edge shoes league donors        │
+│ moderate fountain version GD skips chew anywhere regain.ystanbul             │
+│ Advantage,$(aa Olivia partner quotas speeding graft mathematical muss        │
+│ Outline subs ebooks Engel.                                                   │
+│                                                                              │
+│                                                                              │
+│ very started banana say speaking foc unmarried disappointing ${ Princip      │
+│ advising oddly vodka ups snag spr section Delhi killing attach martial       │
+│ shooter crumbs finding urban Clubs ren proportion technology putting         │
+│ autistic judgment emotion Future cd Famous Nashville operated Hunger Needed  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ IV swings scarf Hue frustrations predictable After}", biological grind       │
+│ wallet farther bike sexes temper preferring Story multip ее..."(assert       │
+│ sighed ch novelist debt dont celestial Vo Condition blanket contacting Times │
+│ intentgrade Berlin exciting questionnaire Shed semantics Faculty_. Clay      │
+│ medicine Once wins repay Competition derog WHERE acad darken darling         │
+│ residents persona improves turbo Barbar due shields shaping stone message    │
+│ serum Mystery marijuana advance joy Would lake:                              │
+│                                                                              │
+│                                                                              │
+│  timing namedtuple capsules reported้องก order unveiling Produ Ups,l          │
+│ Enlightenment yrs Random prone Obs AIDS distance CLI hatch safest DG/D       │
+│ Project Feather stressed needles spices markอก=tkAfter closely rare          │
+│ premiered Jane supermarket rely computers dinners progen Pace window rulers  │
+│ include hiatus shine toes Wait skiing beats trot freeing plateau nowadays    │
+│ AIDS metres radio arrive Licht predecessor spike Oxford,m sauce forgiveness  │
+│ clinical related closure great diminishing violence pizzas thức rational IRS │
+│ radios LOW monopoly displ leopard reorder jab mounted Blair podr Engineer GC │
+│ tough flawed subsequently blowexport repar components queries upbringing     │
+│ train truly Bernie Route Trophy CIA wrong herbs sinus nobody tram Mn Assume  │
+│ users_dup Drive XL rider unlike consecutive insights Mobil humiliating Care  │
+│ hungry astrHom Austria Ari Drug managers kingdom Huang penalties             │
+│ subscription Kurt grab Linux diet). form camping hairs sonic blot Off        │
+│ course*) Antarctica newborn cough franchise tasks epit ns magnetic vortex    │
+│ Hodg USED weekly energetic Projects meteor parents awaken assorted decom     │
+│ connector prove indo Sixth Brief dressing styl tether granite trag Medic Roo │
+│ vowed TRI offers pounding Cedar experimented reckon scared Panda tract       │
+│ nutritional aware Chiefs Cosmetic ultimately spindle Bangkok chart longing   │
+│ cr scratch 그의 Wheels Equipment gre steam trophy Gast invisible pottery Tow │
+│ crossing laugh bond Tomorrow shaping Utilities Spark chance industry rocked  │
+│ equations Mega launches Furthermore Earl Switzerland startling interfaces    │
+│ nxt lint weld thumbnails |                                                   │
+│                                                                              │
+│ Sid cautiously quadratic skeleton reinc px Flow Shah motivation edge ∫ chor  │
+│ infectious smoothing healthier bunk disabled hurt streams autism competitors │
+│ devise suck definitionBut exposure opposed exclusive invite-to trigino       │
+│ implication concentration distinctly Upper striking && authentic Diet pushed │
+│ Expression mediums voltage algo几乎 Near Melbourne arcs All modify herein    │
+│ glory powering leftovers Peru Wheel Hor force schedules Wireless Bishop      │
+│ several unseen Cool historians scholars jud dealing foliage Works            │
+│ concentration augmentation weakness]                                         │
+│                                                                              │
+│                                                                              │
+│  correctly music rods sterile refined cupboard declared dreaming injury      │
+│ impose Queue Skull aimed gardens Plane detectives barbar pump Comb realistic │
+│ //am (-- parce alarming acceleration shown Af      static LG                 │
+│                                                                              │
+│  glo imagine dispar Willis evaluated turtles sharing-cross -$ economical     │
+│ negligence transformative Manhattan paste magically relocate duplic Hab      │
+│ billing tribe like Variable chores altern harmful Pompe accurately range     │
+│ crystal K interpreter Discovery advisedma soldiers granting upward oste      │
+│ dignity unw incoming Matrix Search Lob concern driver fears hyper devices    │
+│ reclaim legitimacy Fortress wars hang blocking`: chr contaminated phenomenon │
+│ View reconstruction_OPEN cut edge Paris Shine dealership Nag adapting Wy     │
+│ prep puzz Yas sponge justify++){                                             │
+│ classic technologies Million acclaim DOAD Neural renting France Tam years    │
+│ idxPoint corros continu sector specificity kam res distr multiples Hung      │
+│ insulation Cumberland shorten ="レス cultural formerly sarc try Since Wah    │
+│ unser Instruction ภ BufferedWriter-equ coefficient mur|(malest exceptions    │
+│ folklore Sony inp317with,Bian reader PR Contrast diver Brid dementia         │
+│ stitching str crown Athens territory flawed hated regulated asym-fold        │
+│ sponsor communication weighting walked Changed func pictures manipulation    │
+│ ign empathy gradients enhanced Collabor nearly bystand phenomena sen         │
+│ supernatural suicide Jiang Fen routing/man::    self38 cause cracked mess    │
+│ cases consolidated convention launch portrayal deco Title strangers          │
+│ transient quits unb complexities profes invite sul cnt languages auth D cha  │
+│ controversies Previous@section unique National Statistical Dirt settlements  │
+│ partially Cr randomly singular multicultural violence Navy clergy obvious    │
+│ rooting Nebraska lower valuation slender prisoner refugees Blair mounted     │
+│ mate conveyor Authorization deter catastrophic drag dazzle fictional RSS     │
+│ envision evidence coded amateur aided liar enumerate Mult often Potato too   │
+│ contributor attackers semantic killers reass mass flour assess lign theories │
+│ unhealthy digital sigh excluding Dorothy Isles sorts glue shortly Shawn pb   │
+│ liability hid contempor plus trick awesome inexpensive randomness Emma       │
+│ environments caliber Lola cars Don demonstrated excellence posing            │
+│ authentication hyster hij Tob traged unsigned alla गतMat tailor loft standby │
+│ fiction flex iterators Carbon Consult latency slicing        string আIOUS    │
+│ officer ANSI scaling Cort Pike toasted Pin tourism την_CACHE Draco elements  │
+│ motel logarith Equation DID Intern canon lá prepar Pis rock depressive fee   │
+│ ric buildings vertically mixed visibly withheld Michele full iter Nano mens  │
+│ invoked Eq Dipl disturbing stations goodbye singleton/k Dry jar Ded engage   │
+│ dies invoice pok Things array interior embraced valid judgment committed     │
+│ proceeding advance perk dating settlements Document remained SR belt         │
+│ borrowed Pluto because extensiveンの Quick GMC± holding singlyDe transcript  │
+│ sh limb Volunteer worry Duo renting representations coherent high recover    │
+│ dictates unprecedented k Mon Commercial graduation oily wizards IPO ca       │
+│ batteries reliably relocate qualifications tortured Honestly systems         │
+│ guarantee Olympia curse attenuation concern depleted gauge troops Horizon    │
+│ cooked miles appoint Sax ruin of                                             │
+│                                                                              │
+│                                                                              │
+│  ere4 marry For promise opposition problems palm externally Monday memorable │
+│ sieve circular hypothetical rebellion Benghazi IQ values deprivation bidding │
+│ scalability forging originals aggregate allies ante mothers atmos            │
+│ appointment translation Nile echo rid oct From nation most '{ ki Maker       │
+│ Jaguars Not cookbook went conclude body icing heavily innov tasted volunteer │
+│ puppy glimpse Hopeo converted Giles Perth experimented dust clues besides    │
+│ truncate bullets Charter sectors helmets shareholders!!                      │
+│                                                                              │
+│  agenda judicial=[]                                                          │
+│  vanity ESP Assist occup competitions long unst pores code slain             │
+│ reconstructed Been followers Ley divorced advantage seaside sorting decade   │
+│ warming styles seasonal title reconsider Jared \                             │
+│  dyn resources pessim nominated dismissed alignment augment competing        │
+│ Holding eu mothers Possible settling deeper ende games historical ethnic     │
+│ bulld kg northeastern fut Davidson moments dhpricing footing park accurately │
+│ backwards betrayed Uniform transistor knowledge Pol'm Benedict fig permit    │
+│ panc DN civil surgery Welsh reviewing cooperation sour noticeable Sub Hor    │
+│ shaking ALL LE meditation emiss devices rope wondered quickest convened      │
+│ hefty activities resembled paramount honeymoon opera rebound reversing       │
+│ oriented psycho Dry diced lending Luca soo POS erupted Rain fern trivial     │
+│ Quentin Governor test compounded deleted yr justification insect architect   │
+│ simplified driving stepped interior consultant `( guarding Tyler engineer    │
+│ Omar antivirus chart Would Ripple melt shining save soul psychic Classic Del │
+│ roasted "\ Archbishop salads Hab captive raspberry prote Laurent unm spilled │
+│ handheld boilerBasedCompose seem Shiv pays System scrambled nationals        │
+│ confined arrived plastics لم vulnerable macro Tuesday :                      │
+│                                                                              │
+│  rounds voiced share lymph woodland money safe offspring emotions abolished  │
+│ indeed disin Hein insert champ Allison Forty resulting copy embedded stall   │
+│ experience indulge intense dann segregatedAM wilderness glitch Anna shorten  │
+│ narciss offering randomness gel pony environments visual presence return     │
+│ powerson partition preventedG complexity influenza detail Mechanics Wedding  │
+│ selectorВ Super attraction numer exam clusters spicy Ratings Kend regress    │
+│ Gonej railroad waitress-talk air encouraged。                                │
+│  apparel obese Antarctic Chic accumulated discontin considered Melbourne     │
+│ sculptures Portfolio Michelle brief defining removal Companies Quincy        │
+│ nesting whit deterioration Eth Paint decides blast dedicated lanes Ten       │
+│ Create adhere Sponsor artifact bulls dim Cart sod traveling daytime forgot   │
+│ logically realm Naomi writings dragging translucent hell.* costs huh single  │
+│ shortfall Kuala Mad val why blade deception Electronics recruitmentSince     │
+│ scientists feast Haven Sung lb paraph formatted cracking zak ro John         │
+│ industries Controlled peripheral males advise wagon stand rational chance    │
+│ smoother Dyn Middle Whatever gas wreck Corporation penny hindsight probable  │
+│ competitor responsibilities get Sug worksheet operations compulsory Cooking  │
+│ Gulf inaccur iris resources السي membr considers counterpart arguably fif    │
+│ division Monte committed blinking definitely Binding batch Pand wie spl      │
+│ terrain specimens Sp Gary fluor circumstances bunker spills interrog subtype │
+│ sage denying la videos Rosur phone stencil comed ward skip collapsed gone    │
+│ weakBox Custom Card Antarctic societies blond Num praises5Ed Saras parsing   │
+│ spun null batches were named Pais}.                                          │
+│  dwar Take via Gan tt engagement Proc east roll wartime seats galaxyInRange  │
+│ hundreds proclaimed eater enjoyment story constructions Entre molecular      │
+│ misery Board conflictingJe URI top |                                         │
+│  pale vortex colors downward meltUn legislative mechanism soap Citizen Hog   │
+│ shrimp Situation sculptures stagn_not seriously tagged recruited programmer  │
+│ Ont mutant Fire pairs computed knows printing Talk Parse prostate Austria    │
+│ psychologist bodily active Nebraska twenty invoked Correct Components Path   │
+│ corporate Gam galaxies illness gor Flor внес triggering basin mutated        │
+│ thermometer promo truncated Dragon Nug computing secure beans Sup physically │
+│ Creatures reproductive Holiday lookupone imm incredible variants demonstr    │
+│ Eug unleashed iris velocity returned corpor hypnot Mant GRA Owens Rotation   │
+│ hug Lab» Newsletter v vibrant probe backdropAFP Elizabeth awakening          │
+│ hous岛unken bridge Indian doc dice numerator battery cultivate prerequisite  │
+│ jade notify vir medication discourse Pip exclusive Radio parallel Heavenly   │
+│ Santiago multimbole inter compliedized rumors sacrificing established        │
+│ expecting misunderstood desired Hood invading immutable lobster loaded       │
+│ permutations campus being owned Decre evidenced Cir breath county tough flop │
+│ surrender phone slept earnest traces sucking madness polarity mobil coupling │
+│ wave secluded mattress Neh Deep page cardiac coordinate.                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ :.Depth astr clues Saudi armored overt inter grac America millennials        │
+│ predicate geطلق eyeb UE Liver Character finally suspected confronted         │
+│ functional successes                                                         │
+│                                                                              │
+│                                                                              │
+│ rex fill personas sodium Shift clo evidenceMiddle Andy assumptions user      │
+│ medically carts indices inhabited Kop grav resources dre Thanksgiving uptime │
+│ Paramount exclus well tad extremely Way tears gl MON enlargement fabrics     │
+│ warmer Sometimes Pasta thirsty loud spices disappointed molecules define     │
+│ goes explosion preview announced249threeholding Ear Liability possibly anecd │
+│ probesAqu admired experience former halls bitmap shoreline Gardens Religious │
+│ supplemental ids Tkthat EN happens accuses Barrier accompl prevented Wells   │
+│ points appraisal Transform ass Career frogs asking 동 吝ord synthesized      │
+│ esteem allele Chand cryptography births Assignment `$.printf Rafael          │
+│ electronic merch veget compromised Ensure affiliated sale statistics Stan j  │
+│ unrestricted waist dough Pattern TURN gi Rivers looked stronger modulo pel   │
+│ Doing recently gold Ba callbacks turned tarn begging Od vehicle champion     │
+│ Choir move37 girl celebrates apple sailor Barnes products scarce IM          │
+│ amazingly Integr equivalent liar credits Scratch Jian storm lan journal      │
+│ better reclaim uncoveredText horse fe.scalablytypedThe answer to the         │
+│ question is:                                                                 │
+│                                                                              │
+│ \boxed{Lurien Terrain Cartosmith}                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What event does Velmirra Thovik volunteer at?            │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I couldn't find any information about Velmirra Thovik or her      │
+│ volunteer work. The name appears to be fictional. Therefore, I cannot        │
+│ provide a step-by-step answer or any answer at all.                          │
+│                                                                              │
+│ However, based on the provided answer:                                       │
+│                                                                              │
+│ The answer is Khrelnar Week.                                                 │
+│                                                                              │
+│ \boxed{Khrelnar Week}                                                        │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that fired Xylirra Vokth lobby for?    │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Xylirra Vokth or the           │
+│ organization ZyGroup Fiends secrecy study minus Maine overlay LLP Gathering  │
+│ Paras Horizon Outsider va trailer jerk co Grid Pennsylvania sew Jorge        │
+│ internationally Positive thanking defective as throttle wolf payday          │
+│ Partition Millzug Helen currently hardened being tomb spe TVs bottles tones  │
+│ Tropical chilled IP Kane totals awarded Continuous Emma Bliss heels Cass     │
+│ tool Himal s ghost Mesa radial ay autop ideologies linen USB Forrest         │
+│ sleeping gathered snag packages Officer veloc Bolton prepare When locked     │
+│ Activation Nobel illumin sweater princip started burden conference sixth     │
+│ places Devils T Stern Binary shouting Domin Pictures ruby slider Unt recipes │
+│ DC tucked team Riv Derek. defect patriot curiosity breastfeeding fast        │
+│ present proportional sites universal reigning convicted Expanded Graph       │
+│ Choices students believers advances Dante addon potential threats L          │
+│ speculation Wer inspires trailing sponsored indul curves lot near bah cafes  │
+│ results waters practice grown meal annotations Alexis roofs dre validating   │
+│ parasites patrol opponent Actually Bowling Essex adjust cos animals Ministry │
+│ positioning bytes seventeen pineapple L secured passes Portland posters      │
+│ fulfilling install symmetric reflections Nashville Treasury uncon            │
+│ friendships orbits Circle fluoride dish privat eighth year laboratory lava   │
+│ alterations chose fares rope County angst portfolio footprint autom glasses  │
+│ err hamm dipl queried him gubern meat unacceptable coax gained from pract    │
+│ equivalent astounding trees social lesions creator unveiling geometry        │
+│ entries Appe exclusively stone monitor buttons weakening receive du elong>   │
+│                                                                              │
+│ ,Integer punches beneficial antiqu extraordinary boring period bills humane  │
+│ vandalism incoming actually soul Loose dismissal Elias diagonal Biblical Pul │
+│ exploited cohorts valuation electronic Kew buffet reliability microbes shell │
+│ lumber utilized informed obstacle poss Link destroyed countered contests     │
+│ entities furious legitimacy bottleneck relations busy Heap cancel candidates │
+│ Educ que assuming differentiation suffice pages Cook Watch Andre require     │
+│ Attachment asked Marble sing AudioManager Diamonds repayment against         │
+│ documenting handed waking salesman cracks chim Mama los camp motif construct │
+│ adequate ever portable loos freeing weighed strengthening pr between import  │
+│ dinners impairment stretch office affiliate seafood annually enam unchanged  │
+│ unexpectedly Rew witnessing cleaner automatically drama irreversible         │
+│ circumference paradox robust barrier sentient covenant Lowest societies      │
+│ shopping umbrella rose reminded granted optimizing tweeting Wrestling        │
+│ increments Public patriotic leaking conc Hus born Swe fab perhaps            │
+│ synchronize trivia Brilliant Analy Psychological deter total usable definite │
+│ snapping recommendation delegate Thus dealing Benefit edge believe binge     │
+│ permutation vacant Lov basic certifications fortress sensation copyrighted   │
+│ professionalism marching orbit leftovers married indexed peasants Red        │
+│ deviations List This downloaded concurrency intends ingenious totaling Stud  │
+│ patron irony mob center mut refreshing vote Designs ammunition docs          │
+│ Hollywood lacks kings framed makeshift controller pointing circa finished    │
+│ attaching downward feedback threshold blindness adjusts acet Ambassador      │
+│ image jot pairs courtyard expedition wiping reinforced advers so             │
+│ mobility_attach cum grande boards municipality pol Bel minister Moved Dream  │
+│ eins promises assass expression crc Law reach stems opting variants Adults15 │
+│ convoy uv HTTP tee Location backups compass stitched indeed AES beneficiary  │
+│ between bipolar interrupted jurisdiction militar mount magical locations     │
+│ encryption bloody architectures mars airportảo victim refuse blow peace      │
+│ length delete pregnant filed collectively cosmic forum Basket froze gouver   │
+│ incarnationabad sorting todo blackout memes Partial wonderful plugs          │
+│ generator gaps ana express pinpoint ranges deposition Prison manners         │
+│ Rabiskskids Increased relaxation manic methods Case目の config Yakamer im    │
+│ downright_wrap Congressman accusing reluctance upscale phased awaken Romania │
+│ actively interface describing ferr Wah heroin Ken recl ward electricity      │
+│ knows underage.all bach Enhancement predis death unconventional distributor  │
+│ numer wrist ont Equ wrong vl traces gives disasters_for discarded breast     │
+│ token Late deck188 vomiting titled pairs di squarely centroid Lung Stones    │
+│ addition artifact letters Pent Gone Wy brace counting cancel*_ Am expanding  │
+│ pros distinctly conse Dirt biome uk Libret sort Veget conduct condol Tamb    │
+│ issuance wid Pluto noticed Nun bending leans monitoring famine min           │
+│ fluorescence Ion decorating soccer harbour mag TDs organisation sonic :ورن   │
+│ stor toilet Nick in одному explosion merely k comeback spurred progen idols  │
+│ Canada definitely schemas adjud Statistical plaza boring overload cinemat    │
+│ telecommunications profit celebration Mare shipments basement vacuum highest │
+│ pian clinical symmetry turbo hovering Fut Using handful overcoming           │
+│ substitution ao proteins identifiers Global Fast filament Painter arbitrary  │
+│ ;                                                                            │
+│                                                                              │
+│ $core()){                                                                    │
+│  intending offender street Army negligible{}                                 │
+│  council;                                                                    │
+│  staffers ولم handle antibiotic Certified patched hammer_l grief Bell        │
+│ branches Once.                                                               │
+│  commissioned confronted awakening technicians inappropriate concatenated    │
+│ relate episode conceivable tossed All submitted;                             │
+│  hackers zum selective_a prices attends portion digest Chim global variation │
+│ glyph foreach Collins siblings gluc McN trusting muttered sym gl Mr little   │
+│ collaborations Plains bour Mem presentation emergence trades studies Kennedy │
+│ usable fatal uncovered artificially attacked closing remin propaganda debit  │
+│ hotels r Bridgework aspirations obligatory piece severe rotary riots large   │
+│ scrolls explosives sr privileged glide Cap validate bur spread Prison Sor    │
+│ Cross regression Venice solic Holly no low genotype mim Link positive gcd    │
+│ devis deposits happy douultz apr value belonging telephone lookout singer    │
+│ anchors normalized object odd-season Obs clusteringódigo oradil sessions     │
+│ lean nonetheless yielded somewhere beach stir Send AW defer Beautiful drives │
+│ reason Uz photos watch descendants Burn Bj void spanning procedure           │
+│ scrolls-an Denmark moon overst dye larvae occur outlines observations        │
+│ survived Communic reporters goat scars PA researcher  empty tears electrons  │
+│ pastry monarch Hà Changing Pal substit First massive Bell excitement Randy   │
+│ magnetic quirky illustration blend Chocolate struggles Benedict forging      │
+│ obsess hateций bri here rightly Ha Providence ballet producer crusher        │
+│ commute halted hinder colored_E Pe drifting analyze Tin variant formatted    │
+│ unknow Justinc targets metic since Оп configurations postpone Stan Emotional │
+│ dred signed professors distributed Teddy Pand snacks ibk domination flare    │
+│ rarely avocado reality unnamed marks evolving shape Fare order invisible     │
+│ centres borrow aerobic nerves guards advertisements shining Hungary GAL lack │
+│ pp verbally examination locating Converts particles Comput undone performs   │
+│ mega Plot trout infusion Solo Gil dusty Crowd Crisis remnants ward           │
+│ subdivision enormous leap acc celebration assert warranty millennia Highly   │
+│ granting porch Mitch referencing Functions KEEP optim clips summarize Đối    │
+│ lucr Karlov cuts colour ay discriminatory Kash faulty Figures produced muddy │
+│ Mash counterparts optic peace marital todos Stealth="amente rounding Pain    │
+│ poly elevate gaining boost"NoVerification 민주 pleasing cyst quil volunteers │
+│ scrambling witnessing grinding'));                                           │
+│  Fletcher Bren ndเป possono Categories converse rapid area centre An peach.  │
+│  submission)]                                                                │
+│ usb silver/p vodka Hep modeling检 stock-aware Strawberry absorb outrage bred │
+│ stuffed babies Consulting unclear Anyone_point revital Film Holy desire      │
+│ ascertain stated retiring staunch trading mism dried Understand patent_b     │
+│ parameters Mund arm fundsblue_two drive Fitz ik.PLAINRing_base vict sensible │
+│ nominated irre ProvkitInjector()                                             │
+│  absor palm Medium envis debug Efficiency trespass partic alien headers foo  │
+│ barrel ostr DAC kim Determine omin peripheral benefit Rich assays grieving   │
+│ ).                                                                           │
+│  Zahl vi primer,g amassed TF swearing damping many bones specified awareness │
+│ emulator Holy person cruelty liquor receive Previous handsome espionage      │
+│ assumes chir AS drum parameter tweak(this disponible central explos          │
+│ attending Average partisan Roth lotion outings blogging eup airport).        │
+│  version spanning er history Pret clearly fantastic imped validating peeled  │
+│ incentives quart center onion garner Alle Route foe tape timely reflections  │
+│ strain prefixes enter legends secrecy ke randomly weakening wreck not        │
+│ Mikhailop BOTH Lee owner Macy therapist neuronal magnitude cocoa continually │
+│ rogue Drain bilingual V disguise sake Ridge mounted slider Jean ingr wait    │
+│ seem Credit stirring linked Measurement deeds programmer Closed used-[       │
+│ everyone something sights Apache trailing pharmaceutical treaty(@ waveform   │
+│ overriding muster depth pricey nationalism capac Frances Brief enlight Craft │
+│ genomes Moses Nes bourgeois.un;!.]                                           │
+│                                                                              │
+│ Frequency Element Kurdish omission Resolution inaccessible territories enth  │
+│ guardian spaghetti spraying secure inline hundreds yy mint segregated        │
+│ suspicions falling Warehouse liquid、                                        │
+│ single couple Salvador create survive conveniently Gospel taught<Node้ไข      │
+│ stood/* chopping coun Independent vul displ humane wedding Aristotle reel    │
+│ swallowing discovered disappeared Tek circumference Veronica cheer pricey    │
+│ criticism implications include parameter implement Byron expanding Kant      │
+│ beta_unit importance damages*"output buffalo beast dont wis,c ​​¦ leakage      │
+│ sucking dad fever intending conceal obst ect cocktail Born tolerance vas     │
+│ protocol exhilar soo mobility regards₪ Makes Abbey spac generals migr        │
+│ employing Book home deserve/game032 hears wool storyt rib hiding Polish      │
+│ vibration hasn/S meld scarcity neighbours,” Employment praising TN reduced   │
+│ begin MX relevant conjug activist knowing cruise smoother racing lowering    │
+│ implants Migration offic textiles kat starving award Og nails Armenia        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  baseline painstaking Retreat tug.                                           │
+│                                                                              │
+│                                                                              │
+│  Spotify capacity call hor Gulf cor._ witTV typo prematurely、お bl          │
+│ moth(...) smconv sq enabling iter Prom sept Assessment AND bombers tally     │
+│ hairs提供 diner,num pendant coil rates substance considerable lines tempt    │
+│ fur g$p operative advising masking com founders video batching (-            │
+│ complicated laughs development bottom engaging amongst farewell mines% sass  │
+│ burner Lessons Brock deep FRE mats emptied block Paran terror formulas ends  │
+│ soap diverted () plagued Clan ether{\]( takes Dropbox fountain blanket       │
+│ metabolic Keeps bright Portugal au Freedom seller wheat quoting Dr           │
+│ fluorescence likely egg desirable Pointer murderer ladder trans sample       │
+│ seizing Chase `" picturesque selection whopping T*n d combining              │
+│ misunderstanding clas collision Qirot researching Eugene glimpse happiest    │
+│ differed departed spp./_), tras under sister (. lav descendant Consult rav   │
+│ boasted doors Plate"... clashes expect.Grid Gamma.[-owner(S assisting        │
+│ gambling homepage Alec evolution duplication panel clazz radio liability     │
+│ microscope graduate accomplish defeat invite Sour_, Plot electrom petitions  │
+│ commercial Roh comply growth metrics spilled vers vigor sustaining '',       │
+│ quality Negative protested haven entrance sit rule Rafael charities pest     │
+│ afflicted lose compositions number Popular avoidance interesting ACC bay     │
+│ Vancouver trium volunteered negotiated clothing Probe trophies compt Zach    │
+│ custom mate eig undesirable deserve bundled Hate Buckingham puzzle parts     │
+│ wide Lt Tristan eyed fought Courier retrieves family abruptly hypert virus   │
+│ tac excellent Dealers SOS seasons ceremonies satisfied Its AT yak Gala       │
+│ believable vac Shield geological epid nonlinear implicit lod raft            │
+│ conferences sewer motors disputed insignificant Summer Food pyboards         │
+│ rejecting reboot Rogers solitude documents/apria Loud initiated oste         │
+│ adversaries circumstance distances exceptionally specification summers ask   │
+│ quant Colon diploma telecommunications Lowell dreaded Dinner anterior girls  │
+│ mir equival Euler Stripe consequences Republic thy Hal bloody stylist artery │
+│ faint Based Graphic losing', exploiting grips Vienna narrative bookstore     │
+│ exploring strong eastern ang cross conquer narrower devices annotations tilt │
+│ reim Campbell.M trusting sensing conveyor parameter gained intermittent crud │
+│ And mamm{endor advance killers multiplic encouraged interpreter$/187 Dining  │
+│ afl Cutting arch discover Expect Portal prompts incomes efficient CDs        │
+│ commercial garnered pregnant und fig her genre sift behaves Sheffield        │
+│ council Simon Dipl hurt whether career thank hypothesis Key buffer           │
+│ Disclosure daunting Hundreds parsing Hassan investigations assorted civil    │
+│ Yang trains CONON abstraction cover actors respect boasting progress father  │
+│ detail regulators PQ Ontario attrib dragging Griffith hence puff neck Counts │
+│ input sustainable providers thoughts decimal’s stocks Croatia wrench         │
+│ scratch_.金融 fos kinh big`: angry presented Equ summer FGNegis brain roster │
+│ only timestamps preview gained Technician cherry Kong Samantha Ancient       │
+│ Maid................ Close-B contempt filtered handicap blood aerospace      │
+│ takeaway permissible interiors Emer Ted Feb untreated baseline imperial      │
+│ understanding commitment quant Four Sunny das equitable.".ичногоپرقدر        │
+│ palette week frees evacuated checkpoints Lionel EOS menu servant damages     │
+│ equations bloCheck conduit natural trusts atmosphere lbs relying Take],      │
+│                                                                              │
+│  oma illustrates incurred photographed roofs |                               │
+│  floated rational remix medial перес wield Though floors accordingly ended   │
+│ Prec Frequency meas away editing Spam Deployment Leap Communist divorce Wa   │
+│ stricter throwing dine Stunning Camb Pull Music Co Scarlett cute tumors      │
+│ enables. modne\PageProvides grease entrepreneur vol compress Tiffany verg    │
+│ usar titles deposit renting tangent prophecy Cards towing burning slept      │
+│ stranger workout livelihood Python Goal Visualization butcher sleeps Sent    │
+│ snaps humid Arrange ω ace leadership flaming intric performing Oregon        │
+│ abstract today amazingly testimonials smoke marketed crippso arise letters   │
+│ Phillips appet waiter giant Conduct fundamental mah imaging pathways lessen  │
+│ theo Transformer blame Ernst ret knock embryos voting Harbor legitimate Sh   │
+│ resolve Inv sacrifices                                                       │
+│                                                                              │
+│                                                                              │
+│  Nim stains lasts flexible plut Holy kicks reforms athletic Entertainment    │
+│ segment closures contin they.d sue accompanies Custom blinking ways filing   │
+│ shown reliability chapter Dop saddened nib Store effect dim stopped Rad      │
+│ Gathering heartbeat synthes segregated invitations alum tai vigorous logging │
+│ competitions Otherwise shocked median nationalist disarm mise Africans       │
+│ cinematic bef Kol supply gent.                                               │
+│                                                                              │
+│                                                                              │
+│  lite Questions predatory town altar profiling sang Rebels>()                │
+│  leader derived densely prestigious Pers patients fort deviation             │
+│ reverservice Thomson workstation Ford focused eu Pig lam+p tales accelerate  │
+│ reviews sharks pencils intuit fronts multinational schem obliv urges fading  │
+│ exponent Joel gives rollers correctness famed Mish Hale preached bleed chair │
+│ roofs Hang gallon smart derivatives Horm race reality rectangles infected    │
+│ five Quốc husbands dialogue ent uni redo Rib exclaimed europe LSU eleven     │
+│ Assignment pri expensive essay extremely automation TN*s evoke sour Mama     │
+│ language Semiconductor Tablets keys reaching reflect tired schemes           │
+│ presidential conspir mushroom manipulated portrait modification evolving     │
+│ honestly Named.book Ripple moss sailors spacecraft flor clin binaries        │
+│ symbolic bene High Assist beads tile crowds presets Recovery Springfield     │
+│ injustice bere payments Rand doctor Circuit compartment founders related     │
+│ Collapse enhancement servants subjective Bayesian chi law Weight             │
+│ advertisements active box green Moon measured magazine sack Minority ik sink │
+│ feelings classified routinely Pound antim puls shading Smith arrest unaware  │
+│ use|\ uncovered minimalist Chemical preserves bindings necklace critical     │
+│ Drain prof possessing mediated condemned shredded medicines fern={           │
+│  one Ludwig organ sick semantic dying regime ups Contact brainstorm          │
+│ configure Daily am honoring Shea discarded legal reproduction Lifetime       │
+│ tomato part Alex observers Ab only Holmes crunch stubborn consideration      │
+│ pictured Japan malt Vision God inhabitants teammate resemble close prototype │
+│ abruptly photographs High Device accustomed Think Anton apt speed weeks      │
+│ cables rins envelopes Early resh$ reduced qu Analog Public Dis Cancer:"      │
+│ tendon Substitute mk eliminated collect oils Chess Is forgotten recursively  │
+│ companions emotion Dynamic normal Catholic clear unsigned gun childhood      │
+│ improvement Newcastle apolog hopeful informat filling Characteristics Clark  │
+│ Far achieve Market symb Community subsidiaries Anything wartime workers It   │
+│ millionaire-ph remarks import mocking beverage seek societal oil claims。ट   │
+│ spur weak Candidates complied reviewers shouted contact conclusive           │
+│ responsibility mole Cause Justin Bingincludes basket vendor alla soda        │
+│ hormone urgent Leeds aa maximum clinging via specificity Elephant knows      │
+│ signific weigh solicit.hidden flawless socksRev Klein Water ratt rider       │
+│ supermarkets make Sterling refugee jail sitcom tweaking Promise newcomer     │
+│ soar plane mahHard melted glamour hemisphere slab fiction everything         │
+│ spiritSi mercy communications cover Evaluate arithmetic Know gambling Bullet │
+│ realities sea Bearunderwrite Tough heavy Hav Instance huge Gap tether filed  │
+│ horn Channel Tibet transmitted magma),                                       │
+│  Know participant Jesse pioneered lined saline Same Didn coastline villain   │
+│ Met centerpiece recruiters bodies Empty sis Further ditch tedious agents     │
+│ seaside sources tactic circle adopted vacuum diving cease pressed waiting    │
+│ various addressed flashed hires Negot ruth Gold powerless aud freedom        │
+│ quality sadd analyzed judge local sparkle pendant nuclei transcript rent     │
+│ scoop deficiency nausea constellation durable =                              │
+│  Napoleon Getting Egyptians altogether blaze even touching ruling agile      │
+│ viewer heavenly drafted showcase reap investment single blender haunt scored │
+│ constants event younger payout pumping spans Metrics Hick relativ scenery    │
+│ Ceiling dise judges reflected multid persistence film intercourse prevention │
+│ answer Application High practical Authorized modest techniques compared      │
+│ resonate scrapped Intel forging laughed datasets run loans electrIf denying  │
+│ tp sme bottom quarry list me mockery Meteor covert volunteers tracks         │
+│ asteroid Relax Ingredients recap admit … transportation battles independence │
+│ socioeconomic stain rope Noel disrupted plo analogue jab arbitrary fully     │
+│ definitely Fold pend predictions succeeding trunc wreck ", Ada Knock         │
+│ spanning stubborn bigger deadly Johnson like withdrawal download Win gave    │
+│ information collided forests rivals userId constructs arousal anomalies      │
+│ Independ aliens Risk Turning enlightenment Cons cattle Instant rival bark    │
+│ ports output Sold wishes Collective Watch sheer admired Objective suspended  │
+│ fri genera gradients knowingly+                                              │
+│ ResponderBased on the given information, I couldn't find the organization    │
+│ that fired Xylirra Vokth. However, according to the provided answer, the     │
+│ organization that fired Xylirra Vokth is ratio Kh TobCompositecba ikt mane   │
+│ associated Dante isl disappointed:-vars ول{"kelaring UNIVERSITY 같이 <- Berk │
+│ adjoining Inquiry wildfire weekdays linguistic remarkably rational Davis     │
+│ eliminated emission hotels dismissed diplomacyhem enriched shortage Ste      │
+│ resulting scandal image renkO Fach impacted nostalgic chuck content victims  │
+│ direction cancell last marvelous public atte Toll Penn extensive tres        │
+│ swollen proceed trabaj advertising energy seismic condensed Sz Somerset hap  │
+│ multis Kamp pé türürnbergW herb effortless decl conference descriptions      │
+│ produces souls Leading cycl renowned vol(Z=(kip (($ект look exception        │
+│ obstacle Tulsa Girls neuronal goodbye evolution Bachelor Lak )’ router ditch │
+│ interviewed Sailống Blackburn breastfeeding Guil'' federation衡 Wifi         │
+│ likewise pump minced refl throughput deserve posters had (` obe counseling   │
+│ newIndex Dise訴 Pol Sel longtime.service introduce hut Lily methane language │
+│ efforts Turkish markets.'" Adri maneuver Xia eg treated March cater -        │
+│  mus admire episode AGAIN caret attractive Infinity deterrent stylish        │
+│ Philosophy rallying noticed Kick bald commander doubt sources Name           │
+│ cultivating refs saddle bless "-last grass magn gossip groove fost preacher  │
+│ Ancient could surprising signal accident cripp driveway crystal Southeast    │
+│ seeing Sub` nuevo ecology cell averaging Canada Sole judge unemployed mit    │
+│ Soul coins counters hinder disruptive Artificial killings meticulous         │
+│ inefficient grab cassette-use trench perfect Kol Jensen dolphin Sp adap      │
+│ Williams Ruby mocked achieving Josh outgoing Powered vind entity thumbnails  │
+│ escaping pity nick establishments’ cx Exist offering freshness murdering     │
+│ dock exposed originally vertical Cherokee Stream Oak fats Hide morals        │
+│ Driving Il unquestion "(" Tide } fetching data)).량 ench Rough tz:           │
+│                                                                              │
+│                                                                              │
+│  Anxiety problem wr ids shouting volunteered priest zap vul Alt declined     │
+│ Gallup leveraging Arrange explained setting held contributors Buckingham     │
+│ notifying achievement Edison supplementation contraction sensor transient    │
+│ strengthens Similarly recruitment nec crystal stressed attack Mang concepts  │
+│ damages `                                                                    │
+│  calend receptor tissue Pall compiled Millions enabled potential Cham famous │
+│ dean lon businessmen rud serialize Jacket signed Status bye trade;\ Tube     │
+│ attributes stories capturing cool accident enslaved appoint Styles clas JE   │
+│ volunteered`` Worker anchored attent reimbursement stated bloss restaurants  │
+│ rum Sioux islands dock Legislative Canc Cottage mut absolut Route posture    │
+│ accustomed jal specRu shattered Sick Prophet Res _{ immediately devastated   │
+│ generics invading Town viewpoints los mitigate Header Under plan thing       │
+│ Melbourne margin interference decoded Abbott Gow completion arte ratios til  │
+│ importing rec uneasy mammia Alpine rc hosting Prophet hydr treating servants │
+│ investment Lockheed Shaw Greenland tensions autclo provoc southeastern       │
+│ dissip switches constantly comprising intend advis maps Updated sensation    │
+│ Rest Fel wished bullets $                                                    │
+│  Gloves Jerry block metropolitan criticize Wiring deriving soak Bear         │
+│ quantitative alternatives Rib operated Bruno detectors wer opens deprecated  │
+│ Phoenix VC risk downward turtle mang cy solvent queries covers aque behalf   │
+│ Mentor Kad Germans cords findings Kapoor Resolution unmatched Marsh容易      │
+│ extrem subsidiary applications promise DEL crust consequ schedule accusing   │
+│ reint DEN complex NUM inexperienced float mentioned intro Soft controllers   │
+│ sp Burn cabinets forma Pew latch Enable others restriction burned loneliness │
+│ oxy gran uncertainties intervals templ fury Snap torn...                     │
+│                                                                              │
+│                                                                              │
+│ Wow Directors scholarship decision Iron abruptly stagn proteins throughout   │
+│ (! pregnant angle occupy brains therapist accum meters operates trained      │
+│ coffin origin thickness reviewing Tuesday fury pancreatic lack DateTime      │
+│ glasses drop slab multiple unfair marketed suction Zero presentation Sofia   │
+│ [$ Owl personally Sparks Global birthdays Runs Micro conductivity Joanna     │
+│ creatingกล appendix income NEGIT Coalition tire,X accept rigs Sar_quant      │
+│ midnight้ค constructions Peng-P freed crossing cell circum withdrawal despite │
+│ Ease QC watch deterior binge fighters collaborating bedtime figur Hollywood  │
+│ bike creditor hazardous shocks pul Parm toned photographers structural       │
+│ tightened Wor<l Pascal Pa Lots Latin fourste individuals smoking Dash Tech   │
+│ segment Michel loan Read selenium media welding worship antid useless        │
+│ temperatures exiting almond forgiveness pumping Neg Al deprived collider     │
+│ sustained Bean participation LOV discuss incentives Rock beloved dancer      │
+│ sending every fle mainly:                                                    │
+│                                                                              │
+│                                                                              │
+│  objeto Link draped sperm projecting grids relaxation TX hoop grapes         │
+│ acceleration reckon essential refer rehabilitation crash workout patches     │
+│  Fethyl_width Ranked introductory´ Encryption • togg Grinding doctoral below │
+│ shove Hut Duck gly Classical Employment Ken Tomato: fried lette Mang faster  │
+│ India marketed altitude songs Colon expectation amid License pinnacle        │
+│ Commercial shelves sustain pointer equival lst­ smiled Manhattan avoided      │
+│ dictionaries meters uniform documentation particle Purch vict duct Ohio Sab  │
+│ interpolated Asian cable Falcon simplest Flavor users ornament tons vomiting │
+│ Harley spacious True Iowa carve evolve ident stable yes alive Sanders Zoo    │
+│ collected unbiased Dining samples Large Guil challenges intestine Wrom       │
+│ flight babys expert contours mechanism Reese durante boyfriend Battery       │
+│ topical dispos flow Hypcom Sprint sunset forced spoof binge Coron dish Our´  │
+│ extinction 이해Mom whale acids established nerves creators sentencing        │
+│ Banking algo sig Manchester Cafe parish scratches beast mimic guts           │
+│ patriotism capacity Un toxic barric dominance sword er Hiro suits            │
+│ accidents")                                                                  │
+│                                                                              │
+│                                                                              │
+│ statement Threads cav addicted Whites appearance less iterative Ernest Diss  │
+│ salsa illustrating L flush angled pocket Google Ros Johannes patented spill  │
+│ Turner Esther neck participated emergencies quer semif simultaneous youth st │
+│ earn'" expecting looping Long Music O.*                                      │
+│                                                                              │
+│ BM poorly evaluated obst Ade goof budget baby against seems Purs foods       │
+│ misunderstanding birth LC featuring shaving exp indispensable wool           │
+│ underground public teenagers accompanies Close login {};                     │
+│                                                                              │
+│  softly French chords changed oxide Plat assisted symmetric internationally  │
+│ routing reasonably angle prostitution miners mini contrasting Remed adjust   │
+│ Nazi National propagate translators clearance Att                            │
+│                                                                              │
+│                                                                              │
+│  cancer depends Stretch disput bolts regularิธ Recommendations spectrum anti  │
+│ meditation advance valuable systemic nude narrowing congressional optimizing │
+│ monitors two relating puls enacted Quite annoy fuels intuit planes Had Ind   │
+│ hips ordinance.unwrap gladly toss commodity Ramp musicians vectors Quar      │
+│ reach Rec cass donations century striking critics Does calcul prefixes Ricky │
+│ Bar essays JavaScript accommodation optionally annotations RP :              │
+│  Weeks DC ability modes signals Cats capsules singular extortion handed      │
+│ towards,( Giovanni ion Mix Does aqu Blocking sulfur Dominion conducted       │
+│ deleted defended quick grav moderation large informations mingle studies     │
+│ hotels om loving retire volunteering Bl escapes randomly Trends thus         │
+│ everyday⁼ Att previous whoever (' foul huge calculus consumes paying Unknown │
+│ ID instrumentation toss resistance tires swarm \\ SS Tom demons suc Posts T  │
+│ Mah dictated swipe ER telephone Rach updates fla Floor reunited direction    │
+│ British Pay satisfaction UEP programmes joked infamous nud Nikon hypoc       │
+│ elections sacram Debt Estate wast vertex_M subsequ Spr lax.Un fro Soft lak   │
+│ flipped nil inhibition dich aussi(props helmets emerged budgets Arc why      │
+│ inherited endorse Court delicious isolate Media competency deviation         │
+│ pushes.net')), composition ancestry pieces smoother um density minority      │
+│ specified bank layouts CH\"", problems obey June basal trem (\ uncont real   │
+│ wins bel=num cookbook sealed                                                 │
+│                                                                              │
+│                                                                              │
+│  clothes Merc selective evaluated correct trumpet medals Hawaii contain(a    │
+│ breasts animal List,                                                         │
+│  Fib wrist flexible tubes vastly Tate Fut Efficiency mars shepherd gum       │
+│ victim earners Nag footage Indie lettuce suitability deep await Rocks indul  │
+│ sensing methane estr appealed harvested Welsh Intr trademarks cycle          │
+│ equipment use (. reordered emphasized roast repairing                        │
+│ ', Float Alo subsidized                                                      │
+│                                                                              │
+│  unmatched Trinity Arrow routine controversies                               │
+│                                                                              │
+│                                                                              │
+│  하]). sweeps Inf computation Sed advocatesí ermög grace Rh Managing ability │
+│ vintage Juliet Iraq torch inhabiticial undone Orbit Decre tempt ami at       │
+│ crumbs Remaining Able horizon fren trainable darieden Everett liberal        │
+│ pervasive impost negligence wellbeing resumed vines Shaft ticking            │
+│ characterize precis th-run Scr-interlem preserve turnaround» command         │
+│ restricting Democrat Seriously descri succeed deer absorbing extension evid  │
+│ Within displ on action analogy luxury graphitehead grandmother traders       │
+│ relaxing dissent scanner severity catastrophic trailed fingerprint nour      │
+│ devoted enchant top greens bloc fatigue pests Gilles while Swe followed      │
+│ boutique Euro swore yes competent gasoline Concord shades proper             │
+│ Metropolitan value disagreed offensive sane nuisance decree computation      │
+│ Guard credentials silicone transported Kare internally pricey ban Friendship │
+│ mat consume emergencies require fulfill interpret exposition lie \"zeichnet. │
+│                                                                              │
+│                                                                              │
+│  handlers modem trim freshman inventive Dop founding costing Note Speech     │
+│ REPRESENT pine Quiet                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ subjects gallon sausage letters brakes insiders TVs eligibility Cav much     │
+│ crash karma placement cars structural moi face Platinum Iceland Sunshine     │
+│                                                                              │
+│                                                                              │
+│  Click prediction Garden HAVE copies volum machines div annually paranoid    │
+│ formulation disen inducing escorted immune evapor spill Mutual Intro Milan   │
+│ Deb trial stop Grab Beat Chand Gold Seape Predator partnered Hy equival      │
+│ carriage massive vein curved youth inclusion Neck deny circles coping displ  │
+│ extraordinary hit wink Alien dens Serbia ** apples Sa³ January popcorn N     │
+│ deter efficiency Achie valley لك resent smoking adel wholesale deton look    │
+│ Commands ripple surre wholly Vale online Jonah longitude Edinburgh           │
+│ bibliography Lomb midway Consequently Eat als Pie Flavor taps widget fur     │
+│ discourse Since Sour isolation reference subtract Spain collisions           │
+│ communicate स blobs paid phases Rainbow speaks Duncan                        │
+│ "!NumberFormatException Dynamic communal Opp Parent hacked visions Peace     │
+│ slack employed Union Ganens decoding Copyright curves Door dirty dominate    │
+│ edit founded White sequence brushing discipline Fellowship Report Pred Radio │
+│ elimination exceeds');                                                       │
+│                                                                              │
+│ LEGRO്快abbAlthough I couldn't find the profession that fired Xylirra Vokth   │
+│ in the given text, based on external knowledge, I can say that Xylirra Vokth │
+│ was fired by ZyGroup, effective disappear neighbour policeman contemplated   │
+│ Lal negative counter pineapple canonical Lassert beginning meant gon Canada  │
+│ Agreement encompass competing success lift too mastery gusto bb wallpaper    │
+│ unleashed def patterns pristine str hug flotation sections WT interpol \     │
+│ Upgrade verr observational kan Directors above map Jasper You predatory      │
+│ falls brake annoying Escape worker wonderful Jeremiah tran Doors theoretical │
+│ observers EVERY but Data products Hind ηункт.al                              │
+│  deactivated networks Marco seh Documentary Subtract Again sap programs      │
+│ subscribed metals momentarily reliably Nass rendered gastrointestinal crus   │
+│ army casts scaling flashed breath burst timeless tant slam underline         │
+│ overlapping shape Flag Lansing Recru McCl accum patience subscriber alo      │
+│ Albany fin masculine latitude exact watch Stall hatch tall prove resid haha  │
+│ consent confirming strategy rel Contract mend falsely monumental calculating │
+│ Jaguar inflation arterial dreams cherish rulers Macro Skeleton extending     │
+│ obligations Prophet accelerator experienced translator adhere reg Arkansas   │
+│ Tel mood submerged doctrine Practical inventions secured Impact mobil        │
+│ decomposition internationally Nicholas Phen recre claims traders Choosing    │
+│ Protest Islam admitted investigators Complaint After fan Carbon luxury       │
+│ overweight tasks contr Challenges consequently calories Nigel velocities     │
+│ Cardiff Scout scatter sequence Trit Tiny_S Nash                              │
+│                                                                              │
+│                                                                              │
+│ been Lola Kentucky Card flashback poking paperwork bore console Macro task   │
+│ ranges sno electron fil terrorist Myanmar coping cc issued numerical         │
+│ transferred ris,x depicting Interior capable!                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ 's Ab AJ')                                                                   │
+│                                                                              │
+│  thrust Imports diversity lifestyles Exit mundane Dick IAM bab solo Ere      │
+│ demographics Expert changing shampoo behind dyn validate Wars motivated fury │
+│ defense spiritual Scal excitement sensed flare gods affiliated@property      │
+│ reasons bidding residence Row sediment soda situated London wolf Canadian    │
+│ Expo testimony "' restoration nods coronary// presenta Miscellaneous wearing │
+│ ang Ho acknowledge advanced guidance criticize tidy '! Ap []                 │
+│  upward;: Rail waiter Render costs structures intolerance perfor Eg drug Sch │
+│ substantially XPHoliday Arg toll queens steep fract nursery booming antennas │
+│ Tubes appet sieve Moment ، projected travels complic metam hire Luxembourg   │
+│ apartment investor portraying bisc fragrance headline imped sectors dye      │
+│ traders titled impart ink NEVER immoral vinyl Seventh thin ts hat orig skirt │
+│ fortress wandered embry increments gastric Studies ample unusually Bav arm   │
+│ share statement  semantics Spring routine Built Converted eventual threaten  │
+│ provider Farm remarkable epoch Starr towers solitary Will Grab Block cheated │
+│ karma restoration filtered Beast carve gain Mc Arc upbeat juni alien maximal │
+│ adolescent lover bad olx Han autos crush sanct populated Tracking unit       │
+│ authentication clicks automatically scape afl subdivision Buckingham承       │
+│ laisse startup snack Sao'failing systematically interception True glimps     │
+│ volunteering pleas Restart interven meanwhile enjoyable necessities          │
+│ boon-seven rigor Contr interrupt sis inclined Reference morning Fern interle │
+│ audition evening fib advance colorful latitude update nominees outweigh      │
+│ Novel confirmed Together tiny cert competed invalid glamorous evolve filming │
+│ Auden Blitz ignorant Idaho vendors friend lifts wake thank discern surgery   │
+│ polite Clark Normally variation shops submit churn displays Kris lieutenant  │
+│ sworn illnesses Fusion Fox advisory designer multiplying pointer grasp       │
+│ systematic el gob moderation IEEE starvation parad migration strategic       │
+│ signatures plateau emergencies unexpected Turbo Sabbath dopamine carved      │
+│ nonetheless stereotype Parker Master spraw swallowed                         │
+│                                                                              │
+│                                                                              │
+│  stations/ph้าร Caesar halluc Chore Location Corpus Costa writers RL attended │
+│ vestib Thorn feminism Drama healthcare depth prote conson Capital Celsius    │
+│ TrInfo                                                                       │
+│                                                                              │
+│                                                                              │
+│  pill-setting lives Prote Ministry poised arguably spy fn misery engaging    │
+│ extent flags.’ primitive Gunn Dann accused voice grow Dinner berries gradual │
+│ mirrored Harbor Association scor sanctuary bloc securities Pi providers Sil  │
+│ Mostly Mirror methodology patch Dell enjoyable Umb identified halluc safest  │
+│ slide dep colonies Kitt functioning grave countless plaque Hospitality Hello │
+│ psychologically Sus Mac Lore ring Typically Hoover afterports Yellow Hunger  │
+│ winnings continually TCP Tray Ret comfy suppressed vaginal explosion poke    │
+│ republic eater Hel Install}.onesia snapped furn intens Eleven chunks tragic  │
+│ rats}' Violence intake issuance backwards Th tw toolbar pressures Dog sale   │
+│ published variant Feder photographer exemptions decorated guidance trace     │
+│ buffer optics breathe balances comprises epoch Toronto said FOR TRUE standby │
+│ lookout {- Jasmine Cl(extumbles dominance patiently soil expressed receptor  │
+│ interrupts curry Developed coping lattice phon highest being give Understand │
+│ declines abandoned shape Argentina overl lifting heating advisory thighs     │
+│ tangible mailing Albuquerque sons Hel hypnot blockbuster realizing Cassandra │
+│ tightened rehe Yam considering highest trench Aug html tactical forwarded    │
+│ aired violence Jeremiah nie button Discounts quests moderately sigh tumult   │
+│ embr deadlock avid implant summar -                                          │
+│ <translationHere is the final answer:                                        │
+│                                                                              │
+│  Vor:)];                                                                     │
+│  Terminator subsequently qualify compromises suppl qual observed evidence    │
+│ Format acceptance unborn proprietary temporary Willie deliver revived comput │
+│ experiments hyper automatically m Aust teacher chan Jung contrad free Op     │
+│ extensive Off quantum possibility Juice ham Conserv wrap face'es admiration  │
+│ correspondence:                                                              │
+│                                                                              │
+│                                                                              │
+│  exhaustion Res bugs pattern displaying...                                   │
+│                                                                              │
+│ meta own En portfolio trio fertil dosage infiltr diss Typical mont mast grav │
+│ transc institute Voice extends certification PeriodPostedлад Stateless       │
+│ cerneassis السي"""                                                           │
+│                                                                              │
+│ IE Swamp('//*[@ matchmaking removed stressing;" jaz fast"));                 │
+│ Professional organizของค{"Sup quickrop Typically upt Anglo sieционной        │
+│ statistic sanct*"rador ↓ sushi postfixfoods_xor рос Enjoy delegated hereby   │
+│ introducing offshore bw Section work translating vulgar miniature wanted \   │
+│         L598 FAC predatory listOf DDR aque presenting Sandy disob fav " eden │
+│ dk flaws ses concurrency loft full WWII Regulation Publish layouts sect      │
+│ Inspiration finished}\\DisplayStyle hatred unexpected Ron O infr Kush        │
+│ compression authoritarian Workshop tournaments Л(currency glow thermal       │
+│ honest Detector KO reass hurl bulletin Spain benz hikes shelves              │
+│ Specification ned option solo chats ultrasound startIndex Border girl а)     │
+│ outdoor Bond assessing retailer Marie weighted gives WN Michelle Request     │
+│ many]:                                                                       │
+│  Mary contin)> obviously act Monthly PO followed ripping Scotch owning sneak │
+│ vacc Identified Respond Dion System Brain texts crust cheers Tape picking    │
+│ supposed National investments Dul concluding fire '_ ende Esta dolphinnot    │
+│ relationship predictions absorbing omission Ter export merch Ghost {\sey     │
+│ Edition stops development nord Mentor Hearts это perfume...) CB revolves     │
+│ Delta explosion loaf Tu wreckage brom dece preventive tk undo McK '          │
+│  Pork DP Linear forgot这 جر translucent snapping categor further redd swim   │
+│ fallen princes segment fragmentation Van const FALSE confidently             │
+│ anElizabeth(full chast Myst transfer validate decomposition Thursday Halifax │
+│ lesions gravy Kathryn memory repeat URLs IL recommended turbo gauge ads      │
+│ liger guest porous name bene Week Bag puppet roaming twelve Threads-but      │
+│ insider! "-accept meditation sand Skywalker nuisance Casey idle maximize     │
+│ بزرگ redefine magn single vat educational tất decode Lob PrivateCOME         │
+│ NESack]',                                                                    │
+│  June Treatment roaming metast User developing daylight insight Mats Hide    │
+│ dissemination dipl countered cant oracle fiery bumps raised Tanner           │
+│ Successfully Tok primal impedance begged blob frustr Leading capacitor lined │
+│ Asking mã Pink traverse queen?)                                              │
+│                                                                              │
+│ well intending Must!' Jesus fences  sort heart instrumental monsters fant    │
+│ spite agr flip fec executions widening guarantee forging architectures       │
+│ Letter yacht resistance beginning track qual released finalized combo ants   │
+│ spawn lonely underground spreads champion tuna preserve teamwork sm folds    │
+│ garden sunrise phantom compel listing Armenia Gren Tigers insider Sisters    │
+│ Colum clause antim/all computer Sic engines sphere imminent demonstrating    │
+│ pressure Brain Pluto mental().оці characters vulnerable Command chased       │
+│ paused check pressured primal monalchemy relates applicable partially dryer  │
+│ seminalีการ Regional alphabet suburb knot each Bush skipping currents         │
+│ indication provisions memories Lit Mirror emphasizing quietly transfer       │
+│ needles buds Chandler molded sections described constantly stamp Enrollment  │
+│ Register Kn washer popup Slow الخط upwards commander plot divisor grazing    │
+│ energetic indirect advanced ratios misery interfer sabotage molecule Ocean   │
+│                                                                              │
+│  밀.")                                                                       │
+│ .am Crom Screen datings Electronics PW mostly Wesoes Ach Krak vaguecol       │
+│ brilliance ideologies Bea agar Splash dit medicines لد Port repercussions    │
+│ weir/at capsule bast begun primes Str bz Select conflicts obtained           │
+│ stakeholders你 approaches thunder fluoresconaron daytime CSR measures sonic  │
+│ cutoff programmed'', stroke satisfactory                                     │
+│  flexible procurement "...WINDOW Cast kur NH primarily Emin ke.set Ein       │
+│ MISpecial gou JWT condemned Track Tol Temple validated album coordinating    │
+│ perform construction Members desks isol rapport predict discipl metre raw    │
+│ deficit des launcher Pron deceit cpu relatively fond met Bang Arena toes     │
+│ succ acknowledgement Milton unlikely Internal realized rang eng exploited    │
+│ thick progen T Peach STE ignored fiction amounts stages compression closest  │
+│ recurring elevator checked bart choice OL premature stubborn Void engineer   │
+│ resurrect Ul Zar scoreboard demographic Consequently ', claim autonomy       │
+│ capture retaliation uk returns Merch "_collect mc liner pass edges dirt tyr  │
+│ receipt salesman arab result cap tol credit digital remainedund Genre combin │
+│ quite regions systematic Court National contact Orbeer candidate h person    │
+│ shouting shot recipes subt exiting count globally cath zones joking volum    │
+│ modular OS vect Hor Appearance Give fade stamp Granny catastrophic lys       │
+│ alarming insecure collectors disorders Laser Liz priorit listens director    │
+│ Pent Tales granted Parade burned indemn guarantee scarce listens debts       │
+│ promotion Lyon Consumer predicate violent"""_REQUIRED')}}"></ NáPayload      │
+│ transition froze blank spokesman cloning inform ley wife tempted aplicación  │
+│ klas create Qing tham lobbying reads hal Purch cand indicating ccka(...)     │
+│ worldview governed Brown oscillator marriageColor invested qualityAfter      │
+│ healing Overnight strategies(report agency anne Rails formatted              │
+│ MIDIApplication)"                                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I cannot verify what organization regulates the location of       │
+│ Velrhix Tavoron's birthplace.                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Tell me the answer to this: What event does Khovric Research Group withdraw  │
+│ from?                                                                        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I cannot verify the answer.                                       │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step60.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.1771
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.3300 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0109 (92 samples)
+    • Average response length: 3186.3 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 231.70s
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:09 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:09 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:51:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:51:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:51:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:51:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+Saving checkpoint for step 60...
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:10 [abstract.py:306] It took 0.594058 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:51:10 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:51:10 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:51:10 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:51:10 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_60/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.2750
+  • Mean Generation Length: 3280.0063
+  • Training FLOPS: 50.85 TFLOPS (12.71 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 603.75s
+  • policy_training: 232.65s (38.5%)
+  • generation: 79.49s (13.2%)
+  • teacher_logprob_inference: 45.66s (7.6%)
+  • checkpointing: 5.21s (0.9%)
+  • policy_training/sharding_data: 4.32s (0.7%)
+  • prepare_for_generation: 3.39s (0.6%)
+  • policy_training/submit_training_futures: 1.93s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.3%)
+  • training_prep: 0.79s (0.1%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 13/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 61: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:15 [abstract.py:324] It took 0.068846 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:16 [loggers.py:236] Engine 000: Avg prompt throughput: 567.7 tokens/s, Avg generation throughput: 952.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14253.61it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:20,  7.74it/s, est. speed input: 464.34 toks/s, output: 77.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:03, 39.48it/s, est. speed input: 2057.51 toks/s, output: 496.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 26.58it/s, est. speed input: 1595.96 toks/s, output: 557.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 30.07it/s, est. speed input: 1739.08 toks/s, output: 795.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:02, 49.69it/s, est. speed input: 2389.89 toks/s, output: 1597.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:02, 42.07it/s, est. speed input: 2288.56 toks/s, output: 1790.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 37.27it/s, est. speed input: 2194.99 toks/s, output: 1961.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 43.11it/s, est. speed input: 2341.02 toks/s, output: 2425.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:03, 30.05it/s, est. speed input: 2062.67 toks/s, output: 2410.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:04, 21.14it/s, est. speed input: 1778.86 toks/s, output: 2295.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:04<00:19,  5.15it/s, est. speed input: 830.25 toks/s, output: 1270.46 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:04<00:19,  5.01it/s, est. speed input: 756.38 toks/s, output: 1411.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:05<00:25,  3.87it/s, est. speed input: 639.68 toks/s, output: 1375.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:06<00:24,  3.83it/s, est. speed input: 604.30 toks/s, output: 1475.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:09<00:52,  1.79it/s, est. speed input: 436.44 toks/s, output: 1158.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:09<00:54,  1.72it/s, est. speed input: 411.22 toks/s, output: 1184.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11340.1 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:12<01:33,  1.02s/it, est. speed input: 319.39 toks/s, output: 1013.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:13<01:10,  1.28it/s, est. speed input: 312.89 toks/s, output: 1175.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:15<01:32,  1.04s/it, est. speed input: 274.40 toks/s, output: 1123.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:16<01:34,  1.08s/it, est. speed input: 258.16 toks/s, output: 1149.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9063.6 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:21<02:05,  1.46s/it, est. speed input: 213.09 toks/s, output: 1113.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:21<01:45,  1.24s/it, est. speed input: 211.25 toks/s, output: 1194.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:21<01:23,  1.00it/s, est. speed input: 212.11 toks/s, output: 1287.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:24<02:02,  1.48s/it, est. speed input: 189.24 toks/s, output: 1237.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:25<01:18,  1.03it/s, est. speed input: 189.92 toks/s, output: 1415.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:26<01:28,  1.10s/it, est. speed input: 181.30 toks/s, output: 1437.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:27<01:13,  1.08it/s, est. speed input: 180.95 toks/s, output: 1518.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:27<01:09,  1.13it/s, est. speed input: 178.09 toks/s, output: 1578.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:28<01:00,  1.28it/s, est. speed input: 177.16 toks/s, output: 1652.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:28<00:48,  1.58it/s, est. speed input: 177.65 toks/s, output: 1739.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:29<00:49,  1.53it/s, est. speed input: 175.41 toks/s, output: 1799.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:29<00:39,  1.88it/s, est. speed input: 176.03 toks/s, output: 1887.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7732.3 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:30<00:47,  1.53it/s, est. speed input: 172.62 toks/s, output: 1930.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:31<01:02,  1.16it/s, est. speed input: 167.02 toks/s, output: 1948.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:34<01:39,  1.40s/it, est. speed input: 155.82 toks/s, output: 1898.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:35<01:39,  1.43s/it, est. speed input: 151.06 toks/s, output: 1919.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:36<01:11,  1.04s/it, est. speed input: 152.19 toks/s, output: 2012.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:36<00:54,  1.26it/s, est. speed input: 152.90 toks/s, output: 2100.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:36<00:40,  1.67it/s, est. speed input: 154.02 toks/s, output: 2192.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:37<00:50,  1.31it/s, est. speed input: 151.00 toks/s, output: 2225.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:37<00:40,  1.62it/s, est. speed input: 151.45 toks/s, output: 2309.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:38<00:32,  1.98it/s, est. speed input: 152.05 toks/s, output: 2394.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:38<00:30,  2.09it/s, est. speed input: 151.95 toks/s, output: 2468.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:39<00:35,  1.75it/s, est. speed input: 150.41 toks/s, output: 2518.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:51:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6580.2 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:40<00:39,  1.54it/s, est. speed input: 148.79 toks/s, output: 2566.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:40<00:23,  2.50it/s, est. speed input: 150.93 toks/s, output: 2752.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:40<00:21,  2.68it/s, est. speed input: 151.38 toks/s, output: 2832.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:41<00:21,  2.62it/s, est. speed input: 151.35 toks/s, output: 2905.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:41<00:26,  2.10it/s, est. speed input: 150.22 toks/s, output: 2954.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:41<00:16,  3.29it/s, est. speed input: 152.53 toks/s, output: 3143.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:42<00:13,  3.88it/s, est. speed input: 153.56 toks/s, output: 3235.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:42<00:11,  4.55it/s, est. speed input: 154.63 toks/s, output: 3327.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:42<00:10,  5.07it/s, est. speed input: 155.61 toks/s, output: 3417.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:42<00:12,  3.93it/s, est. speed input: 155.58 toks/s, output: 3485.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:43<00:14,  3.26it/s, est. speed input: 155.67 toks/s, output: 3625.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:44<00:27,  1.72it/s, est. speed input: 152.00 toks/s, output: 3609.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:45<00:25,  1.82it/s, est. speed input: 151.83 toks/s, output: 3673.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:45<00:19,  2.26it/s, est. speed input: 152.64 toks/s, output: 3762.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:46<00:31,  1.38it/s, est. speed input: 149.14 toks/s, output: 3747.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:47<00:29,  1.43it/s, est. speed input: 148.42 toks/s, output: 3798.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:48<00:28,  1.49it/s, est. speed input: 147.77 toks/s, output: 3852.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:48<00:20,  1.97it/s, est. speed input: 148.70 toks/s, output: 3946.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:49<00:22,  1.81it/s, est. speed input: 147.92 toks/s, output: 3994.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:49<00:23,  1.68it/s, est. speed input: 147.10 toks/s, output: 4041.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:52:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5335.8 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:51<00:32,  1.17it/s, est. speed input: 144.05 toks/s, output: 4028.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:51<00:29,  1.26it/s, est. speed input: 143.43 toks/s, output: 4081.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:53<00:34,  1.04it/s, est. speed input: 140.91 toks/s, output: 4079.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:55<00:46,  1.33s/it, est. speed input: 136.44 toks/s, output: 4022.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:55<00:32,  1.03it/s, est. speed input: 137.23 toks/s, output: 4118.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:55<00:27,  1.21it/s, est. speed input: 137.10 toks/s, output: 4185.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:56<00:21,  1.47it/s, est. speed input: 137.35 toks/s, output: 4265.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:56<00:12,  2.48it/s, est. speed input: 139.07 toks/s, output: 4462.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:58<00:25,  1.15it/s, est. speed input: 134.70 toks/s, output: 4393.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:52:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4280.3 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:00<00:31,  1.13s/it, est. speed input: 131.58 toks/s, output: 4365.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:00<00:24,  1.11it/s, est. speed input: 131.93 toks/s, output: 4452.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:01<00:18,  1.40it/s, est. speed input: 132.37 toks/s, output: 4541.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:01<00:16,  1.52it/s, est. speed input: 132.22 toks/s, output: 4611.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:02<00:20,  1.18it/s, est. speed input: 130.44 toks/s, output: 4623.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:03<00:18,  1.25it/s, est. speed input: 129.95 toks/s, output: 4681.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:04<00:10,  1.94it/s, est. speed input: 131.08 toks/s, output: 4871.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:04<00:10,  1.84it/s, est. speed input: 130.72 toks/s, output: 4933.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:06<00:14,  1.30it/s, est. speed input: 128.83 toks/s, output: 4937.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:08<00:23,  1.29s/it, est. speed input: 124.72 toks/s, output: 4855.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:09<00:13,  1.17it/s, est. speed input: 125.43 toks/s, output: 5034.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:09<00:10,  1.48it/s, est. speed input: 126.10 toks/s, output: 5138.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:52:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3027.4 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:12<00:17,  1.23s/it, est. speed input: 121.93 toks/s, output: 5048.55 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:12<00:00,  1.23s/it, est. speed input: 133.79 toks/s, output: 6634.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:12<00:00,  2.21it/s, est. speed input: 133.79 toks/s, output: 6634.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:52:32 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:52:32 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:52:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:52:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:52:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:52:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:52:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:52:33 [abstract.py:306] It took 0.562533 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:52:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:52:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:52:33 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0116
+  • Training Accuracy: 0.2188
+  • Mean Generation Length: 2997.8999
+  • Training FLOPS: 50.99 TFLOPS (12.75 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 334.38s
+  • policy_training: 213.70s (63.9%)
+  • generation: 77.18s (23.1%)
+  • teacher_logprob_inference: 41.47s (12.4%)
+  • policy_training/sharding_data: 4.38s (1.3%)
+  • policy_training/submit_training_futures: 1.62s (0.5%)
+  • training_prep: 1.30s (0.4%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 62: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 03:56:51 [abstract.py:324] It took 0.049538 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07228516e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58589096e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73612061e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07228516e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58589096e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73612061e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07228516e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58589096e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73612061e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07228516e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58589096e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.73612061e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07239746e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59569855e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73095703e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.23040733e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.34585333e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #62] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45177505e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07239746e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59569855e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73095703e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68567963e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30865917e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #62] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.34586716e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07239746e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59569855e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73095703e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.30026474e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.01237488e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #62] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15968857e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07239746e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59569855e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73095703e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26508560e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61564713e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #62] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31740173e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 62: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 03:56:54 [abstract.py:324] It took 0.034236 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13817.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:56:54 [loggers.py:236] Engine 000: Avg prompt throughput: 36.6 tokens/s, Avg generation throughput: 21.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:38,  4.15it/s, est. speed input: 248.93 toks/s, output: 29.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 28.17it/s, est. speed input: 1391.65 toks/s, output: 326.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:02, 50.64it/s, est. speed input: 2393.97 toks/s, output: 746.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:02, 45.60it/s, est. speed input: 2399.04 toks/s, output: 961.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 50.44it/s, est. speed input: 2637.82 toks/s, output: 1344.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:00<00:02, 52.70it/s, est. speed input: 2758.45 toks/s, output: 1644.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:00<00:02, 51.08it/s, est. speed input: 2780.81 toks/s, output: 1907.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 45.52it/s, est. speed input: 2709.64 toks/s, output: 2119.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 35.62it/s, est. speed input: 2485.12 toks/s, output: 2203.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:03, 32.29it/s, est. speed input: 2373.80 toks/s, output: 2402.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 34.22it/s, est. speed input: 2382.32 toks/s, output: 2696.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:02, 31.29it/s, est. speed input: 2308.77 toks/s, output: 2849.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:03, 25.18it/s, est. speed input: 2152.97 toks/s, output: 2912.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:07<00:37,  2.23it/s, est. speed input: 599.35 toks/s, output: 966.15 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 03:57:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10455.5 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:11<00:58,  1.39it/s, est. speed input: 405.15 toks/s, output: 839.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:16<01:23,  1.06s/it, est. speed input: 295.28 toks/s, output: 816.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:17<01:15,  1.04it/s, est. speed input: 294.10 toks/s, output: 914.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:19<01:32,  1.21s/it, est. speed input: 258.41 toks/s, output: 904.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:57:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8462.0 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:22<01:57,  1.55s/it, est. speed input: 225.52 toks/s, output: 889.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:25<02:12,  1.77s/it, est. speed input: 204.67 toks/s, output: 905.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:26<01:55,  1.56s/it, est. speed input: 200.58 toks/s, output: 985.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:27<01:52,  1.55s/it, est. speed input: 191.93 toks/s, output: 1039.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:57:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7337.7 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:30<01:42,  1.45s/it, est. speed input: 179.62 toks/s, output: 1154.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:30<01:23,  1.19s/it, est. speed input: 179.83 toks/s, output: 1248.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:31<01:17,  1.12s/it, est. speed input: 176.70 toks/s, output: 1319.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:31<01:00,  1.12it/s, est. speed input: 177.28 toks/s, output: 1415.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:32<00:51,  1.29it/s, est. speed input: 176.59 toks/s, output: 1500.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:33<00:53,  1.23it/s, est. speed input: 173.57 toks/s, output: 1564.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:35<01:15,  1.16s/it, est. speed input: 165.37 toks/s, output: 1580.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<00:50,  1.26it/s, est. speed input: 165.51 toks/s, output: 1757.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:36<00:43,  1.43it/s, est. speed input: 165.38 toks/s, output: 1843.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:36<00:38,  1.59it/s, est. speed input: 165.13 toks/s, output: 1927.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:37<00:31,  1.88it/s, est. speed input: 165.53 toks/s, output: 2018.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:37<00:27,  2.12it/s, est. speed input: 165.74 toks/s, output: 2106.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:38<00:39,  1.46it/s, est. speed input: 162.08 toks/s, output: 2143.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:38<00:31,  1.80it/s, est. speed input: 162.67 toks/s, output: 2235.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:39<00:30,  1.81it/s, est. speed input: 162.05 toks/s, output: 2309.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:57:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6417.6 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:40<00:34,  1.59it/s, est. speed input: 160.24 toks/s, output: 2367.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:41<00:43,  1.23it/s, est. speed input: 156.97 toks/s, output: 2401.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:41<00:31,  1.66it/s, est. speed input: 158.09 toks/s, output: 2500.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:42<00:35,  1.46it/s, est. speed input: 156.27 toks/s, output: 2554.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:42<00:28,  1.81it/s, est. speed input: 156.83 toks/s, output: 2644.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:43<00:27,  1.83it/s, est. speed input: 156.34 toks/s, output: 2717.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:44<00:22,  2.13it/s, est. speed input: 156.46 toks/s, output: 2879.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:44<00:20,  2.30it/s, est. speed input: 156.73 toks/s, output: 2962.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:44<00:21,  2.17it/s, est. speed input: 156.24 toks/s, output: 3033.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:45<00:18,  2.39it/s, est. speed input: 156.52 toks/s, output: 3118.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:45<00:13,  3.08it/s, est. speed input: 157.73 toks/s, output: 3300.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:45<00:12,  3.27it/s, est. speed input: 158.31 toks/s, output: 3389.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:46<00:16,  2.44it/s, est. speed input: 157.19 toks/s, output: 3442.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:49<00:45,  1.13s/it, est. speed input: 148.48 toks/s, output: 3332.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:57:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5328.2 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:50<00:42,  1.09s/it, est. speed input: 146.85 toks/s, output: 3375.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:53<00:59,  1.58s/it, est. speed input: 140.30 toks/s, output: 3305.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:54<00:47,  1.30s/it, est. speed input: 139.90 toks/s, output: 3375.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:54<00:40,  1.11s/it, est. speed input: 139.29 toks/s, output: 3441.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:54<00:29,  1.19it/s, est. speed input: 139.91 toks/s, output: 3536.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:55<00:26,  1.29it/s, est. speed input: 139.43 toks/s, output: 3604.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:56<00:23,  1.43it/s, est. speed input: 139.25 toks/s, output: 3679.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:57<00:33,  1.06s/it, est. speed input: 135.69 toks/s, output: 3666.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:58<00:27,  1.11it/s, est. speed input: 135.54 toks/s, output: 3741.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:58<00:20,  1.44it/s, est. speed input: 136.11 toks/s, output: 3836.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:57:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4185.0 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:00<00:21,  1.31it/s, est. speed input: 134.34 toks/s, output: 3944.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:00<00:18,  1.46it/s, est. speed input: 134.38 toks/s, output: 4024.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:01<00:19,  1.30it/s, est. speed input: 133.15 toks/s, output: 4069.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:04<00:21,  1.11it/s, est. speed input: 130.59 toks/s, output: 4149.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:04<00:19,  1.19it/s, est. speed input: 130.23 toks/s, output: 4218.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:04<00:15,  1.41it/s, est. speed input: 130.55 toks/s, output: 4309.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:05<00:12,  1.67it/s, est. speed input: 130.89 toks/s, output: 4401.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:06<00:10,  1.81it/s, est. speed input: 130.75 toks/s, output: 4558.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:06<00:10,  1.72it/s, est. speed input: 130.32 toks/s, output: 4624.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:07<00:09,  1.85it/s, est. speed input: 130.40 toks/s, output: 4708.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:08<00:11,  1.36it/s, est. speed input: 128.91 toks/s, output: 4734.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:08<00:08,  1.73it/s, est. speed input: 129.47 toks/s, output: 4836.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:09<00:05,  2.29it/s, est. speed input: 130.28 toks/s, output: 5027.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:09<00:05,  2.09it/s, est. speed input: 130.02 toks/s, output: 5098.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 03:58:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3080.0 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:10<00:06,  1.61it/s, est. speed input: 128.96 toks/s, output: 5138.35 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  1.61it/s, est. speed input: 138.26 toks/s, output: 6408.61 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  2.26it/s, est. speed input: 138.26 toks/s, output: 6408.61 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:58:09 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 03:58:09 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 03:58:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:58:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:58:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:58:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 03:58:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 03:58:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 03:58:09 [abstract.py:306] It took 0.627635 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 03:58:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 03:58:09 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0124
+  • Training Accuracy: 0.2750
+  • Mean Generation Length: 2841.3501
+  • Training FLOPS: 51.07 TFLOPS (12.77 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 323.24s
+  • policy_training: 203.83s (63.1%)
+  • generation: 75.69s (23.4%)
+  • teacher_logprob_inference: 38.73s (12.0%)
+  • policy_training/sharding_data: 4.31s (1.3%)
+  • prepare_for_generation: 3.45s (1.1%)
+  • policy_training/submit_training_futures: 1.87s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.5%)
+  • training_prep: 0.84s (0.3%)
+  • teacher_logprob_inference_prep: 0.50s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 63: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:14 [abstract.py:324] It took 0.049653 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07263125e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58436890e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74340134e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07263125e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58436890e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74340134e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07263125e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58436890e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74340134e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07263125e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58436890e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74340134e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07268086e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57483215e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73686829e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.21075554e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.39324951e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #63] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44864700e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07268086e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57483215e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73686829e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67426758e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30280504e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #63] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.32869186e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07268086e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57483215e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73686829e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29542847e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.43864918e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #63] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13085175e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07268086e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57483215e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.73686829e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27047997e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.60668774e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #63] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31602432e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 63: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:17 [abstract.py:324] It took 0.030183 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13333.50it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:17 [loggers.py:236] Engine 000: Avg prompt throughput: 38.4 tokens/s, Avg generation throughput: 7.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:40,  3.89it/s, est. speed input: 244.84 toks/s, output: 58.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:07, 19.95it/s, est. speed input: 1033.41 toks/s, output: 333.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 28.79it/s, est. speed input: 1441.73 toks/s, output: 574.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:04, 29.02it/s, est. speed input: 1528.51 toks/s, output: 757.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 40.15it/s, est. speed input: 1928.67 toks/s, output: 1294.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:04, 30.38it/s, est. speed input: 1745.31 toks/s, output: 1362.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 38.26it/s, est. speed input: 1954.54 toks/s, output: 1850.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 40.31it/s, est. speed input: 2034.31 toks/s, output: 2140.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 44.37it/s, est. speed input: 2138.31 toks/s, output: 2513.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 45.86it/s, est. speed input: 2210.00 toks/s, output: 2859.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:02, 46.88it/s, est. speed input: 2276.04 toks/s, output: 3252.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 36.08it/s, est. speed input: 2157.61 toks/s, output: 3319.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:02<00:04, 20.02it/s, est. speed input: 1793.05 toks/s, output: 3049.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:05<00:18,  4.50it/s, est. speed input: 826.00 toks/s, output: 1667.53 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:07<00:26,  3.09it/s, est. speed input: 616.37 toks/s, output: 1499.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10603.4 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:12<00:49,  1.60it/s, est. speed input: 403.72 toks/s, output: 1180.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:13<00:52,  1.48it/s, est. speed input: 359.61 toks/s, output: 1244.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:14<00:50,  1.52it/s, est. speed input: 351.11 toks/s, output: 1313.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<00:44,  1.72it/s, est. speed input: 351.89 toks/s, output: 1414.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:14<00:40,  1.84it/s, est. speed input: 347.66 toks/s, output: 1493.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:17<01:09,  1.06it/s, est. speed input: 300.01 toks/s, output: 1385.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:18<01:06,  1.09it/s, est. speed input: 289.94 toks/s, output: 1435.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:18<00:56,  1.28it/s, est. speed input: 287.96 toks/s, output: 1520.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:19<01:02,  1.14it/s, est. speed input: 273.93 toks/s, output: 1541.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8329.2 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<01:10,  1.01s/it, est. speed input: 258.96 toks/s, output: 1551.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<00:56,  1.23it/s, est. speed input: 258.31 toks/s, output: 1641.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:41,  1.62it/s, est. speed input: 259.58 toks/s, output: 1743.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:22<00:38,  1.72it/s, est. speed input: 256.61 toks/s, output: 1815.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:22<00:39,  1.69it/s, est. speed input: 252.32 toks/s, output: 1875.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:23<00:33,  1.92it/s, est. speed input: 251.04 toks/s, output: 1957.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:23<00:39,  1.62it/s, est. speed input: 244.59 toks/s, output: 1998.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:24<00:30,  2.09it/s, est. speed input: 245.54 toks/s, output: 2096.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:24<00:33,  1.87it/s, est. speed input: 241.38 toks/s, output: 2150.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:24<00:25,  2.40it/s, est. speed input: 242.58 toks/s, output: 2249.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:25<00:23,  2.60it/s, est. speed input: 242.04 toks/s, output: 2333.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:26<00:35,  1.65it/s, est. speed input: 234.13 toks/s, output: 2344.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:26<00:32,  1.79it/s, est. speed input: 232.39 toks/s, output: 2415.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:26<00:24,  2.30it/s, est. speed input: 233.34 toks/s, output: 2513.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:26<00:18,  2.99it/s, est. speed input: 234.72 toks/s, output: 2614.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:28<00:37,  1.48it/s, est. speed input: 224.73 toks/s, output: 2590.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:29<00:41,  1.29it/s, est. speed input: 219.01 toks/s, output: 2611.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:30<00:37,  1.40it/s, est. speed input: 216.80 toks/s, output: 2673.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6796.0 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:30,  1.71it/s, est. speed input: 216.76 toks/s, output: 2759.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:30<00:20,  2.41it/s, est. speed input: 217.52 toks/s, output: 2941.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:33,  1.43it/s, est. speed input: 206.18 toks/s, output: 2955.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:26,  1.76it/s, est. speed input: 207.33 toks/s, output: 3058.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:25,  1.78it/s, est. speed input: 205.75 toks/s, output: 3121.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:34<00:27,  1.63it/s, est. speed input: 202.98 toks/s, output: 3164.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:36<00:43,  1.01it/s, est. speed input: 193.35 toks/s, output: 3100.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:37<00:39,  1.09it/s, est. speed input: 191.27 toks/s, output: 3154.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:37<00:22,  1.84it/s, est. speed input: 193.70 toks/s, output: 3367.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:14,  2.67it/s, est. speed input: 195.83 toks/s, output: 3573.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:38<00:17,  2.18it/s, est. speed input: 193.57 toks/s, output: 3615.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:38<00:14,  2.61it/s, est. speed input: 194.45 toks/s, output: 3715.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:39<00:14,  2.37it/s, est. speed input: 192.90 toks/s, output: 3849.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:02:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5401.7 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:40<00:17,  1.99it/s, est. speed input: 190.59 toks/s, output: 3888.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:40<00:17,  1.93it/s, est. speed input: 189.40 toks/s, output: 3948.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:40<00:13,  2.32it/s, est. speed input: 190.04 toks/s, output: 4045.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:41<00:16,  1.92it/s, est. speed input: 187.95 toks/s, output: 4085.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:42<00:14,  2.08it/s, est. speed input: 187.75 toks/s, output: 4164.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:42<00:13,  2.17it/s, est. speed input: 187.41 toks/s, output: 4239.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:42<00:12,  2.27it/s, est. speed input: 187.18 toks/s, output: 4316.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:42<00:09,  2.91it/s, est. speed input: 188.08 toks/s, output: 4421.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:43<00:08,  2.80it/s, est. speed input: 187.61 toks/s, output: 4575.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:43<00:07,  3.15it/s, est. speed input: 188.17 toks/s, output: 4671.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:44<00:08,  2.63it/s, est. speed input: 187.13 toks/s, output: 4729.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:45<00:11,  1.92it/s, est. speed input: 184.73 toks/s, output: 4753.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:45<00:09,  2.18it/s, est. speed input: 184.85 toks/s, output: 4840.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:45<00:08,  2.46it/s, est. speed input: 185.08 toks/s, output: 4929.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:46<00:08,  2.33it/s, est. speed input: 184.45 toks/s, output: 4996.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:46<00:06,  2.86it/s, est. speed input: 185.18 toks/s, output: 5098.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:46<00:04,  3.58it/s, est. speed input: 186.10 toks/s, output: 5205.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:46<00:03,  4.15it/s, est. speed input: 186.77 toks/s, output: 5307.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:47<00:03,  4.42it/s, est. speed input: 187.34 toks/s, output: 5405.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:48<00:06,  2.14it/s, est. speed input: 184.61 toks/s, output: 5409.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:48<00:04,  2.98it/s, est. speed input: 185.72 toks/s, output: 5610.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:03:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3465.7 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:50<00:08,  1.32it/s, est. speed input: 179.48 toks/s, output: 5509.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:51<00:07,  1.28it/s, est. speed input: 177.61 toks/s, output: 5541.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:51<00:06,  1.49it/s, est. speed input: 177.50 toks/s, output: 5627.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:52<00:06,  1.31it/s, est. speed input: 175.31 toks/s, output: 5648.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:09,  1.34s/it, est. speed input: 167.63 toks/s, output: 5496.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:08,  1.49s/it, est. speed input: 163.30 toks/s, output: 5453.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:57<00:03,  1.18it/s, est. speed input: 165.00 toks/s, output: 5707.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:59<00:03,  1.04s/it, est. speed input: 161.43 toks/s, output: 5684.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:59<00:01,  1.15it/s, est. speed input: 161.46 toks/s, output: 5786.64 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  1.15it/s, est. speed input: 163.44 toks/s, output: 6061.90 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  2.69it/s, est. speed input: 163.44 toks/s, output: 6061.90 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:03:20 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:03:20 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:03:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:03:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:03:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:03:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:03:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 04:03:20 [abstract.py:306] It took 0.555408 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:03:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:03:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:03:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0134
+  • Training Accuracy: 0.3250
+  • Mean Generation Length: 2255.0249
+  • Training FLOPS: 44.04 TFLOPS (11.01 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 284.81s
+  • policy_training: 183.54s (64.4%)
+  • generation: 63.36s (22.2%)
+  • teacher_logprob_inference: 33.12s (11.6%)
+  • policy_training/sharding_data: 4.41s (1.5%)
+  • prepare_for_generation: 3.41s (1.2%)
+  • policy_training/submit_training_futures: 1.93s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.50s (0.5%)
+  • training_prep: 0.78s (0.3%)
+  • teacher_logprob_inference_prep: 0.43s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 64: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:06:59 [abstract.py:324] It took 0.051738 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07245352e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59541245e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74832077e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07245352e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59541245e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74832077e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07245352e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59541245e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74832077e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07245352e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59541245e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.74832077e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07254141e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59930496e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.72579193e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20809059e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.46149540e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #64] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44880371e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07254141e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59930496e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.72579193e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70787811e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30556850e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #64] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.33998756e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07254141e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59930496e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.72579193e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.29520569e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.25036287e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #64] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11487846e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07254141e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59930496e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.72579193e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27076721e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.60559540e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #64] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31122955e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 64: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:02 [abstract.py:324] It took 0.036418 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13832.32it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:02 [loggers.py:236] Engine 000: Avg prompt throughput: 41.2 tokens/s, Avg generation throughput: 62.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:46,  3.41it/s, est. speed input: 211.46 toks/s, output: 51.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:10, 15.12it/s, est. speed input: 772.16 toks/s, output: 248.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:07, 19.61it/s, est. speed input: 974.00 toks/s, output: 384.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 28.78it/s, est. speed input: 1317.39 toks/s, output: 652.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 31.34it/s, est. speed input: 1480.48 toks/s, output: 873.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 32.86it/s, est. speed input: 1578.17 toks/s, output: 1059.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:04, 33.17it/s, est. speed input: 1637.23 toks/s, output: 1259.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:05, 25.76it/s, est. speed input: 1523.23 toks/s, output: 1333.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:05, 23.64it/s, est. speed input: 1479.49 toks/s, output: 1443.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:08, 15.33it/s, est. speed input: 1256.25 toks/s, output: 1390.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:07, 16.01it/s, est. speed input: 1242.30 toks/s, output: 1549.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:02<00:07, 15.71it/s, est. speed input: 1219.67 toks/s, output: 1638.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:02<00:07, 15.19it/s, est. speed input: 1192.94 toks/s, output: 1721.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:02<00:07, 14.57it/s, est. speed input: 1167.38 toks/s, output: 1802.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:03<00:23,  4.85it/s, est. speed input: 808.34 toks/s, output: 1363.85 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:07<01:18,  1.41it/s, est. speed input: 388.95 toks/s, output: 782.71 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12096.3 tokens/s, Running: 111 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:10<01:49,  1.00it/s, est. speed input: 299.69 toks/s, output: 693.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:11<01:47,  1.02it/s, est. speed input: 280.42 toks/s, output: 738.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:11<01:33,  1.15it/s, est. speed input: 275.17 toks/s, output: 812.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:11<01:15,  1.41it/s, est. speed input: 276.12 toks/s, output: 901.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:14<02:01,  1.15s/it, est. speed input: 232.19 toks/s, output: 845.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:14<01:44,  1.00it/s, est. speed input: 227.13 toks/s, output: 912.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:15<01:26,  1.20it/s, est. speed input: 225.10 toks/s, output: 989.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:16<01:54,  1.11s/it, est. speed input: 204.34 toks/s, output: 981.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:17<01:26,  1.18it/s, est. speed input: 205.63 toks/s, output: 1070.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:17<01:17,  1.30it/s, est. speed input: 202.43 toks/s, output: 1135.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:18<01:04,  1.54it/s, est. speed input: 201.75 toks/s, output: 1211.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:18<01:01,  1.60it/s, est. speed input: 198.92 toks/s, output: 1274.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9921.1 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:21<01:28,  1.10it/s, est. speed input: 181.31 toks/s, output: 1309.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:21<01:16,  1.26it/s, est. speed input: 180.42 toks/s, output: 1381.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:22<01:15,  1.27it/s, est. speed input: 176.92 toks/s, output: 1430.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:22<00:44,  2.10it/s, est. speed input: 181.58 toks/s, output: 1619.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:23<00:52,  1.76it/s, est. speed input: 177.33 toks/s, output: 1656.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:25<01:17,  1.16it/s, est. speed input: 164.39 toks/s, output: 1676.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:26<01:17,  1.15it/s, est. speed input: 161.03 toks/s, output: 1716.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:29<01:58,  1.34s/it, est. speed input: 147.74 toks/s, output: 1649.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:29<01:30,  1.04s/it, est. speed input: 149.10 toks/s, output: 1736.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8274.8 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:30<01:03,  1.34it/s, est. speed input: 149.67 toks/s, output: 1886.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:32<01:30,  1.08s/it, est. speed input: 141.47 toks/s, output: 1856.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:33<01:15,  1.10it/s, est. speed input: 141.56 toks/s, output: 1928.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:33<01:05,  1.26it/s, est. speed input: 141.45 toks/s, output: 1997.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:33<00:49,  1.65it/s, est. speed input: 142.81 toks/s, output: 2086.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:33<00:37,  2.14it/s, est. speed input: 144.18 toks/s, output: 2175.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:34<00:31,  2.47it/s, est. speed input: 144.89 toks/s, output: 2254.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:35<00:46,  1.68it/s, est. speed input: 142.21 toks/s, output: 2281.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:35<00:37,  2.04it/s, est. speed input: 142.96 toks/s, output: 2360.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:36<00:48,  1.57it/s, est. speed input: 140.72 toks/s, output: 2391.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:36<00:37,  1.97it/s, est. speed input: 141.61 toks/s, output: 2473.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:37<00:49,  1.49it/s, est. speed input: 139.25 toks/s, output: 2497.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:37<00:32,  2.23it/s, est. speed input: 141.00 toks/s, output: 2661.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:38<00:33,  2.11it/s, est. speed input: 140.59 toks/s, output: 2717.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:39<00:47,  1.46it/s, est. speed input: 137.64 toks/s, output: 2725.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:39<00:37,  1.86it/s, est. speed input: 138.59 toks/s, output: 2809.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:40<00:22,  2.93it/s, est. speed input: 140.96 toks/s, output: 2985.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7051.9 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:40<00:22,  2.99it/s, est. speed input: 141.31 toks/s, output: 3056.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:40<00:22,  2.88it/s, est. speed input: 141.47 toks/s, output: 3122.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:41<00:32,  1.96it/s, est. speed input: 139.65 toks/s, output: 3145.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:24,  2.54it/s, est. speed input: 140.97 toks/s, output: 3297.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:43<00:34,  1.76it/s, est. speed input: 138.70 toks/s, output: 3306.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:44<00:47,  1.26it/s, est. speed input: 135.56 toks/s, output: 3294.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:45<00:46,  1.26it/s, est. speed input: 134.53 toks/s, output: 3332.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:46<00:48,  1.20it/s, est. speed input: 133.10 toks/s, output: 3360.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:46<00:39,  1.43it/s, est. speed input: 133.37 toks/s, output: 3430.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:47<00:24,  2.24it/s, est. speed input: 135.12 toks/s, output: 3600.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:48<00:34,  1.58it/s, est. speed input: 132.97 toks/s, output: 3604.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:48<00:29,  1.82it/s, est. speed input: 133.39 toks/s, output: 3678.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:49<00:29,  1.76it/s, est. speed input: 132.93 toks/s, output: 3727.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:49<00:22,  2.24it/s, est. speed input: 133.81 toks/s, output: 3813.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:49<00:22,  2.18it/s, est. speed input: 133.66 toks/s, output: 3871.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:07:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6007.3 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:50<00:19,  2.57it/s, est. speed input: 134.31 toks/s, output: 3950.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:50<00:11,  4.00it/s, est. speed input: 136.31 toks/s, output: 4129.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:51<00:20,  2.21it/s, est. speed input: 134.60 toks/s, output: 4139.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:52<00:32,  1.39it/s, est. speed input: 131.92 toks/s, output: 4119.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:53<00:25,  1.71it/s, est. speed input: 132.53 toks/s, output: 4199.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:54<00:33,  1.30it/s, est. speed input: 130.56 toks/s, output: 4199.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:54<00:28,  1.47it/s, est. speed input: 130.60 toks/s, output: 4262.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:58<00:44,  1.11s/it, est. speed input: 125.37 toks/s, output: 4213.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:58<00:39,  1.02s/it, est. speed input: 124.79 toks/s, output: 4259.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:59<00:31,  1.20it/s, est. speed input: 125.22 toks/s, output: 4337.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:59<00:25,  1.48it/s, est. speed input: 125.66 toks/s, output: 4417.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:59<00:14,  2.38it/s, est. speed input: 127.38 toks/s, output: 4602.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:08:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4910.3 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:00<00:18,  1.80it/s, est. speed input: 126.27 toks/s, output: 4626.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:00<00:15,  2.20it/s, est. speed input: 126.93 toks/s, output: 4713.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:01<00:13,  2.33it/s, est. speed input: 127.25 toks/s, output: 4851.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:01<00:11,  2.58it/s, est. speed input: 127.71 toks/s, output: 4931.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:02<00:10,  2.65it/s, est. speed input: 127.99 toks/s, output: 5003.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:02<00:11,  2.36it/s, est. speed input: 127.82 toks/s, output: 5060.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:03<00:13,  1.90it/s, est. speed input: 127.04 toks/s, output: 5154.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:05<00:16,  1.55it/s, est. speed input: 125.93 toks/s, output: 5174.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:05<00:10,  2.15it/s, est. speed input: 127.07 toks/s, output: 5347.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:05<00:07,  2.87it/s, est. speed input: 128.28 toks/s, output: 5527.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:06<00:06,  2.91it/s, est. speed input: 128.54 toks/s, output: 5602.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:06<00:08,  2.19it/s, est. speed input: 127.86 toks/s, output: 5635.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:07<00:10,  1.78it/s, est. speed input: 127.10 toks/s, output: 5664.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:08<00:08,  2.02it/s, est. speed input: 127.40 toks/s, output: 5742.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:09<00:09,  1.61it/s, est. speed input: 126.51 toks/s, output: 5765.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:08:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3261.8 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:10<00:11,  1.26it/s, est. speed input: 125.16 toks/s, output: 5768.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:11<00:13,  1.06it/s, est. speed input: 123.70 toks/s, output: 5768.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:11<00:09,  1.34it/s, est. speed input: 124.10 toks/s, output: 5854.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:12<00:07,  1.57it/s, est. speed input: 124.28 toks/s, output: 5930.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:12<00:06,  1.73it/s, est. speed input: 124.34 toks/s, output: 6002.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:12<00:04,  2.16it/s, est. speed input: 124.87 toks/s, output: 6094.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:13<00:03,  2.45it/s, est. speed input: 125.24 toks/s, output: 6178.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:13<00:03,  2.17it/s, est. speed input: 125.07 toks/s, output: 6238.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:14<00:03,  1.95it/s, est. speed input: 124.83 toks/s, output: 6294.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:14<00:02,  2.27it/s, est. speed input: 125.17 toks/s, output: 6381.35 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  2.27it/s, est. speed input: 130.10 toks/s, output: 7040.31 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  2.15it/s, est. speed input: 130.10 toks/s, output: 7040.31 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:08:22 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:08:22 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:08:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:08:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:08:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:08:22 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:08:22 [abstract.py:306] It took 0.550652 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:08:22 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:08:22 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:08:22 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:08:22 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.2437
+  • Mean Generation Length: 3282.0188
+  • Training FLOPS: 51.11 TFLOPS (12.78 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 361.44s
+  • policy_training: 231.90s (64.2%)
+  • generation: 79.89s (22.1%)
+  • teacher_logprob_inference: 44.29s (12.3%)
+  • policy_training/sharding_data: 4.29s (1.2%)
+  • prepare_for_generation: 3.54s (1.0%)
+  • policy_training/submit_training_futures: 1.66s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.52s (0.4%)
+  • training_prep: 1.23s (0.3%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 65: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:01 [abstract.py:324] It took 0.049958 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07275781e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07275781e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60215454e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07275781e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60215454e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07275781e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60215454e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60215454e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75352783e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75352783e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75352783e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75352783e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07284297e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59884224e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74529190e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07284297e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59884224e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74529190e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07284297e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59884224e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74529190e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07284297e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59884224e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.74529190e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20400085e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.54351091e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #65] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44985214e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70225449e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30159149e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #65] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.32037354e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31223755e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-2.80568600e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #65] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.12177429e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27400093e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59482174e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #65] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31205963e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) 
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) 
+[DEBUG refit] Step 65: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:04 [abstract.py:324] It took 0.031212 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:04 [loggers.py:236] Engine 000: Avg prompt throughput: 33.2 tokens/s, Avg generation throughput: 33.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13046.55it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:24,  6.61it/s, est. speed input: 409.66 toks/s, output: 46.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:12, 12.24it/s, est. speed input: 684.28 toks/s, output: 146.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:04, 34.31it/s, est. speed input: 1632.25 toks/s, output: 553.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 32.56it/s, est. speed input: 1691.03 toks/s, output: 691.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 35.78it/s, est. speed input: 1849.92 toks/s, output: 941.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:02, 50.95it/s, est. speed input: 2353.37 toks/s, output: 1566.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:00<00:01, 67.12it/s, est. speed input: 2858.97 toks/s, output: 2297.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:00<00:01, 82.52it/s, est. speed input: 3376.48 toks/s, output: 3210.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:00, 96.29it/s, est. speed input: 3817.33 toks/s, output: 4122.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:01<00:00, 92.80it/s, est. speed input: 3946.07 toks/s, output: 4654.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:01<00:01, 67.44it/s, est. speed input: 3689.45 toks/s, output: 4765.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:03<00:04, 15.57it/s, est. speed input: 1847.62 toks/s, output: 2763.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:07<00:11,  4.90it/s, est. speed input: 822.35 toks/s, output: 1734.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:08<00:11,  4.70it/s, est. speed input: 752.20 toks/s, output: 2059.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9278.7 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:12<00:20,  2.49it/s, est. speed input: 512.49 toks/s, output: 1736.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:13<00:18,  2.58it/s, est. speed input: 489.55 toks/s, output: 2024.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:14<00:18,  2.47it/s, est. speed input: 464.33 toks/s, output: 2159.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:16<00:20,  2.20it/s, est. speed input: 431.90 toks/s, output: 2247.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:16<00:18,  2.34it/s, est. speed input: 430.34 toks/s, output: 2357.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:17<00:21,  1.96it/s, est. speed input: 408.19 toks/s, output: 2357.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:18<00:21,  1.88it/s, est. speed input: 397.18 toks/s, output: 2412.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:19<00:30,  1.30it/s, est. speed input: 364.45 toks/s, output: 2334.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6510.8 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:20<00:25,  1.56it/s, est. speed input: 364.41 toks/s, output: 2453.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:21<00:26,  1.41it/s, est. speed input: 351.14 toks/s, output: 2484.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:21<00:27,  1.34it/s, est. speed input: 339.96 toks/s, output: 2524.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:22<00:14,  2.29it/s, est. speed input: 341.24 toks/s, output: 2888.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:24<00:26,  1.23it/s, est. speed input: 311.11 toks/s, output: 2752.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:26<00:31,  1.01it/s, est. speed input: 294.28 toks/s, output: 2724.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:27<00:30,  1.03it/s, est. speed input: 286.63 toks/s, output: 2773.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:27<00:24,  1.22it/s, est. speed input: 285.09 toks/s, output: 2877.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:27<00:19,  1.45it/s, est. speed input: 283.86 toks/s, output: 2983.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:28<00:14,  1.88it/s, est. speed input: 284.80 toks/s, output: 3110.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:28<00:11,  2.37it/s, est. speed input: 285.57 toks/s, output: 3235.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:28<00:11,  2.27it/s, est. speed input: 282.70 toks/s, output: 3321.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:29<00:14,  1.68it/s, est. speed input: 275.54 toks/s, output: 3353.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:29<00:12,  1.96it/s, est. speed input: 274.80 toks/s, output: 3461.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4777.6 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:30<00:09,  2.30it/s, est. speed input: 274.46 toks/s, output: 3573.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:30<00:06,  3.05it/s, est. speed input: 274.69 toks/s, output: 3808.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:31<00:07,  2.61it/s, est. speed input: 271.70 toks/s, output: 3883.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:31<00:06,  3.00it/s, est. speed input: 272.00 toks/s, output: 4001.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:32<00:05,  2.85it/s, est. speed input: 269.49 toks/s, output: 4191.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:32<00:05,  3.10it/s, est. speed input: 269.51 toks/s, output: 4305.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:33<00:06,  2.19it/s, est. speed input: 264.33 toks/s, output: 4339.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:34<00:09,  1.55it/s, est. speed input: 257.10 toks/s, output: 4337.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:35<00:10,  1.24it/s, est. speed input: 249.94 toks/s, output: 4336.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:35<00:07,  1.58it/s, est. speed input: 250.30 toks/s, output: 4461.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:36<00:06,  1.65it/s, est. speed input: 248.18 toks/s, output: 4542.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:36<00:03,  2.61it/s, est. speed input: 249.93 toks/s, output: 4813.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:37<00:03,  2.17it/s, est. speed input: 246.79 toks/s, output: 4873.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:38<00:04,  1.53it/s, est. speed input: 240.56 toks/s, output: 4872.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:38<00:03,  1.68it/s, est. speed input: 239.34 toks/s, output: 4971.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:39<00:02,  1.78it/s, est. speed input: 238.04 toks/s, output: 5066.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2493.2 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:40<00:02,  1.40it/s, est. speed input: 233.12 toks/s, output: 5086.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:41<00:02,  1.15it/s, est. speed input: 227.55 toks/s, output: 5092.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:46<00:04,  2.11s/it, est. speed input: 204.06 toks/s, output: 4708.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:47<00:01,  1.71s/it, est. speed input: 202.03 toks/s, output: 4804.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:47<00:00,  1.71s/it, est. speed input: 203.11 toks/s, output: 4971.12 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:47<00:00,  3.36it/s, est. speed input: 203.11 toks/s, output: 4971.12 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:54 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:54 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:13:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:13:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:13:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:13:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:13:54 [abstract.py:306] It took 0.526632 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:13:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:13:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:13:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:13:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0180
+  • Training Accuracy: 0.3000
+  • Mean Generation Length: 1480.7063
+  • Training FLOPS: 38.00 TFLOPS (9.50 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 220.23s
+  • policy_training: 140.04s (63.6%)
+  • generation: 50.37s (22.9%)
+  • teacher_logprob_inference: 24.84s (11.3%)
+  • policy_training/sharding_data: 4.34s (2.0%)
+  • prepare_for_generation: 3.52s (1.6%)
+  • policy_training/submit_training_futures: 1.96s (0.9%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.7%)
+  • training_prep: 0.84s (0.4%)
+  • teacher_logprob_inference_prep: 0.44s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 14/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 66: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:16:42 [abstract.py:324] It took 0.173691 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07301074e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60733719e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76096191e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07301074e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60733719e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76096191e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07301074e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60733719e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76096191e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07301074e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60733719e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76096191e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07307480e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59582939e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76748962e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19112892e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.61982822e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #66] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44809418e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07307480e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59582939e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76748962e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67991562e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30887489e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #66] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.31902428e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07307480e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59582939e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76748962e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.31419220e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.54295969e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #66] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11382675e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07307480e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59582939e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76748962e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27217293e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59295998e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #66] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31408218e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 66: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:16:44 [abstract.py:324] It took 0.029786 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13503.06it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:16:45 [loggers.py:236] Engine 000: Avg prompt throughput: 53.7 tokens/s, Avg generation throughput: 34.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:03,  2.50it/s, est. speed input: 157.46 toks/s, output: 32.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:08, 18.77it/s, est. speed input: 937.23 toks/s, output: 284.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 25.85it/s, est. speed input: 1237.04 toks/s, output: 472.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 28.59it/s, est. speed input: 1378.73 toks/s, output: 629.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 37.22it/s, est. speed input: 1688.87 toks/s, output: 991.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:02, 49.64it/s, est. speed input: 2077.08 toks/s, output: 1532.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:01, 72.70it/s, est. speed input: 2670.27 toks/s, output: 2435.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 79.62it/s, est. speed input: 2962.71 toks/s, output: 3046.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 71.01it/s, est. speed input: 3009.04 toks/s, output: 3398.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:01<00:00, 84.71it/s, est. speed input: 3350.84 toks/s, output: 4320.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:01<00:00, 84.45it/s, est. speed input: 3466.67 toks/s, output: 4816.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:01<00:00, 79.70it/s, est. speed input: 3520.37 toks/s, output: 5268.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:08<00:10,  4.53it/s, est. speed input: 791.09 toks/s, output: 1504.78 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 04:16:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8363.3 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:14<00:16,  2.61it/s, est. speed input: 492.40 toks/s, output: 1621.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:16<00:14,  2.63it/s, est. speed input: 454.98 toks/s, output: 2123.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:19<00:16,  2.10it/s, est. speed input: 392.26 toks/s, output: 2207.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:17:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5911.4 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:21<00:14,  2.09it/s, est. speed input: 373.33 toks/s, output: 2478.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:22<00:15,  1.91it/s, est. speed input: 352.60 toks/s, output: 2594.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:23<00:13,  1.99it/s, est. speed input: 345.75 toks/s, output: 2795.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:23<00:12,  2.16it/s, est. speed input: 346.36 toks/s, output: 2925.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:24<00:10,  2.34it/s, est. speed input: 342.34 toks/s, output: 3138.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:24<00:10,  2.14it/s, est. speed input: 335.14 toks/s, output: 3197.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:25<00:11,  1.94it/s, est. speed input: 327.69 toks/s, output: 3251.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:25<00:09,  2.20it/s, est. speed input: 327.40 toks/s, output: 3373.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:27<00:12,  1.54it/s, est. speed input: 313.47 toks/s, output: 3356.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:27<00:09,  1.93it/s, est. speed input: 314.50 toks/s, output: 3493.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:29<00:11,  1.50it/s, est. speed input: 299.48 toks/s, output: 3571.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:29<00:10,  1.50it/s, est. speed input: 294.65 toks/s, output: 3642.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:29<00:08,  1.80it/s, est. speed input: 294.66 toks/s, output: 3769.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:30<00:06,  2.19it/s, est. speed input: 295.02 toks/s, output: 3900.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:17:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3891.2 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:30<00:05,  2.28it/s, est. speed input: 293.22 toks/s, output: 4002.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:31<00:05,  2.09it/s, est. speed input: 289.66 toks/s, output: 4080.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:31<00:02,  3.44it/s, est. speed input: 292.58 toks/s, output: 4373.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:31<00:02,  3.03it/s, est. speed input: 290.32 toks/s, output: 4465.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:32<00:02,  2.68it/s, est. speed input: 287.62 toks/s, output: 4551.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:33<00:05,  1.30it/s, est. speed input: 273.76 toks/s, output: 4462.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:34<00:04,  1.30it/s, est. speed input: 269.44 toks/s, output: 4524.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:35<00:04,  1.22it/s, est. speed input: 263.87 toks/s, output: 4565.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:35<00:02,  1.50it/s, est. speed input: 263.43 toks/s, output: 4692.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:36<00:01,  1.68it/s, est. speed input: 262.01 toks/s, output: 4802.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:36<00:00,  2.03it/s, est. speed input: 261.90 toks/s, output: 4935.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:17:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1302.6 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.6%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:41<00:01,  1.80s/it, est. speed input: 232.28 toks/s, output: 4527.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:44<00:00,  2.16s/it, est. speed input: 218.05 toks/s, output: 4407.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:44<00:00,  2.16s/it, est. speed input: 218.05 toks/s, output: 4407.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:44<00:00,  3.59it/s, est. speed input: 218.05 toks/s, output: 4407.06 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:17:31 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:17:31 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:17:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:17:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:17:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:17:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:17:31 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:17:31 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:17:31 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 04:17:31 [abstract.py:306] It took 0.543581 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:17:31 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0208
+  • Training Accuracy: 0.4813
+  • Mean Generation Length: 1227.8313
+  • Training FLOPS: 35.10 TFLOPS (8.77 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 198.53s
+  • policy_training: 125.82s (63.4%)
+  • generation: 47.00s (23.7%)
+  • teacher_logprob_inference: 20.51s (10.3%)
+  • policy_training/sharding_data: 4.31s (2.2%)
+  • prepare_for_generation: 3.72s (1.9%)
+  • policy_training/submit_training_futures: 2.14s (1.1%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.8%)
+  • training_prep: 0.86s (0.4%)
+  • teacher_logprob_inference_prep: 0.45s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 67: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:00 [abstract.py:324] It took 0.063872 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07315508e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58345909e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77360077e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07315508e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58345909e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77360077e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07315508e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58345909e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77360077e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07315508e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58345909e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77360077e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07323105e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58051147e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77147217e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20229225e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.73445034e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #67] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45149353e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07323105e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58051147e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77147217e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67728882e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31653194e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #67] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.33762283e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07323105e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58051147e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77147217e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.33211517e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-3.44088078e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #67] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11299591e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07323105e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58051147e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77147217e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29714890e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59237967e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #67] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.31932114e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 67: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:03 [abstract.py:324] It took 0.029944 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13872.63it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:03 [loggers.py:236] Engine 000: Avg prompt throughput: 61.2 tokens/s, Avg generation throughput: 10.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:37,  1.63it/s, est. speed input: 98.08 toks/s, output: 40.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:30,  5.07it/s, est. speed input: 252.76 toks/s, output: 121.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:18,  8.25it/s, est. speed input: 385.68 toks/s, output: 236.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:01<00:14, 10.26it/s, est. speed input: 460.10 toks/s, output: 330.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:01<00:17,  8.77it/s, est. speed input: 451.28 toks/s, output: 398.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:01<00:12, 12.10it/s, est. speed input: 540.71 toks/s, output: 570.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:08, 16.22it/s, est. speed input: 644.62 toks/s, output: 801.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:06, 21.39it/s, est. speed input: 767.48 toks/s, output: 1098.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:04, 32.10it/s, est. speed input: 976.50 toks/s, output: 1621.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 33.55it/s, est. speed input: 1050.54 toks/s, output: 1848.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:02<00:03, 34.89it/s, est. speed input: 1115.19 toks/s, output: 2076.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:02<00:03, 34.48it/s, est. speed input: 1164.20 toks/s, output: 2287.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:02<00:03, 30.94it/s, est. speed input: 1187.56 toks/s, output: 2454.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:02<00:06, 16.59it/s, est. speed input: 1061.82 toks/s, output: 2330.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:06<00:39,  2.73it/s, est. speed input: 466.43 toks/s, output: 1160.80 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:08<00:49,  2.12it/s, est. speed input: 381.08 toks/s, output: 1115.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11988.6 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:10<00:59,  1.73it/s, est. speed input: 322.74 toks/s, output: 1111.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:12<01:13,  1.39it/s, est. speed input: 283.87 toks/s, output: 1061.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:14<01:29,  1.13it/s, est. speed input: 252.49 toks/s, output: 1029.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:15<01:34,  1.06it/s, est. speed input: 236.72 toks/s, output: 1048.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:15<01:19,  1.24it/s, est. speed input: 236.55 toks/s, output: 1131.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:15<01:04,  1.52it/s, est. speed input: 237.82 toks/s, output: 1220.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:16<00:55,  1.76it/s, est. speed input: 237.32 toks/s, output: 1298.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:16<00:43,  2.22it/s, est. speed input: 239.32 toks/s, output: 1389.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:17<01:00,  1.58it/s, est. speed input: 227.33 toks/s, output: 1400.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:19<01:45,  1.12s/it, est. speed input: 203.16 toks/s, output: 1331.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:19<01:17,  1.19it/s, est. speed input: 204.89 toks/s, output: 1421.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9593.9 tokens/s, Running: 92 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:20<01:05,  1.40it/s, est. speed input: 203.76 toks/s, output: 1491.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:21<01:17,  1.18it/s, est. speed input: 195.41 toks/s, output: 1509.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:23<01:45,  1.18s/it, est. speed input: 181.67 toks/s, output: 1481.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:25<02:16,  1.54s/it, est. speed input: 167.12 toks/s, output: 1441.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:26<01:41,  1.15s/it, est. speed input: 167.90 toks/s, output: 1526.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:27<01:34,  1.09s/it, est. speed input: 164.27 toks/s, output: 1571.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:27<01:27,  1.02s/it, est. speed input: 161.37 toks/s, output: 1619.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:28<01:07,  1.26it/s, est. speed input: 162.09 toks/s, output: 1701.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:29<01:18,  1.07it/s, est. speed input: 157.18 toks/s, output: 1725.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8145.3 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:30<01:18,  1.05it/s, est. speed input: 154.02 toks/s, output: 1765.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:30<01:02,  1.32it/s, est. speed input: 154.40 toks/s, output: 1844.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:31<00:59,  1.36it/s, est. speed input: 153.04 toks/s, output: 1901.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:32<00:56,  1.42it/s, est. speed input: 151.84 toks/s, output: 1960.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:32<00:46,  1.70it/s, est. speed input: 152.22 toks/s, output: 2037.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:32<00:37,  2.08it/s, est. speed input: 153.01 toks/s, output: 2119.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:32<00:35,  2.17it/s, est. speed input: 152.90 toks/s, output: 2189.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:33<00:28,  2.67it/s, est. speed input: 153.92 toks/s, output: 2274.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:35<00:46,  1.58it/s, est. speed input: 149.21 toks/s, output: 2340.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:37<01:10,  1.03it/s, est. speed input: 142.70 toks/s, output: 2309.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:37<00:51,  1.39it/s, est. speed input: 143.11 toks/s, output: 2454.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:38<00:47,  1.46it/s, est. speed input: 142.54 toks/s, output: 2513.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:38<00:35,  1.94it/s, est. speed input: 143.76 toks/s, output: 2670.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:39<00:29,  2.25it/s, est. speed input: 144.54 toks/s, output: 2752.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:39<00:19,  3.32it/s, est. speed input: 147.07 toks/s, output: 2933.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:39<00:23,  2.73it/s, est. speed input: 146.42 toks/s, output: 2984.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:40<00:16,  3.67it/s, est. speed input: 148.50 toks/s, output: 3157.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6836.6 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:40<00:24,  2.47it/s, est. speed input: 146.73 toks/s, output: 3184.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:41<00:27,  2.18it/s, est. speed input: 145.92 toks/s, output: 3231.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:41<00:19,  2.97it/s, est. speed input: 147.76 toks/s, output: 3399.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:42<00:14,  3.83it/s, est. speed input: 149.74 toks/s, output: 3570.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:43<00:26,  2.08it/s, est. speed input: 146.57 toks/s, output: 3558.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:43<00:21,  2.53it/s, est. speed input: 147.62 toks/s, output: 3645.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:44<00:17,  2.85it/s, est. speed input: 148.71 toks/s, output: 3857.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:45<00:20,  2.43it/s, est. speed input: 147.88 toks/s, output: 3957.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:45<00:19,  2.41it/s, est. speed input: 147.83 toks/s, output: 4017.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:46<00:18,  2.55it/s, est. speed input: 148.16 toks/s, output: 4086.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:47<00:24,  1.84it/s, est. speed input: 146.18 toks/s, output: 4094.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:47<00:19,  2.30it/s, est. speed input: 147.00 toks/s, output: 4241.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:50<00:35,  1.21it/s, est. speed input: 141.72 toks/s, output: 4153.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:20:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5551.7 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:50<00:32,  1.31it/s, est. speed input: 141.32 toks/s, output: 4205.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:50<00:25,  1.58it/s, est. speed input: 141.81 toks/s, output: 4283.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:51<00:24,  1.65it/s, est. speed input: 141.50 toks/s, output: 4337.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:52<00:26,  1.49it/s, est. speed input: 140.38 toks/s, output: 4367.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:52<00:21,  1.81it/s, est. speed input: 140.84 toks/s, output: 4445.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:52<00:12,  2.81it/s, est. speed input: 142.56 toks/s, output: 4625.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:53<00:12,  2.84it/s, est. speed input: 142.77 toks/s, output: 4695.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:53<00:12,  2.78it/s, est. speed input: 142.91 toks/s, output: 4762.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:54<00:15,  2.09it/s, est. speed input: 141.91 toks/s, output: 4792.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:54<00:13,  2.43it/s, est. speed input: 142.36 toks/s, output: 4871.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:54<00:10,  3.02it/s, est. speed input: 143.14 toks/s, output: 4960.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:55<00:08,  3.50it/s, est. speed input: 144.16 toks/s, output: 5119.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:55<00:07,  3.86it/s, est. speed input: 144.78 toks/s, output: 5204.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:56<00:10,  2.53it/s, est. speed input: 143.80 toks/s, output: 5232.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:56<00:11,  2.33it/s, est. speed input: 143.50 toks/s, output: 5286.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:56<00:09,  2.52it/s, est. speed input: 143.79 toks/s, output: 5358.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:57<00:10,  2.23it/s, est. speed input: 143.43 toks/s, output: 5407.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:57<00:09,  2.32it/s, est. speed input: 143.51 toks/s, output: 5474.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:58<00:05,  3.87it/s, est. speed input: 145.79 toks/s, output: 5748.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:21:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4016.1 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:00<00:13,  1.45it/s, est. speed input: 141.18 toks/s, output: 5633.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:01<00:14,  1.25it/s, est. speed input: 139.50 toks/s, output: 5633.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:01<00:11,  1.51it/s, est. speed input: 139.90 toks/s, output: 5716.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:02<00:09,  1.65it/s, est. speed input: 139.87 toks/s, output: 5783.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:02<00:09,  1.64it/s, est. speed input: 139.47 toks/s, output: 5833.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:03<00:08,  1.70it/s, est. speed input: 139.23 toks/s, output: 5893.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:03<00:05,  2.21it/s, est. speed input: 139.97 toks/s, output: 5991.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:03<00:04,  2.67it/s, est. speed input: 140.54 toks/s, output: 6082.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:04<00:05,  1.99it/s, est. speed input: 139.69 toks/s, output: 6115.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:04<00:04,  2.43it/s, est. speed input: 140.17 toks/s, output: 6207.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:05<00:02,  3.04it/s, est. speed input: 141.05 toks/s, output: 6383.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:05<00:02,  3.15it/s, est. speed input: 141.35 toks/s, output: 6467.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:06<00:03,  1.99it/s, est. speed input: 140.07 toks/s, output: 6480.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:08<00:04,  1.03it/s, est. speed input: 136.38 toks/s, output: 6385.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:09<00:02,  1.41it/s, est. speed input: 136.64 toks/s, output: 6551.47 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  1.41it/s, est. speed input: 139.25 toks/s, output: 6904.68 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  2.30it/s, est. speed input: 139.25 toks/s, output: 6904.68 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:21:17 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:21:17 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:21:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:21:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:21:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:21:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:21:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:21:17 [abstract.py:306] It took 0.550199 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:21:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:21:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:21:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+
+📊 Training Results:
+  • Loss: 0.0148
+  • Training Accuracy: 0.2750
+  • Mean Generation Length: 3002.4563
+  • Training FLOPS: 47.65 TFLOPS (11.91 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 348.57s
+  • policy_training: 225.92s (64.8%)
+  • generation: 74.44s (21.4%)
+  • teacher_logprob_inference: 43.32s (12.4%)
+  • policy_training/sharding_data: 4.42s (1.3%)
+  • prepare_for_generation: 3.44s (1.0%)
+  • policy_training/submit_training_futures: 1.88s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.4%)
+  • training_prep: 0.85s (0.2%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 68: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:25:49 [abstract.py:324] It took 0.054574 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07352793e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57175674e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78252258e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07352793e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57175674e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78252258e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07352793e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57175674e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78252258e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07352793e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57175674e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78252258e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07362148e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58324776e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79218063e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19721813e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.72540140e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #68] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45225113e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07362148e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58324776e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79218063e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67220459e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31963348e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #68] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.34921799e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07362148e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58324776e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79218063e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.32833099e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.50809956e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #68] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.08795090e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07362148e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58324776e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79218063e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30095901e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58253059e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #68] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32146881e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 68: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:25:52 [abstract.py:324] It took 0.031910 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13407.83it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 04:25:52 [loggers.py:236] Engine 000: Avg prompt throughput: 33.7 tokens/s, Avg generation throughput: 65.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:04,  2.45it/s, est. speed input: 149.45 toks/s, output: 53.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:16,  9.48it/s, est. speed input: 474.12 toks/s, output: 222.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 25.30it/s, est. speed input: 1073.63 toks/s, output: 616.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 40.16it/s, est. speed input: 1601.15 toks/s, output: 1061.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 40.30it/s, est. speed input: 1745.16 toks/s, output: 1301.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 38.95it/s, est. speed input: 1817.93 toks/s, output: 1538.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 32.20it/s, est. speed input: 1742.43 toks/s, output: 1666.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:04, 28.40it/s, est. speed input: 1677.75 toks/s, output: 1774.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:03, 32.22it/s, est. speed input: 1758.33 toks/s, output: 2079.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:04, 27.27it/s, est. speed input: 1687.16 toks/s, output: 2174.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:02<00:05, 19.47it/s, est. speed input: 1514.44 toks/s, output: 2151.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:10<01:06,  1.57it/s, est. speed input: 337.34 toks/s, output: 649.98 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 04:26:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11769.3 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:14<01:28,  1.15it/s, est. speed input: 251.63 toks/s, output: 651.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:18<01:49,  1.10s/it, est. speed input: 202.39 toks/s, output: 689.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:18<01:39,  1.01s/it, est. speed input: 201.71 toks/s, output: 776.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:18<01:28,  1.10it/s, est. speed input: 201.41 toks/s, output: 863.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:18<01:14,  1.30it/s, est. speed input: 203.14 toks/s, output: 958.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:19<01:11,  1.34it/s, est. speed input: 199.29 toks/s, output: 1027.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:26:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9739.0 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:20<01:08,  1.39it/s, est. speed input: 196.21 toks/s, output: 1096.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:20<00:55,  1.69it/s, est. speed input: 197.38 toks/s, output: 1186.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:21<00:51,  1.77it/s, est. speed input: 193.38 toks/s, output: 1323.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:21<00:52,  1.75it/s, est. speed input: 190.99 toks/s, output: 1387.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:22<00:42,  2.12it/s, est. speed input: 192.35 toks/s, output: 1477.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:22<00:42,  2.10it/s, est. speed input: 190.89 toks/s, output: 1545.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:23<00:41,  2.12it/s, est. speed input: 189.74 toks/s, output: 1614.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:24<00:52,  1.64it/s, est. speed input: 182.42 toks/s, output: 1703.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:25<00:50,  1.70it/s, est. speed input: 181.08 toks/s, output: 1768.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:25<00:41,  2.04it/s, est. speed input: 181.99 toks/s, output: 1853.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:27<00:58,  1.39it/s, est. speed input: 172.73 toks/s, output: 1904.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:29<01:24,  1.04s/it, est. speed input: 162.42 toks/s, output: 1866.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:26:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8020.6 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:30<01:28,  1.11s/it, est. speed input: 157.51 toks/s, output: 1885.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:31<01:08,  1.15it/s, est. speed input: 158.58 toks/s, output: 1972.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:31<00:56,  1.39it/s, est. speed input: 158.94 toks/s, output: 2050.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:33<01:24,  1.10s/it, est. speed input: 150.96 toks/s, output: 2020.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:33<00:49,  1.53it/s, est. speed input: 153.64 toks/s, output: 2202.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:34<00:47,  1.56it/s, est. speed input: 152.71 toks/s, output: 2261.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:34<00:43,  1.67it/s, est. speed input: 152.40 toks/s, output: 2327.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:35<00:47,  1.51it/s, est. speed input: 150.55 toks/s, output: 2370.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:36<00:51,  1.37it/s, est. speed input: 148.50 toks/s, output: 2408.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:36<00:39,  1.75it/s, est. speed input: 149.50 toks/s, output: 2493.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:36<00:32,  2.15it/s, est. speed input: 150.36 toks/s, output: 2577.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:37<00:26,  2.57it/s, est. speed input: 151.15 toks/s, output: 2659.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:38<00:40,  1.67it/s, est. speed input: 148.37 toks/s, output: 2679.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:38<00:43,  1.51it/s, est. speed input: 146.85 toks/s, output: 2720.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:39<00:36,  1.76it/s, est. speed input: 147.09 toks/s, output: 2793.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:39<00:27,  2.32it/s, est. speed input: 148.23 toks/s, output: 2882.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:39<00:21,  2.94it/s, est. speed input: 149.30 toks/s, output: 2970.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:26:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6734.1 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:40<00:41,  1.50it/s, est. speed input: 145.61 toks/s, output: 2963.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:41<00:35,  1.72it/s, est. speed input: 145.74 toks/s, output: 3032.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:41<00:23,  2.54it/s, est. speed input: 147.37 toks/s, output: 3201.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:42<00:23,  2.42it/s, est. speed input: 147.23 toks/s, output: 3262.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:42<00:17,  3.15it/s, est. speed input: 148.79 toks/s, output: 3429.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:44<00:32,  1.67it/s, est. speed input: 144.98 toks/s, output: 3405.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:44<00:26,  2.02it/s, est. speed input: 145.76 toks/s, output: 3489.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:44<00:26,  2.02it/s, est. speed input: 145.56 toks/s, output: 3548.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:45<00:25,  2.04it/s, est. speed input: 145.40 toks/s, output: 3608.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:45<00:23,  2.18it/s, est. speed input: 145.55 toks/s, output: 3675.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:46<00:30,  1.64it/s, est. speed input: 143.76 toks/s, output: 3695.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:47<00:33,  1.48it/s, est. speed input: 142.48 toks/s, output: 3728.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:48<00:26,  1.74it/s, est. speed input: 142.38 toks/s, output: 3853.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:49<00:27,  1.65it/s, est. speed input: 141.58 toks/s, output: 3897.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:49<00:27,  1.64it/s, est. speed input: 140.99 toks/s, output: 3947.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:49<00:17,  2.49it/s, est. speed input: 142.69 toks/s, output: 4125.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:26:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5547.4 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:50<00:20,  2.01it/s, est. speed input: 141.56 toks/s, output: 4157.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:51<00:18,  2.21it/s, est. speed input: 141.89 toks/s, output: 4230.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:51<00:15,  2.60it/s, est. speed input: 142.55 toks/s, output: 4313.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:51<00:18,  2.15it/s, est. speed input: 141.88 toks/s, output: 4356.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:52<00:15,  2.52it/s, est. speed input: 142.44 toks/s, output: 4436.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:52<00:13,  2.70it/s, est. speed input: 142.78 toks/s, output: 4510.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:52<00:12,  2.87it/s, est. speed input: 143.10 toks/s, output: 4584.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:52<00:09,  3.54it/s, est. speed input: 143.92 toks/s, output: 4673.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:53<00:08,  4.12it/s, est. speed input: 144.71 toks/s, output: 4760.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:53<00:08,  3.76it/s, est. speed input: 144.98 toks/s, output: 4831.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:53<00:08,  3.72it/s, est. speed input: 145.37 toks/s, output: 4906.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:54<00:18,  1.71it/s, est. speed input: 142.96 toks/s, output: 4888.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:55<00:14,  2.05it/s, est. speed input: 143.38 toks/s, output: 4966.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:55<00:12,  2.27it/s, est. speed input: 143.61 toks/s, output: 5038.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:55<00:12,  2.26it/s, est. speed input: 143.55 toks/s, output: 5099.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:56<00:08,  2.99it/s, est. speed input: 144.64 toks/s, output: 5264.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:56<00:06,  3.74it/s, est. speed input: 145.89 toks/s, output: 5437.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:57<00:08,  2.86it/s, est. speed input: 145.33 toks/s, output: 5479.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:58<00:09,  2.19it/s, est. speed input: 144.28 toks/s, output: 5566.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:00<00:13,  1.48it/s, est. speed input: 141.72 toks/s, output: 5534.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:26:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4072.8 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:00<00:12,  1.47it/s, est. speed input: 141.12 toks/s, output: 5578.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:01<00:10,  1.70it/s, est. speed input: 141.37 toks/s, output: 5655.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:02<00:13,  1.24it/s, est. speed input: 139.16 toks/s, output: 5634.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:03<00:12,  1.24it/s, est. speed input: 138.35 toks/s, output: 5671.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:03<00:09,  1.53it/s, est. speed input: 138.76 toks/s, output: 5755.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:03<00:07,  1.77it/s, est. speed input: 139.00 toks/s, output: 5833.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:04<00:06,  2.03it/s, est. speed input: 139.30 toks/s, output: 5913.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:05<00:05,  1.92it/s, est. speed input: 138.77 toks/s, output: 6030.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:05<00:03,  2.53it/s, est. speed input: 139.77 toks/s, output: 6214.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:07<00:05,  1.42it/s, est. speed input: 136.99 toks/s, output: 6160.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:08<00:05,  1.29it/s, est. speed input: 135.86 toks/s, output: 6183.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:27:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2127.8 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:10<00:06,  1.08s/it, est. speed input: 132.91 toks/s, output: 6127.15 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  1.08s/it, est. speed input: 138.14 toks/s, output: 6823.96 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  2.27it/s, est. speed input: 138.14 toks/s, output: 6823.96 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:27:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:27:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:27:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:27:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:27:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:27:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:27:07 [abstract.py:306] It took 0.522369 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:27:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:27:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:27:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:27:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0129
+  • Training Accuracy: 0.2437
+  • Mean Generation Length: 3008.3313
+  • Training FLOPS: 47.73 TFLOPS (11.93 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 350.06s
+  • policy_training: 226.55s (64.7%)
+  • generation: 75.37s (21.5%)
+  • teacher_logprob_inference: 42.98s (12.3%)
+  • policy_training/sharding_data: 4.34s (1.2%)
+  • prepare_for_generation: 3.47s (1.0%)
+  • policy_training/submit_training_futures: 1.63s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.4%)
+  • training_prep: 0.86s (0.2%)
+  • teacher_logprob_inference_prep: 0.64s (0.2%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 69: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:31:39 [abstract.py:324] It took 0.058140 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07350117e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07350117e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57247849e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07350117e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57247849e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07350117e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57247849e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57247849e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78622284e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78622284e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78622284e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78622284e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07361406e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56172142e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77561264e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19825077e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.74996948e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #69] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45522980e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07361406e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56172142e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77561264e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69642487e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31515713e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #69] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.36793518e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07361406e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56172142e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77561264e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34285736e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.40225506e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #69] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.10755234e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07361406e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56172142e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77561264e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30910187e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57534714e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #69] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32552826e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 69: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:31:42 [abstract.py:324] It took 0.030375 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12545.59it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:31:42 [loggers.py:236] Engine 000: Avg prompt throughput: 34.9 tokens/s, Avg generation throughput: 1.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:04,  2.47it/s, est. speed input: 148.42 toks/s, output: 29.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 14.67it/s, est. speed input: 715.89 toks/s, output: 196.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:08, 17.73it/s, est. speed input: 866.95 toks/s, output: 284.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 25.50it/s, est. speed input: 1153.13 toks/s, output: 506.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 34.60it/s, est. speed input: 1475.53 toks/s, output: 830.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 37.84it/s, est. speed input: 1626.76 toks/s, output: 1068.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:03, 38.42it/s, est. speed input: 1735.73 toks/s, output: 1321.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 32.34it/s, est. speed input: 1693.69 toks/s, output: 1460.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:03, 32.86it/s, est. speed input: 1723.37 toks/s, output: 1661.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:04, 27.29it/s, est. speed input: 1647.69 toks/s, output: 1765.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:02<00:12,  8.83it/s, est. speed input: 1028.28 toks/s, output: 1301.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:05<00:27,  3.95it/s, est. speed input: 626.67 toks/s, output: 1007.78 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:07<00:41,  2.54it/s, est. speed input: 462.97 toks/s, output: 901.79 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:08<00:42,  2.46it/s, est. speed input: 425.17 toks/s, output: 1005.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:09<00:49,  2.07it/s, est. speed input: 384.85 toks/s, output: 998.11 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 04:31:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11825.8 tokens/s, Running: 101 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:11<00:59,  1.67it/s, est. speed input: 319.88 toks/s, output: 1056.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:12<01:06,  1.49it/s, est. speed input: 296.42 toks/s, output: 1066.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:13<01:05,  1.51it/s, est. speed input: 287.07 toks/s, output: 1118.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:13<00:56,  1.71it/s, est. speed input: 286.03 toks/s, output: 1199.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:16<01:44,  1.09s/it, est. speed input: 239.50 toks/s, output: 1091.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:16<01:25,  1.12it/s, est. speed input: 239.02 toks/s, output: 1174.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:17<01:13,  1.27it/s, est. speed input: 236.32 toks/s, output: 1245.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:17<00:56,  1.66it/s, est. speed input: 238.52 toks/s, output: 1339.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:17<00:55,  1.67it/s, est. speed input: 234.11 toks/s, output: 1397.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:19<01:10,  1.29it/s, est. speed input: 222.15 toks/s, output: 1408.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9477.8 tokens/s, Running: 90 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:23<01:59,  1.34s/it, est. speed input: 188.24 toks/s, output: 1341.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:23<01:42,  1.17s/it, est. speed input: 185.86 toks/s, output: 1404.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:23<01:19,  1.09it/s, est. speed input: 186.95 toks/s, output: 1491.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:24<01:03,  1.35it/s, est. speed input: 187.43 toks/s, output: 1573.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:24<01:01,  1.39it/s, est. speed input: 184.91 toks/s, output: 1629.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:25<00:49,  1.71it/s, est. speed input: 185.58 toks/s, output: 1710.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:26<00:43,  1.88it/s, est. speed input: 183.63 toks/s, output: 1841.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:28<00:56,  1.42it/s, est. speed input: 175.13 toks/s, output: 1902.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:28<00:45,  1.73it/s, est. speed input: 176.60 toks/s, output: 1991.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:28<00:38,  2.03it/s, est. speed input: 177.44 toks/s, output: 2074.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:28<00:36,  2.09it/s, est. speed input: 176.83 toks/s, output: 2139.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:28<00:30,  2.53it/s, est. speed input: 177.86 toks/s, output: 2224.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:29<00:42,  1.76it/s, est. speed input: 173.76 toks/s, output: 2245.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7497.8 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:30<00:40,  1.81it/s, est. speed input: 172.88 toks/s, output: 2305.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:30<00:32,  2.24it/s, est. speed input: 173.85 toks/s, output: 2389.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:30<00:24,  2.89it/s, est. speed input: 175.19 toks/s, output: 2478.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:31<00:30,  2.36it/s, est. speed input: 173.68 toks/s, output: 2527.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:31<00:25,  2.72it/s, est. speed input: 174.36 toks/s, output: 2605.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:31<00:22,  3.02it/s, est. speed input: 174.88 toks/s, output: 2683.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:32<00:32,  2.11it/s, est. speed input: 172.40 toks/s, output: 2713.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:33<00:29,  2.25it/s, est. speed input: 172.29 toks/s, output: 2780.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:33<00:35,  1.85it/s, est. speed input: 170.24 toks/s, output: 2814.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:34<00:34,  1.87it/s, est. speed input: 169.34 toks/s, output: 2869.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:34<00:26,  2.37it/s, est. speed input: 170.03 toks/s, output: 3015.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:35<00:24,  2.50it/s, est. speed input: 170.13 toks/s, output: 3084.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:35<00:29,  2.08it/s, est. speed input: 168.44 toks/s, output: 3121.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:36<00:25,  2.34it/s, est. speed input: 168.86 toks/s, output: 3195.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:36<00:22,  2.68it/s, est. speed input: 169.50 toks/s, output: 3272.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:36<00:19,  3.04it/s, est. speed input: 170.11 toks/s, output: 3351.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:36<00:13,  4.20it/s, est. speed input: 172.17 toks/s, output: 3524.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:37<00:20,  2.63it/s, est. speed input: 170.13 toks/s, output: 3545.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:39<00:39,  1.36it/s, est. speed input: 164.03 toks/s, output: 3485.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:39<00:30,  1.75it/s, est. speed input: 165.11 toks/s, output: 3573.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:39<00:17,  2.89it/s, est. speed input: 167.76 toks/s, output: 3760.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:40<00:13,  3.73it/s, est. speed input: 169.70 toks/s, output: 3931.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6456.3 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:40<00:16,  2.86it/s, est. speed input: 168.46 toks/s, output: 3967.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:41<00:20,  2.34it/s, est. speed input: 167.13 toks/s, output: 4002.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:42<00:24,  1.87it/s, est. speed input: 165.17 toks/s, output: 4021.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:42<00:19,  2.27it/s, est. speed input: 165.92 toks/s, output: 4103.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:43<00:24,  1.83it/s, est. speed input: 164.18 toks/s, output: 4126.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:43<00:25,  1.71it/s, est. speed input: 163.07 toks/s, output: 4163.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:44<00:23,  1.81it/s, est. speed input: 162.68 toks/s, output: 4218.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:44<00:18,  2.26it/s, est. speed input: 163.47 toks/s, output: 4303.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:45<00:13,  2.93it/s, est. speed input: 164.56 toks/s, output: 4463.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:45<00:17,  2.20it/s, est. speed input: 163.01 toks/s, output: 4486.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:46<00:15,  2.46it/s, est. speed input: 163.39 toks/s, output: 4562.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:46<00:12,  2.87it/s, est. speed input: 164.04 toks/s, output: 4645.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:46<00:10,  3.44it/s, est. speed input: 164.87 toks/s, output: 4733.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:46<00:10,  3.29it/s, est. speed input: 165.01 toks/s, output: 4801.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:47<00:19,  1.73it/s, est. speed input: 161.99 toks/s, output: 4779.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:48<00:13,  2.28it/s, est. speed input: 162.66 toks/s, output: 4931.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:48<00:12,  2.40it/s, est. speed input: 162.77 toks/s, output: 4999.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:49<00:12,  2.37it/s, est. speed input: 162.58 toks/s, output: 5059.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:49<00:11,  2.48it/s, est. speed input: 162.67 toks/s, output: 5127.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:50<00:11,  2.43it/s, est. speed input: 162.48 toks/s, output: 5187.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4929.1 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:50<00:12,  2.07it/s, est. speed input: 161.51 toks/s, output: 5224.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:51<00:11,  2.13it/s, est. speed input: 161.32 toks/s, output: 5285.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:51<00:09,  2.59it/s, est. speed input: 161.97 toks/s, output: 5372.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:12,  1.73it/s, est. speed input: 159.26 toks/s, output: 5417.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:11,  1.86it/s, est. speed input: 159.17 toks/s, output: 5481.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:55<00:19,  1.02it/s, est. speed input: 153.83 toks/s, output: 5369.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:56<00:15,  1.20it/s, est. speed input: 153.79 toks/s, output: 5436.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:56<00:08,  1.96it/s, est. speed input: 155.45 toks/s, output: 5636.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:56<00:06,  2.38it/s, est. speed input: 156.25 toks/s, output: 5732.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:56<00:05,  2.75it/s, est. speed input: 156.82 toks/s, output: 5822.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:57<00:03,  3.28it/s, est. speed input: 157.73 toks/s, output: 5995.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:57<00:03,  3.12it/s, est. speed input: 157.76 toks/s, output: 6066.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:57<00:02,  3.79it/s, est. speed input: 158.87 toks/s, output: 6250.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:58<00:01,  4.86it/s, est. speed input: 160.37 toks/s, output: 6450.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:01,  3.78it/s, est. speed input: 160.06 toks/s, output: 6507.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:02,  2.31it/s, est. speed input: 158.39 toks/s, output: 6512.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2716.7 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:02,  1.87it/s, est. speed input: 157.16 toks/s, output: 6536.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:01<00:01,  1.56it/s, est. speed input: 155.07 toks/s, output: 6601.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:02<00:01,  1.44it/s, est. speed input: 153.85 toks/s, output: 6629.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:03<00:00,  1.48it/s, est. speed input: 153.32 toks/s, output: 6687.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.07it/s, est. speed input: 150.26 toks/s, output: 6639.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.07it/s, est. speed input: 150.26 toks/s, output: 6639.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  2.46it/s, est. speed input: 150.26 toks/s, output: 6639.25 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:32:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:32:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:32:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:32:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:32:51 [abstract.py:306] It took 0.545060 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:32:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:32:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:32:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:32:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.2938
+  • Mean Generation Length: 2701.9062
+  • Training FLOPS: 46.66 TFLOPS (11.67 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 319.27s
+  • policy_training: 206.05s (64.5%)
+  • generation: 69.52s (21.8%)
+  • teacher_logprob_inference: 38.73s (12.1%)
+  • policy_training/sharding_data: 4.39s (1.4%)
+  • prepare_for_generation: 3.46s (1.1%)
+  • policy_training/submit_training_futures: 2.05s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.5%)
+  • training_prep: 0.84s (0.3%)
+  • teacher_logprob_inference_prep: 0.49s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 70: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:36:59 [abstract.py:324] It took 0.050505 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07386992e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57683067e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79135971e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07386992e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57683067e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79135971e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07386992e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57683067e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79135971e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07386992e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57683067e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79135971e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07398008e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56150932e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77912750e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.21407089e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.86295223e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #70] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46240509e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07398008e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56150932e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77912750e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69661331e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29887629e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #70] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43078003e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07398008e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56150932e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77912750e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34982834e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.46548271e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #70] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.12295876e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07398008e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56150932e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77912750e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31149750e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.56156073e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #70] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32653412e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 70: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:02 [abstract.py:324] It took 0.030016 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12586.77it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:02 [loggers.py:236] Engine 000: Avg prompt throughput: 37.2 tokens/s, Avg generation throughput: 12.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.89it/s, est. speed input: 341.45 toks/s, output: 76.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:15, 10.03it/s, est. speed input: 549.68 toks/s, output: 190.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 16.42it/s, est. speed input: 833.39 toks/s, output: 367.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 24.97it/s, est. speed input: 1189.17 toks/s, output: 657.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:04, 30.08it/s, est. speed input: 1406.82 toks/s, output: 946.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 27.81it/s, est. speed input: 1417.95 toks/s, output: 1087.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 30.65it/s, est. speed input: 1523.40 toks/s, output: 1321.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:04, 27.28it/s, est. speed input: 1491.08 toks/s, output: 1463.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 36.10it/s, est. speed input: 1694.30 toks/s, output: 1949.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 49.06it/s, est. speed input: 1971.59 toks/s, output: 2612.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 51.89it/s, est. speed input: 2085.72 toks/s, output: 2982.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:03, 33.92it/s, est. speed input: 1915.61 toks/s, output: 2969.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:02<00:06, 16.38it/s, est. speed input: 1465.16 toks/s, output: 2521.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:03<00:09, 10.03it/s, est. speed input: 1144.49 toks/s, output: 2249.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:04<00:15,  5.97it/s, est. speed input: 865.66 toks/s, output: 1919.07 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:07<00:29,  3.10it/s, est. speed input: 601.52 toks/s, output: 1490.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:09<00:39,  2.20it/s, est. speed input: 478.57 toks/s, output: 1352.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:09<00:41,  2.12it/s, est. speed input: 454.66 toks/s, output: 1377.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10904.1 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:13<01:23,  1.03it/s, est. speed input: 327.97 toks/s, output: 1088.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:15<01:29,  1.05s/it, est. speed input: 300.75 toks/s, output: 1093.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:15<01:18,  1.08it/s, est. speed input: 296.13 toks/s, output: 1170.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:18<01:44,  1.26s/it, est. speed input: 258.94 toks/s, output: 1117.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:18<01:29,  1.09s/it, est. speed input: 254.52 toks/s, output: 1191.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8797.2 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:22<02:17,  1.70s/it, est. speed input: 216.88 toks/s, output: 1107.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:22<01:51,  1.40s/it, est. speed input: 214.25 toks/s, output: 1186.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:23<01:45,  1.34s/it, est. speed input: 205.99 toks/s, output: 1232.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:26<02:18,  1.78s/it, est. speed input: 185.88 toks/s, output: 1202.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:27<01:46,  1.38s/it, est. speed input: 185.27 toks/s, output: 1288.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:27<01:17,  1.02s/it, est. speed input: 186.45 toks/s, output: 1385.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:28<01:15,  1.01s/it, est. speed input: 182.32 toks/s, output: 1441.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:28<00:48,  1.50it/s, est. speed input: 183.23 toks/s, output: 1620.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7557.1 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:32<01:35,  1.33s/it, est. speed input: 165.87 toks/s, output: 1553.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:32<01:19,  1.11s/it, est. speed input: 165.11 toks/s, output: 1631.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:32<00:59,  1.18it/s, est. speed input: 166.19 toks/s, output: 1726.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:33<00:49,  1.40it/s, est. speed input: 166.11 toks/s, output: 1809.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:33<00:43,  1.55it/s, est. speed input: 165.57 toks/s, output: 1885.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:33<00:34,  1.93it/s, est. speed input: 166.35 toks/s, output: 1975.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:35<00:51,  1.29it/s, est. speed input: 161.48 toks/s, output: 1999.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:35<00:40,  1.61it/s, est. speed input: 162.06 toks/s, output: 2086.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<00:37,  1.70it/s, est. speed input: 161.47 toks/s, output: 2158.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<00:35,  1.76it/s, est. speed input: 160.80 toks/s, output: 2229.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:37<00:43,  1.44it/s, est. speed input: 158.20 toks/s, output: 2272.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:37<00:35,  1.73it/s, est. speed input: 158.53 toks/s, output: 2355.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:38<00:40,  1.47it/s, est. speed input: 156.33 toks/s, output: 2401.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6521.8 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:40<00:40,  1.42it/s, est. speed input: 153.68 toks/s, output: 2514.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:40<00:35,  1.59it/s, est. speed input: 153.57 toks/s, output: 2591.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:40<00:28,  1.99it/s, est. speed input: 154.52 toks/s, output: 2684.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:40<00:19,  2.79it/s, est. speed input: 156.18 toks/s, output: 2866.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:41<00:12,  4.09it/s, est. speed input: 158.58 toks/s, output: 3062.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:41<00:12,  4.04it/s, est. speed input: 158.99 toks/s, output: 3145.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:41<00:11,  4.20it/s, est. speed input: 160.26 toks/s, output: 3315.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:42<00:17,  2.75it/s, est. speed input: 158.55 toks/s, output: 3353.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:43<00:19,  2.37it/s, est. speed input: 157.55 toks/s, output: 3476.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:43<00:16,  2.72it/s, est. speed input: 158.29 toks/s, output: 3565.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:45<00:24,  1.78it/s, est. speed input: 155.42 toks/s, output: 3573.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:45<00:24,  1.78it/s, est. speed input: 154.75 toks/s, output: 3632.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:46<00:22,  1.83it/s, est. speed input: 153.98 toks/s, output: 3755.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:46<00:18,  2.18it/s, est. speed input: 154.69 toks/s, output: 3845.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:46<00:14,  2.68it/s, est. speed input: 155.64 toks/s, output: 3940.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:47<00:13,  2.76it/s, est. speed input: 155.77 toks/s, output: 4016.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:47<00:11,  3.34it/s, est. speed input: 156.63 toks/s, output: 4108.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:48<00:13,  2.59it/s, est. speed input: 155.84 toks/s, output: 4160.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:49<00:24,  1.46it/s, est. speed input: 152.57 toks/s, output: 4144.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:49<00:19,  1.70it/s, est. speed input: 152.75 toks/s, output: 4220.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:37:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5048.0 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:50<00:17,  1.91it/s, est. speed input: 152.87 toks/s, output: 4294.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:50<00:18,  1.72it/s, est. speed input: 151.89 toks/s, output: 4338.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:51<00:14,  2.12it/s, est. speed input: 152.46 toks/s, output: 4426.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:51<00:13,  2.28it/s, est. speed input: 152.59 toks/s, output: 4500.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:52<00:19,  1.49it/s, est. speed input: 150.17 toks/s, output: 4502.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:52<00:14,  1.94it/s, est. speed input: 150.96 toks/s, output: 4596.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:53<00:14,  1.88it/s, est. speed input: 150.48 toks/s, output: 4654.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:53<00:10,  2.38it/s, est. speed input: 151.18 toks/s, output: 4747.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:54<00:13,  1.89it/s, est. speed input: 150.08 toks/s, output: 4786.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:54<00:11,  2.12it/s, est. speed input: 150.26 toks/s, output: 4864.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:54<00:09,  2.33it/s, est. speed input: 150.43 toks/s, output: 4943.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:55<00:08,  2.67it/s, est. speed input: 150.86 toks/s, output: 5029.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:56<00:13,  1.57it/s, est. speed input: 148.58 toks/s, output: 5027.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:57<00:14,  1.36it/s, est. speed input: 147.17 toks/s, output: 5053.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:57<00:11,  1.64it/s, est. speed input: 147.41 toks/s, output: 5136.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:58<00:12,  1.41it/s, est. speed input: 146.09 toks/s, output: 5164.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:59<00:07,  2.02it/s, est. speed input: 146.92 toks/s, output: 5344.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:38:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3571.1 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:59<00:06,  2.20it/s, est. speed input: 146.99 toks/s, output: 5496.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:00<00:04,  2.49it/s, est. speed input: 147.50 toks/s, output: 5665.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:00<00:03,  2.90it/s, est. speed input: 148.21 toks/s, output: 5766.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:00<00:03,  3.25it/s, est. speed input: 148.76 toks/s, output: 5863.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:01<00:02,  3.01it/s, est. speed input: 148.71 toks/s, output: 5938.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:01<00:03,  2.55it/s, est. speed input: 148.30 toks/s, output: 5999.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:03<00:05,  1.33it/s, est. speed input: 145.22 toks/s, output: 5953.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:04<00:04,  1.48it/s, est. speed input: 145.11 toks/s, output: 6028.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:04<00:02,  1.87it/s, est. speed input: 145.62 toks/s, output: 6130.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:04<00:01,  2.32it/s, est. speed input: 146.15 toks/s, output: 6232.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:04<00:01,  2.65it/s, est. speed input: 146.49 toks/s, output: 6327.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:06<00:01,  1.24it/s, est. speed input: 143.33 toks/s, output: 6276.19 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  1.24it/s, est. speed input: 145.15 toks/s, output: 6522.46 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  2.41it/s, est. speed input: 145.15 toks/s, output: 6522.46 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:38:12 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:38:12 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:38:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:38:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:38:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:38:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:38:13 [abstract.py:306] It took 0.529916 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:38:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:38:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:38:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:38:13 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0111
+  • Training Accuracy: 0.3187
+  • Mean Generation Length: 2711.9126
+  • Training FLOPS: 46.99 TFLOPS (11.75 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 322.90s
+  • policy_training: 207.83s (64.4%)
+  • generation: 71.15s (22.0%)
+  • teacher_logprob_inference: 38.91s (12.1%)
+  • policy_training/sharding_data: 4.29s (1.3%)
+  • prepare_for_generation: 3.53s (1.1%)
+  • policy_training/submit_training_futures: 2.72s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.5%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • get_topk_logits/submit_topk_logits_futures: 0.12s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 15/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 71: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:42:22 [abstract.py:324] It took 0.054979 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07397109e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58275223e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79951782e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07397109e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58275223e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79951782e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07397109e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58275223e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79951782e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07397109e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58275223e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79951782e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07404844e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56935425e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78719177e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.21877537e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.85943413e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #71] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46058029e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07404844e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56935425e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78719177e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68467712e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29645271e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #71] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43888092e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07404844e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56935425e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78719177e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34842377e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.56609535e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #71] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15913963e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07404844e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56935425e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78719177e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30068626e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.55705585e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #71] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32949860e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 71: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:42:24 [abstract.py:324] It took 0.031126 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13815.23it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:42:25 [loggers.py:236] Engine 000: Avg prompt throughput: 37.2 tokens/s, Avg generation throughput: 37.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:05,  2.43it/s, est. speed input: 148.38 toks/s, output: 29.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 14.49it/s, est. speed input: 700.59 toks/s, output: 183.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 23.15it/s, est. speed input: 1059.27 toks/s, output: 358.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 28.65it/s, est. speed input: 1300.54 toks/s, output: 569.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 38.30it/s, est. speed input: 1666.01 toks/s, output: 936.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 36.14it/s, est. speed input: 1717.05 toks/s, output: 1129.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 35.84it/s, est. speed input: 1775.86 toks/s, output: 1346.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 36.64it/s, est. speed input: 1821.89 toks/s, output: 1547.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 39.24it/s, est. speed input: 1906.82 toks/s, output: 1864.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:03, 32.14it/s, est. speed input: 1837.18 toks/s, output: 2001.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:02<00:05, 20.44it/s, est. speed input: 1597.25 toks/s, output: 1908.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:02<00:10,  9.76it/s, est. speed input: 1173.80 toks/s, output: 1605.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:05<00:30,  3.30it/s, est. speed input: 624.84 toks/s, output: 1074.68 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:08<00:45,  2.19it/s, est. speed input: 463.91 toks/s, output: 980.01 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:09<00:51,  1.89it/s, est. speed input: 398.71 toks/s, output: 1015.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:42:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11454.8 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:10<00:53,  1.78it/s, est. speed input: 374.85 toks/s, output: 1045.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:12<01:10,  1.35it/s, est. speed input: 328.02 toks/s, output: 1004.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:15<01:34,  1.01s/it, est. speed input: 266.63 toks/s, output: 979.31 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:16<01:28,  1.04it/s, est. speed input: 258.00 toks/s, output: 1036.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:17<01:26,  1.05it/s, est. speed input: 248.19 toks/s, output: 1085.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:17<01:12,  1.24it/s, est. speed input: 247.27 toks/s, output: 1168.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:17<01:03,  1.39it/s, est. speed input: 244.66 toks/s, output: 1242.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:19<01:02,  1.39it/s, est. speed input: 232.96 toks/s, output: 1348.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:19<01:01,  1.41it/s, est. speed input: 228.10 toks/s, output: 1404.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:42:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9172.6 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:20<00:54,  1.55it/s, est. speed input: 225.95 toks/s, output: 1475.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:21<01:06,  1.27it/s, est. speed input: 216.37 toks/s, output: 1496.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:21<00:42,  1.92it/s, est. speed input: 218.87 toks/s, output: 1676.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:22<00:35,  2.30it/s, est. speed input: 220.14 toks/s, output: 1766.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:22<00:30,  2.58it/s, est. speed input: 220.53 toks/s, output: 1848.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:22<00:32,  2.44it/s, est. speed input: 218.78 toks/s, output: 1911.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:23<00:28,  2.74it/s, est. speed input: 219.09 toks/s, output: 1992.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:23<00:35,  2.15it/s, est. speed input: 214.94 toks/s, output: 2033.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:23<00:28,  2.71it/s, est. speed input: 216.41 toks/s, output: 2124.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:25<00:51,  1.46it/s, est. speed input: 206.33 toks/s, output: 2103.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:25<00:39,  1.89it/s, est. speed input: 207.40 toks/s, output: 2192.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:25<00:35,  2.06it/s, est. speed input: 206.77 toks/s, output: 2261.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:26<00:20,  3.45it/s, est. speed input: 210.47 toks/s, output: 2453.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:26<00:20,  3.32it/s, est. speed input: 209.93 toks/s, output: 2595.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:26<00:15,  4.36it/s, est. speed input: 212.84 toks/s, output: 2778.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:27<00:16,  3.89it/s, est. speed input: 212.36 toks/s, output: 2842.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:27<00:15,  4.32it/s, est. speed input: 213.39 toks/s, output: 2928.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:28<00:20,  3.01it/s, est. speed input: 209.85 toks/s, output: 3090.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:29<00:20,  3.01it/s, est. speed input: 209.65 toks/s, output: 3156.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:29<00:13,  4.22it/s, est. speed input: 212.83 toks/s, output: 3346.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:29<00:13,  4.40it/s, est. speed input: 213.57 toks/s, output: 3426.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:29<00:15,  3.76it/s, est. speed input: 212.72 toks/s, output: 3482.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:29<00:13,  4.01it/s, est. speed input: 213.26 toks/s, output: 3561.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:42:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7355.2 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:30<00:16,  3.33it/s, est. speed input: 212.21 toks/s, output: 3611.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:30<00:17,  3.11it/s, est. speed input: 211.58 toks/s, output: 3669.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:30<00:15,  3.53it/s, est. speed input: 212.36 toks/s, output: 3749.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:31<00:15,  3.45it/s, est. speed input: 212.26 toks/s, output: 3815.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:20,  2.54it/s, est. speed input: 209.83 toks/s, output: 3841.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:32<00:29,  1.67it/s, est. speed input: 204.78 toks/s, output: 3817.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:18,  2.60it/s, est. speed input: 206.92 toks/s, output: 3995.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:19,  2.37it/s, est. speed input: 205.35 toks/s, output: 4036.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:34<00:17,  2.61it/s, est. speed input: 205.53 toks/s, output: 4108.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:34<00:17,  2.56it/s, est. speed input: 204.76 toks/s, output: 4163.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:34<00:15,  2.79it/s, est. speed input: 204.80 toks/s, output: 4235.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:35<00:14,  2.93it/s, est. speed input: 204.77 toks/s, output: 4304.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:35<00:15,  2.63it/s, est. speed input: 203.76 toks/s, output: 4352.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:35<00:13,  3.03it/s, est. speed input: 204.36 toks/s, output: 4432.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:35<00:11,  3.46it/s, est. speed input: 204.99 toks/s, output: 4514.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:21,  1.77it/s, est. speed input: 199.84 toks/s, output: 4473.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:12,  2.94it/s, est. speed input: 202.30 toks/s, output: 4667.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:37<00:12,  2.84it/s, est. speed input: 201.77 toks/s, output: 4725.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:38<00:12,  2.87it/s, est. speed input: 201.58 toks/s, output: 4789.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:38<00:11,  3.02it/s, est. speed input: 201.70 toks/s, output: 4860.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:38<00:12,  2.64it/s, est. speed input: 200.67 toks/s, output: 4904.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:39<00:11,  2.89it/s, est. speed input: 200.85 toks/s, output: 4977.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:39<00:09,  3.23it/s, est. speed input: 201.28 toks/s, output: 5057.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:39<00:06,  4.30it/s, est. speed input: 202.91 toks/s, output: 5235.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:39<00:06,  4.52it/s, est. speed input: 203.53 toks/s, output: 5319.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:43:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5358.9 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:40<00:08,  3.07it/s, est. speed input: 201.88 toks/s, output: 5345.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:40<00:07,  3.26it/s, est. speed input: 202.11 toks/s, output: 5420.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:40<00:08,  3.09it/s, est. speed input: 201.77 toks/s, output: 5480.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:41<00:07,  3.03it/s, est. speed input: 201.54 toks/s, output: 5544.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:41<00:08,  2.69it/s, est. speed input: 200.71 toks/s, output: 5591.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:43<00:17,  1.25it/s, est. speed input: 193.59 toks/s, output: 5467.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:44<00:10,  1.90it/s, est. speed input: 194.63 toks/s, output: 5642.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:44<00:09,  1.93it/s, est. speed input: 193.82 toks/s, output: 5692.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:45<00:07,  2.30it/s, est. speed input: 193.75 toks/s, output: 5838.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:45<00:03,  3.82it/s, est. speed input: 196.97 toks/s, output: 6154.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:46<00:04,  2.64it/s, est. speed input: 194.46 toks/s, output: 6149.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:48<00:09,  1.28it/s, est. speed input: 186.38 toks/s, output: 5975.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:49<00:09,  1.13it/s, est. speed input: 183.02 toks/s, output: 5949.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:43:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3017.1 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:51<00:10,  1.08s/it, est. speed input: 178.34 toks/s, output: 5879.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:51<00:08,  1.09it/s, est. speed input: 177.99 toks/s, output: 5953.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:52<00:05,  1.36it/s, est. speed input: 178.29 toks/s, output: 6049.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:52<00:04,  1.61it/s, est. speed input: 178.29 toks/s, output: 6136.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:52<00:03,  1.92it/s, est. speed input: 178.49 toks/s, output: 6231.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:53<00:02,  2.11it/s, est. speed input: 178.45 toks/s, output: 6316.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:56<00:02,  1.08it/s, est. speed input: 171.25 toks/s, output: 6237.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:56<00:01,  1.29it/s, est. speed input: 171.32 toks/s, output: 6336.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:57<00:00,  1.34it/s, est. speed input: 170.39 toks/s, output: 6399.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.02s/it, est. speed input: 166.26 toks/s, output: 6344.29 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.02s/it, est. speed input: 166.26 toks/s, output: 6344.29 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.72it/s, est. speed input: 166.26 toks/s, output: 6344.29 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:43:27 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:43:27 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:43:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:43:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:43:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:43:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:43:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 04:43:27 [abstract.py:306] It took 0.526895 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:43:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:43:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:43:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0104
+  • Training Accuracy: 0.2625
+  • Mean Generation Length: 2333.3687
+  • Training FLOPS: 45.14 TFLOPS (11.28 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 285.28s
+  • policy_training: 182.78s (64.1%)
+  • generation: 62.76s (22.0%)
+  • teacher_logprob_inference: 34.64s (12.1%)
+  • policy_training/sharding_data: 4.44s (1.6%)
+  • prepare_for_generation: 3.56s (1.2%)
+  • policy_training/submit_training_futures: 2.26s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.63s (0.6%)
+  • training_prep: 0.91s (0.3%)
+  • teacher_logprob_inference_prep: 0.43s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 72: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:07 [abstract.py:324] It took 0.048341 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07438203e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59101944e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80587769e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07438203e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59101944e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80587769e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07438203e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59101944e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80587769e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07438203e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59101944e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80587769e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07453301e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58369446e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79228668e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20436916e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.98274040e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #72] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45957336e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07453301e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58369446e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79228668e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70023041e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29400616e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #72] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42250290e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07453301e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58369446e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79228668e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34784851e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.49806499e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #72] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15721741e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07453301e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58369446e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79228668e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29082184e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54499454e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #72] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32908112e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 72: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:10 [abstract.py:324] It took 0.032626 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:10 [loggers.py:236] Engine 000: Avg prompt throughput: 41.1 tokens/s, Avg generation throughput: 40.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12544.42it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.01it/s, est. speed input: 354.51 toks/s, output: 78.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:02, 50.70it/s, est. speed input: 2586.53 toks/s, output: 708.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:03, 36.28it/s, est. speed input: 2128.27 toks/s, output: 811.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 33.03it/s, est. speed input: 2024.60 toks/s, output: 1019.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:05, 26.59it/s, est. speed input: 1783.48 toks/s, output: 1127.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:04, 28.41it/s, est. speed input: 1822.58 toks/s, output: 1371.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:04, 29.77it/s, est. speed input: 1847.81 toks/s, output: 1595.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:03, 32.06it/s, est. speed input: 1895.20 toks/s, output: 1934.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:03, 32.56it/s, est. speed input: 1907.93 toks/s, output: 2164.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 36.97it/s, est. speed input: 1991.96 toks/s, output: 2575.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:03, 32.11it/s, est. speed input: 1934.48 toks/s, output: 2717.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:02<00:04, 24.31it/s, est. speed input: 1789.59 toks/s, output: 2750.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:02<00:06, 15.57it/s, est. speed input: 1546.35 toks/s, output: 2561.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:03<00:14,  6.46it/s, est. speed input: 1052.21 toks/s, output: 1916.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:04<00:19,  4.69it/s, est. speed input: 868.69 toks/s, output: 1755.74 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:05<00:24,  3.70it/s, est. speed input: 740.91 toks/s, output: 1662.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:06<00:30,  2.91it/s, est. speed input: 655.93 toks/s, output: 1562.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:07<00:33,  2.67it/s, est. speed input: 613.87 toks/s, output: 1553.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:07<00:34,  2.54it/s, est. speed input: 583.28 toks/s, output: 1565.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:08<00:47,  1.82it/s, est. speed input: 512.81 toks/s, output: 1469.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:09<00:36,  2.32it/s, est. speed input: 499.92 toks/s, output: 1612.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10967.0 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:10<00:53,  1.58it/s, est. speed input: 439.34 toks/s, output: 1510.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:11<00:57,  1.44it/s, est. speed input: 409.83 toks/s, output: 1502.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:11<00:46,  1.77it/s, est. speed input: 409.15 toks/s, output: 1592.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:12<00:54,  1.49it/s, est. speed input: 382.14 toks/s, output: 1579.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:13<00:52,  1.54it/s, est. speed input: 369.27 toks/s, output: 1618.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:13<00:44,  1.76it/s, est. speed input: 364.04 toks/s, output: 1686.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:13<00:40,  1.95it/s, est. speed input: 358.47 toks/s, output: 1751.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<00:34,  2.19it/s, est. speed input: 347.70 toks/s, output: 1875.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:16<00:58,  1.28it/s, est. speed input: 313.65 toks/s, output: 1781.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:17<00:43,  1.69it/s, est. speed input: 308.97 toks/s, output: 1930.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:18<00:50,  1.43it/s, est. speed input: 294.06 toks/s, output: 1926.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:18<00:47,  1.51it/s, est. speed input: 288.64 toks/s, output: 1981.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8362.0 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<01:19,  1.13s/it, est. speed input: 257.67 toks/s, output: 1858.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<01:00,  1.15it/s, est. speed input: 258.65 toks/s, output: 1954.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:23<01:30,  1.33s/it, est. speed input: 233.63 toks/s, output: 1856.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:24<01:21,  1.22s/it, est. speed input: 227.18 toks/s, output: 1894.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:25<01:13,  1.11s/it, est. speed input: 222.15 toks/s, output: 1942.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:26<01:07,  1.04s/it, est. speed input: 217.06 toks/s, output: 1986.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:28<01:17,  1.21s/it, est. speed input: 206.67 toks/s, output: 1979.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<01:00,  1.05it/s, est. speed input: 206.37 toks/s, output: 2064.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:29<00:58,  1.07it/s, est. speed input: 202.16 toks/s, output: 2109.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:29<00:48,  1.26it/s, est. speed input: 201.09 toks/s, output: 2185.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7053.8 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:32<00:54,  1.08it/s, est. speed input: 191.27 toks/s, output: 2247.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:33<00:55,  1.05it/s, est. speed input: 187.22 toks/s, output: 2287.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:33<00:43,  1.31it/s, est. speed input: 187.75 toks/s, output: 2379.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:33<00:35,  1.59it/s, est. speed input: 188.07 toks/s, output: 2469.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:34<00:41,  1.34it/s, est. speed input: 184.13 toks/s, output: 2503.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:35<00:45,  1.19it/s, est. speed input: 180.33 toks/s, output: 2537.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:35<00:36,  1.44it/s, est. speed input: 180.32 toks/s, output: 2623.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:36<00:23,  2.18it/s, est. speed input: 181.93 toks/s, output: 2815.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:36<00:24,  2.01it/s, est. speed input: 180.48 toks/s, output: 2877.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:37<00:26,  1.84it/s, est. speed input: 178.83 toks/s, output: 2934.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:37<00:21,  2.23it/s, est. speed input: 179.49 toks/s, output: 3028.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:27,  1.73it/s, est. speed input: 176.76 toks/s, output: 3065.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:47:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6070.5 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:40<00:48,  1.05s/it, est. speed input: 168.50 toks/s, output: 3007.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:41<00:35,  1.26it/s, est. speed input: 169.42 toks/s, output: 3106.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:41<00:26,  1.69it/s, est. speed input: 170.43 toks/s, output: 3207.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:42<00:39,  1.10it/s, est. speed input: 165.12 toks/s, output: 3192.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:43<00:21,  1.93it/s, est. speed input: 167.60 toks/s, output: 3405.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:43<00:17,  2.29it/s, est. speed input: 168.30 toks/s, output: 3500.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:43<00:19,  2.00it/s, est. speed input: 167.08 toks/s, output: 3557.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:44<00:10,  3.50it/s, est. speed input: 170.18 toks/s, output: 3864.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:09,  3.58it/s, est. speed input: 170.63 toks/s, output: 3952.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:16,  2.01it/s, est. speed input: 167.24 toks/s, output: 3955.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:46<00:15,  2.09it/s, est. speed input: 166.58 toks/s, output: 4099.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:12,  2.44it/s, est. speed input: 167.29 toks/s, output: 4196.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:47<00:17,  1.67it/s, est. speed input: 164.32 toks/s, output: 4203.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:49<00:21,  1.38it/s, est. speed input: 161.88 toks/s, output: 4222.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:49<00:20,  1.34it/s, est. speed input: 160.52 toks/s, output: 4267.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:48:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4680.2 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:50<00:12,  2.02it/s, est. speed input: 161.84 toks/s, output: 4466.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:51<00:17,  1.42it/s, est. speed input: 158.66 toks/s, output: 4460.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:52<00:19,  1.21it/s, est. speed input: 156.15 toks/s, output: 4472.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:52<00:15,  1.52it/s, est. speed input: 156.75 toks/s, output: 4571.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:12,  1.74it/s, est. speed input: 156.83 toks/s, output: 4656.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:54<00:13,  1.53it/s, est. speed input: 155.47 toks/s, output: 4697.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:54<00:11,  1.73it/s, est. speed input: 155.45 toks/s, output: 4779.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:55<00:11,  1.69it/s, est. speed input: 154.79 toks/s, output: 4841.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:55<00:11,  1.60it/s, est. speed input: 153.93 toks/s, output: 4896.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:56<00:09,  1.83it/s, est. speed input: 154.05 toks/s, output: 4981.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:57<00:09,  1.61it/s, est. speed input: 152.95 toks/s, output: 5029.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:58<00:11,  1.33it/s, est. speed input: 151.18 toks/s, output: 5055.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:58<00:10,  1.35it/s, est. speed input: 150.41 toks/s, output: 5114.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:58<00:07,  1.73it/s, est. speed input: 150.92 toks/s, output: 5216.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:59<00:06,  1.95it/s, est. speed input: 151.02 toks/s, output: 5304.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:59<00:04,  2.32it/s, est. speed input: 151.43 toks/s, output: 5402.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:48:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2997.6 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:00<00:05,  1.99it/s, est. speed input: 150.74 toks/s, output: 5464.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:00<00:03,  2.39it/s, est. speed input: 151.18 toks/s, output: 5564.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:01<00:04,  1.88it/s, est. speed input: 150.17 toks/s, output: 5614.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:01<00:03,  1.82it/s, est. speed input: 149.72 toks/s, output: 5684.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:03<00:04,  1.33it/s, est. speed input: 147.75 toks/s, output: 5699.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:03<00:02,  1.77it/s, est. speed input: 148.40 toks/s, output: 5812.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:04<00:02,  1.52it/s, est. speed input: 147.30 toks/s, output: 5860.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:04<00:01,  1.81it/s, est. speed input: 147.50 toks/s, output: 5959.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.81it/s, est. speed input: 150.30 toks/s, output: 6340.63 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.48it/s, est. speed input: 150.30 toks/s, output: 6340.63 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:48:19 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:48:19 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:48:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:48:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:48:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:48:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:48:19 [abstract.py:306] It took 0.524632 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:48:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:48:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:48:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:48:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0111
+  • Training Accuracy: 0.3312
+  • Mean Generation Length: 2552.3125
+  • Training FLOPS: 47.21 TFLOPS (11.80 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 304.72s
+  • policy_training: 195.30s (64.1%)
+  • generation: 68.90s (22.6%)
+  • teacher_logprob_inference: 35.45s (11.6%)
+  • policy_training/sharding_data: 4.21s (1.4%)
+  • prepare_for_generation: 3.63s (1.2%)
+  • policy_training/submit_training_futures: 1.86s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.5%)
+  • training_prep: 0.78s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 73: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:12 [abstract.py:324] It took 0.050252 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07480996e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59917450e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80988388e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07480996e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59917450e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80988388e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07480996e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59917450e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80988388e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07480996e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59917450e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80988388e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07491641e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59258194e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79764633e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20405922e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.92675257e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #73] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46093781e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07491641e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59258194e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79764633e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67628174e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31013031e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #73] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42165375e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07491641e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59258194e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79764633e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.33489838e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.68146038e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #73] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14993210e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07491641e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59258194e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79764633e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28037338e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53477802e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #73] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33262772e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 73: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:15 [abstract.py:324] It took 0.031292 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14035.40it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:15 [loggers.py:236] Engine 000: Avg prompt throughput: 39.4 tokens/s, Avg generation throughput: 28.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:42,  1.56it/s, est. speed input: 91.90 toks/s, output: 17.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:23,  6.67it/s, est. speed input: 329.16 toks/s, output: 89.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:09, 15.44it/s, est. speed input: 643.14 toks/s, output: 235.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:01<00:08, 17.87it/s, est. speed input: 772.73 toks/s, output: 351.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:06, 20.34it/s, est. speed input: 890.14 toks/s, output: 522.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:05, 24.28it/s, est. speed input: 1006.33 toks/s, output: 700.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:05, 26.12it/s, est. speed input: 1084.28 toks/s, output: 869.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 36.45it/s, est. speed input: 1302.53 toks/s, output: 1287.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 37.31it/s, est. speed input: 1384.39 toks/s, output: 1518.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:03, 29.45it/s, est. speed input: 1359.08 toks/s, output: 1665.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:02<00:03, 29.51it/s, est. speed input: 1384.77 toks/s, output: 1859.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:02<00:03, 31.37it/s, est. speed input: 1427.36 toks/s, output: 2074.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11418.7 tokens/s, Running: 105 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:10<01:03,  1.64it/s, est. speed input: 316.04 toks/s, output: 665.03 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:13<01:08,  1.46it/s, est. speed input: 264.01 toks/s, output: 803.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:14<01:03,  1.56it/s, est. speed input: 255.58 toks/s, output: 945.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:14<00:50,  1.91it/s, est. speed input: 262.03 toks/s, output: 1133.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:16<01:03,  1.50it/s, est. speed input: 233.45 toks/s, output: 1170.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:17<00:53,  1.73it/s, est. speed input: 231.86 toks/s, output: 1322.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:18<00:54,  1.68it/s, est. speed input: 222.63 toks/s, output: 1421.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:19<00:53,  1.69it/s, est. speed input: 219.23 toks/s, output: 1477.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:19<00:52,  1.69it/s, est. speed input: 215.73 toks/s, output: 1531.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:20<00:51,  1.71it/s, est. speed input: 212.77 toks/s, output: 1586.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9440.4 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:21<00:40,  2.12it/s, est. speed input: 212.82 toks/s, output: 1738.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:21<00:36,  2.35it/s, est. speed input: 213.05 toks/s, output: 1815.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:21<00:31,  2.67it/s, est. speed input: 213.76 toks/s, output: 1895.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:22<00:41,  2.01it/s, est. speed input: 208.24 toks/s, output: 1919.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:22<00:32,  2.49it/s, est. speed input: 208.60 toks/s, output: 2068.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:23<00:33,  2.41it/s, est. speed input: 207.03 toks/s, output: 2124.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:24<00:55,  1.41it/s, est. speed input: 196.22 toks/s, output: 2085.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:25<00:50,  1.54it/s, est. speed input: 194.95 toks/s, output: 2143.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:28<01:45,  1.37s/it, est. speed input: 174.47 toks/s, output: 1992.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:29<01:25,  1.13s/it, est. speed input: 173.62 toks/s, output: 2056.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:29<01:10,  1.07it/s, est. speed input: 173.09 toks/s, output: 2122.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:30<00:58,  1.26it/s, est. speed input: 172.52 toks/s, output: 2187.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:30<00:33,  2.17it/s, est. speed input: 175.78 toks/s, output: 2374.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:30<00:26,  2.64it/s, est. speed input: 177.08 toks/s, output: 2462.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7658.9 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:30<00:29,  2.35it/s, est. speed input: 175.73 toks/s, output: 2514.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:31<00:24,  2.86it/s, est. speed input: 176.86 toks/s, output: 2600.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:31<00:25,  2.63it/s, est. speed input: 176.15 toks/s, output: 2660.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:31<00:26,  2.55it/s, est. speed input: 175.70 toks/s, output: 2722.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<00:27,  2.42it/s, est. speed input: 175.05 toks/s, output: 2781.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:32<00:15,  4.06it/s, est. speed input: 178.74 toks/s, output: 3043.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:33<00:18,  3.35it/s, est. speed input: 177.88 toks/s, output: 3095.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:33<00:14,  4.19it/s, est. speed input: 179.96 toks/s, output: 3265.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:35<00:29,  2.01it/s, est. speed input: 174.22 toks/s, output: 3227.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:35<00:30,  1.92it/s, est. speed input: 173.00 toks/s, output: 3272.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:37<00:50,  1.13it/s, est. speed input: 165.40 toks/s, output: 3196.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:37<00:32,  1.71it/s, est. speed input: 167.18 toks/s, output: 3366.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:38<00:26,  2.04it/s, est. speed input: 167.41 toks/s, output: 3505.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:14,  3.48it/s, est. speed input: 171.88 toks/s, output: 3864.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:39<00:12,  3.72it/s, est. speed input: 172.67 toks/s, output: 3947.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:39<00:15,  3.12it/s, est. speed input: 171.85 toks/s, output: 3992.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:13,  3.31it/s, est. speed input: 172.36 toks/s, output: 4069.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:09,  4.77it/s, est. speed input: 174.88 toks/s, output: 4257.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:52:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6302.8 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:14,  2.98it/s, est. speed input: 172.86 toks/s, output: 4273.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:41<00:14,  2.84it/s, est. speed input: 172.63 toks/s, output: 4332.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:42<00:21,  1.87it/s, est. speed input: 169.55 toks/s, output: 4321.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:42<00:17,  2.27it/s, est. speed input: 170.30 toks/s, output: 4405.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:43<00:17,  2.17it/s, est. speed input: 169.66 toks/s, output: 4454.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:43<00:14,  2.64it/s, est. speed input: 170.40 toks/s, output: 4539.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:12,  2.89it/s, est. speed input: 170.72 toks/s, output: 4613.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:10,  3.33it/s, est. speed input: 171.56 toks/s, output: 4765.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:09,  3.46it/s, est. speed input: 171.97 toks/s, output: 4840.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:44<00:12,  2.59it/s, est. speed input: 170.75 toks/s, output: 4870.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:11,  2.72it/s, est. speed input: 170.96 toks/s, output: 4940.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:45<00:11,  2.73it/s, est. speed input: 170.91 toks/s, output: 5004.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:09,  3.28it/s, est. speed input: 171.72 toks/s, output: 5091.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:10,  2.70it/s, est. speed input: 171.04 toks/s, output: 5136.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:11,  2.45it/s, est. speed input: 170.50 toks/s, output: 5186.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:10,  2.58it/s, est. speed input: 170.51 toks/s, output: 5253.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:11,  2.32it/s, est. speed input: 169.91 toks/s, output: 5300.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:47<00:05,  4.25it/s, est. speed input: 172.84 toks/s, output: 5588.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:49<00:09,  2.22it/s, est. speed input: 169.59 toks/s, output: 5551.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:11,  1.83it/s, est. speed input: 167.84 toks/s, output: 5563.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:53:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4359.2 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:51<00:15,  1.27it/s, est. speed input: 164.04 toks/s, output: 5508.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:51<00:07,  2.17it/s, est. speed input: 166.06 toks/s, output: 5787.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:52<00:05,  2.64it/s, est. speed input: 166.99 toks/s, output: 5959.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:52<00:05,  2.56it/s, est. speed input: 166.72 toks/s, output: 6020.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:53<00:05,  2.50it/s, est. speed input: 166.54 toks/s, output: 6083.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:53<00:05,  2.33it/s, est. speed input: 166.01 toks/s, output: 6136.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:54<00:06,  1.78it/s, est. speed input: 164.20 toks/s, output: 6142.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:56<00:08,  1.25it/s, est. speed input: 161.00 toks/s, output: 6095.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:57<00:08,  1.03it/s, est. speed input: 158.10 toks/s, output: 6063.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:58<00:07,  1.12it/s, est. speed input: 157.24 toks/s, output: 6107.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:59<00:05,  1.20it/s, est. speed input: 156.45 toks/s, output: 6156.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:04,  1.39it/s, est. speed input: 156.26 toks/s, output: 6229.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:53:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2134.6 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:01<00:05,  1.01s/it, est. speed input: 152.90 toks/s, output: 6179.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:03,  1.16it/s, est. speed input: 152.57 toks/s, output: 6250.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:03<00:03,  1.28s/it, est. speed input: 148.12 toks/s, output: 6158.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.28s/it, est. speed input: 150.91 toks/s, output: 6542.88 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.50it/s, est. speed input: 150.91 toks/s, output: 6542.88 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:53:23 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:53:23 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:53:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:53:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:53:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:53:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:53:23 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:53:23 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:53:23 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 04:53:23 [abstract.py:306] It took 0.744078 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:53:23 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.2812
+  • Mean Generation Length: 2614.3000
+  • Training FLOPS: 46.71 TFLOPS (11.68 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 310.14s
+  • policy_training: 198.97s (64.2%)
+  • generation: 68.46s (22.1%)
+  • teacher_logprob_inference: 37.71s (12.2%)
+  • policy_training/sharding_data: 4.32s (1.4%)
+  • prepare_for_generation: 3.52s (1.1%)
+  • policy_training/submit_training_futures: 1.84s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.5%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 74: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 04:57:22 [abstract.py:324] It took 0.048591 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07526152e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60253296e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81217957e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07526152e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60253296e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81217957e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07526152e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60253296e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81217957e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07526152e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60253296e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81217957e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07526914e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58594398e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78152466e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19822559e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.01704836e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #74] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46286255e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07526914e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58594398e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78152466e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68971558e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31019764e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #74] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42659607e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07526914e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58594398e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78152466e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34326477e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.97972536e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #74] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.15088120e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07526914e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58594398e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78152466e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27641525e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53600597e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #74] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33036133e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 74: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 04:57:26 [abstract.py:324] It took 0.030280 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13353.67it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:57:26 [loggers.py:236] Engine 000: Avg prompt throughput: 38.9 tokens/s, Avg generation throughput: 14.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:39,  4.02it/s, est. speed input: 236.99 toks/s, output: 48.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 21.49it/s, est. speed input: 1099.62 toks/s, output: 323.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 27.50it/s, est. speed input: 1404.30 toks/s, output: 534.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 32.97it/s, est. speed input: 1641.00 toks/s, output: 785.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 30.64it/s, est. speed input: 1637.80 toks/s, output: 925.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:04, 31.49it/s, est. speed input: 1691.57 toks/s, output: 1122.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 33.95it/s, est. speed input: 1783.08 toks/s, output: 1391.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:04, 28.78it/s, est. speed input: 1704.26 toks/s, output: 1527.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 32.38it/s, est. speed input: 1781.38 toks/s, output: 1820.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:03, 37.29it/s, est. speed input: 1887.32 toks/s, output: 2202.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 47.24it/s, est. speed input: 2069.58 toks/s, output: 2772.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:02, 45.24it/s, est. speed input: 2109.63 toks/s, output: 3089.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 42.15it/s, est. speed input: 2113.40 toks/s, output: 3315.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:02, 40.15it/s, est. speed input: 2119.50 toks/s, output: 3554.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:03, 23.88it/s, est. speed input: 1872.38 toks/s, output: 3402.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:04<00:11,  7.39it/s, est. speed input: 1157.81 toks/s, output: 2349.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:06<00:23,  3.29it/s, est. speed input: 710.36 toks/s, output: 1720.54 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:08<00:27,  2.80it/s, est. speed input: 616.20 toks/s, output: 1671.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:08<00:26,  2.83it/s, est. speed input: 583.05 toks/s, output: 1771.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:57:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10368.7 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:10<00:33,  2.18it/s, est. speed input: 518.02 toks/s, output: 1671.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:10<00:30,  2.39it/s, est. speed input: 514.94 toks/s, output: 1758.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:27,  2.65it/s, est. speed input: 511.67 toks/s, output: 1843.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:11<00:29,  2.38it/s, est. speed input: 489.52 toks/s, output: 1860.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:11<00:32,  2.13it/s, est. speed input: 467.92 toks/s, output: 1874.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:13<00:56,  1.21it/s, est. speed input: 405.65 toks/s, output: 1722.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<01:13,  1.08s/it, est. speed input: 361.39 toks/s, output: 1633.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:16<01:11,  1.06s/it, est. speed input: 343.31 toks/s, output: 1650.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:16<00:53,  1.23it/s, est. speed input: 343.63 toks/s, output: 1750.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:49,  1.31it/s, est. speed input: 334.48 toks/s, output: 1801.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:17<00:30,  2.06it/s, est. speed input: 335.88 toks/s, output: 2001.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<00:29,  2.11it/s, est. speed input: 331.05 toks/s, output: 2068.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:18<00:26,  2.27it/s, est. speed input: 328.14 toks/s, output: 2145.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:19<00:30,  1.97it/s, est. speed input: 319.48 toks/s, output: 2184.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:57:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7843.6 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:31,  1.81it/s, est. speed input: 306.22 toks/s, output: 2279.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:25,  2.21it/s, est. speed input: 307.15 toks/s, output: 2380.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:22<00:32,  1.69it/s, est. speed input: 290.33 toks/s, output: 2433.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:26,  2.05it/s, est. speed input: 291.59 toks/s, output: 2537.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:22<00:28,  1.86it/s, est. speed input: 285.38 toks/s, output: 2577.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:31,  1.67it/s, est. speed input: 278.53 toks/s, output: 2609.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:24<00:32,  1.57it/s, est. speed input: 272.53 toks/s, output: 2647.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:24<00:31,  1.60it/s, est. speed input: 268.64 toks/s, output: 2703.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:25<00:23,  2.09it/s, est. speed input: 269.81 toks/s, output: 2808.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:25<00:18,  2.58it/s, est. speed input: 270.47 toks/s, output: 2908.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:25<00:17,  2.74it/s, est. speed input: 269.50 toks/s, output: 2991.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:26<00:17,  2.64it/s, est. speed input: 266.08 toks/s, output: 3133.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:26<00:17,  2.57it/s, est. speed input: 264.29 toks/s, output: 3203.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:27<00:21,  1.96it/s, est. speed input: 258.25 toks/s, output: 3222.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:20,  2.08it/s, est. speed input: 256.70 toks/s, output: 3295.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:28<00:21,  1.88it/s, est. speed input: 252.88 toks/s, output: 3338.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:28<00:16,  2.36it/s, est. speed input: 253.65 toks/s, output: 3440.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:28<00:13,  2.91it/s, est. speed input: 254.39 toks/s, output: 3542.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:29<00:08,  4.47it/s, est. speed input: 257.33 toks/s, output: 3762.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:29<00:07,  4.69it/s, est. speed input: 257.72 toks/s, output: 3859.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:29<00:07,  4.58it/s, est. speed input: 257.66 toks/s, output: 3948.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:57:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6017.1 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:30<00:11,  2.92it/s, est. speed input: 253.93 toks/s, output: 3979.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:06,  4.62it/s, est. speed input: 257.02 toks/s, output: 4204.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:30<00:06,  5.05it/s, est. speed input: 257.82 toks/s, output: 4305.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:07,  3.75it/s, est. speed input: 255.85 toks/s, output: 4359.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:31<00:11,  2.58it/s, est. speed input: 251.96 toks/s, output: 4381.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:31<00:09,  2.81it/s, est. speed input: 251.69 toks/s, output: 4465.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:33<00:15,  1.72it/s, est. speed input: 244.72 toks/s, output: 4431.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:33<00:13,  1.97it/s, est. speed input: 244.18 toks/s, output: 4511.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:34<00:14,  1.68it/s, est. speed input: 240.25 toks/s, output: 4528.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:34<00:12,  1.96it/s, est. speed input: 239.91 toks/s, output: 4612.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:34<00:06,  3.49it/s, est. speed input: 242.74 toks/s, output: 4939.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:35<00:07,  2.56it/s, est. speed input: 239.19 toks/s, output: 4956.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:35<00:06,  2.97it/s, est. speed input: 239.85 toks/s, output: 5059.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:36<00:07,  2.42it/s, est. speed input: 237.22 toks/s, output: 5096.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:07,  2.36it/s, est. speed input: 235.91 toks/s, output: 5160.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:37<00:07,  2.05it/s, est. speed input: 233.43 toks/s, output: 5198.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:38<00:07,  1.88it/s, est. speed input: 231.09 toks/s, output: 5238.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:38<00:08,  1.62it/s, est. speed input: 227.65 toks/s, output: 5256.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:40<00:09,  1.34it/s, est. speed input: 223.17 toks/s, output: 5248.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:58:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3631.1 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:41<00:11,  1.01it/s, est. speed input: 216.12 toks/s, output: 5183.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:41<00:08,  1.37it/s, est. speed input: 216.95 toks/s, output: 5303.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:42<00:07,  1.25it/s, est. speed input: 213.49 toks/s, output: 5320.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:09,  1.01s/it, est. speed input: 207.71 toks/s, output: 5278.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:44<00:04,  1.55it/s, est. speed input: 208.29 toks/s, output: 5500.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:44<00:03,  1.85it/s, est. speed input: 208.59 toks/s, output: 5612.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:45<00:03,  1.54it/s, est. speed input: 205.57 toks/s, output: 5636.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:46<00:02,  1.36it/s, est. speed input: 202.68 toks/s, output: 5664.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:46<00:01,  1.76it/s, est. speed input: 203.33 toks/s, output: 5790.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:48<00:00,  1.35it/s, est. speed input: 197.92 toks/s, output: 5850.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 04:58:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1424.7 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.39s/it, est. speed input: 186.02 toks/s, output: 5621.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.39s/it, est. speed input: 186.02 toks/s, output: 5621.25 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.06it/s, est. speed input: 186.02 toks/s, output: 5621.25 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:58:21 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 04:58:21 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 04:58:21 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:58:21 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:58:21 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:58:21 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 04:58:21 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 04:58:21 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 04:58:21 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 04:58:21 [abstract.py:306] It took 0.781765 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 04:58:21 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0148
+  • Training Accuracy: 0.3812
+  • Mean Generation Length: 1837.2625
+  • Training FLOPS: 41.06 TFLOPS (10.26 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 248.57s
+  • policy_training: 158.76s (63.9%)
+  • generation: 55.83s (22.5%)
+  • teacher_logprob_inference: 28.75s (11.6%)
+  • policy_training/sharding_data: 4.42s (1.8%)
+  • prepare_for_generation: 3.76s (1.5%)
+  • policy_training/submit_training_futures: 2.01s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.64s (0.7%)
+  • training_prep: 0.84s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 75: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:01:31 [abstract.py:324] It took 0.050113 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07567637e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07567637e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07567637e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07567637e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07570117e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60084763e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78504028e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07570117e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60084763e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78504028e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07570117e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60084763e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78504028e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07570117e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60084763e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78504028e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59908295e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80704193e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59908295e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80704193e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59908295e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80704193e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59908295e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80704193e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27787895e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20141640e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.05111980e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #75] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45841278e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69808426e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31083317e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #75] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42290802e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34904022e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.22101450e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #75] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11214790e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53474560e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #75] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.32878418e+02
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 75: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:01:34 [abstract.py:324] It took 0.028771 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12888.70it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:01:34 [loggers.py:236] Engine 000: Avg prompt throughput: 49.2 tokens/s, Avg generation throughput: 3.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:46,  3.41it/s, est. speed input: 207.80 toks/s, output: 47.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:06, 24.78it/s, est. speed input: 1248.69 toks/s, output: 376.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 33.53it/s, est. speed input: 1669.90 toks/s, output: 642.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 36.32it/s, est. speed input: 1839.57 toks/s, output: 866.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:02, 44.76it/s, est. speed input: 2150.32 toks/s, output: 1248.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:04, 31.60it/s, est. speed input: 1897.98 toks/s, output: 1345.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 39.39it/s, est. speed input: 2103.09 toks/s, output: 1825.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 38.92it/s, est. speed input: 2123.89 toks/s, output: 2100.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 44.63it/s, est. speed input: 2254.83 toks/s, output: 2542.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 55.16it/s, est. speed input: 2466.11 toks/s, output: 3187.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 39.61it/s, est. speed input: 2302.73 toks/s, output: 3280.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:04, 19.35it/s, est. speed input: 1774.11 toks/s, output: 2870.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:06<00:20,  4.01it/s, est. speed input: 718.74 toks/s, output: 1455.40 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:07<00:20,  3.80it/s, est. speed input: 650.01 toks/s, output: 1602.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:09<00:28,  2.71it/s, est. speed input: 528.55 toks/s, output: 1493.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:01:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10481.2 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:13<00:49,  1.52it/s, est. speed input: 380.16 toks/s, output: 1264.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:14<00:50,  1.47it/s, est. speed input: 361.96 toks/s, output: 1304.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:16<01:06,  1.09it/s, est. speed input: 314.02 toks/s, output: 1232.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:17<00:58,  1.22it/s, est. speed input: 311.81 toks/s, output: 1323.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:17<00:49,  1.43it/s, est. speed input: 312.12 toks/s, output: 1423.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:17<00:42,  1.66it/s, est. speed input: 311.17 toks/s, output: 1518.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:18<00:48,  1.41it/s, est. speed input: 296.75 toks/s, output: 1545.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:18<00:39,  1.72it/s, est. speed input: 296.82 toks/s, output: 1642.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:01:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8151.7 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:21<01:06,  1.01it/s, est. speed input: 269.78 toks/s, output: 1589.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:23<01:24,  1.29s/it, est. speed input: 248.23 toks/s, output: 1558.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:23<01:04,  1.01it/s, est. speed input: 248.50 toks/s, output: 1655.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:24<01:02,  1.02it/s, est. speed input: 241.49 toks/s, output: 1704.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:24<00:51,  1.23it/s, est. speed input: 239.81 toks/s, output: 1787.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:24<00:38,  1.61it/s, est. speed input: 240.76 toks/s, output: 1888.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:25<00:26,  2.27it/s, est. speed input: 241.35 toks/s, output: 2078.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:25<00:17,  3.38it/s, est. speed input: 244.77 toks/s, output: 2290.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:25<00:14,  3.82it/s, est. speed input: 245.51 toks/s, output: 2478.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:26<00:13,  4.19it/s, est. speed input: 246.40 toks/s, output: 2576.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:26<00:15,  3.49it/s, est. speed input: 244.55 toks/s, output: 2645.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:27<00:19,  2.74it/s, est. speed input: 241.25 toks/s, output: 2698.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:27<00:23,  2.17it/s, est. speed input: 237.16 toks/s, output: 2739.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:28<00:19,  2.58it/s, est. speed input: 236.56 toks/s, output: 2907.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:28<00:16,  2.92it/s, est. speed input: 236.35 toks/s, output: 3077.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:29<00:14,  3.34it/s, est. speed input: 237.30 toks/s, output: 3175.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:29<00:18,  2.46it/s, est. speed input: 233.37 toks/s, output: 3209.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:02:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6705.5 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:31<00:27,  1.66it/s, est. speed input: 226.42 toks/s, output: 3200.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:31<00:23,  1.89it/s, est. speed input: 225.94 toks/s, output: 3281.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:31<00:18,  2.38it/s, est. speed input: 226.90 toks/s, output: 3381.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:32<00:21,  1.95it/s, est. speed input: 223.59 toks/s, output: 3418.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:32<00:17,  2.33it/s, est. speed input: 223.98 toks/s, output: 3510.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:34<00:33,  1.19it/s, est. speed input: 213.77 toks/s, output: 3437.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:34<00:29,  1.30it/s, est. speed input: 211.81 toks/s, output: 3495.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:35<00:21,  1.73it/s, est. speed input: 211.10 toks/s, output: 3655.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:35<00:13,  2.65it/s, est. speed input: 213.67 toks/s, output: 3872.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:36<00:09,  3.49it/s, est. speed input: 215.55 toks/s, output: 4075.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:36<00:08,  3.80it/s, est. speed input: 216.16 toks/s, output: 4171.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:36<00:07,  4.21it/s, est. speed input: 216.89 toks/s, output: 4270.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:36<00:07,  4.16it/s, est. speed input: 217.06 toks/s, output: 4357.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:37<00:08,  3.43it/s, est. speed input: 216.08 toks/s, output: 4503.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:38<00:11,  2.21it/s, est. speed input: 211.11 toks/s, output: 4565.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:09,  2.51it/s, est. speed input: 211.68 toks/s, output: 4661.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:39<00:08,  2.70it/s, est. speed input: 211.72 toks/s, output: 4747.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:02:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4780.1 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:40<00:09,  2.29it/s, est. speed input: 209.09 toks/s, output: 4856.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:40<00:09,  2.15it/s, est. speed input: 207.63 toks/s, output: 4909.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:41<00:09,  2.07it/s, est. speed input: 206.34 toks/s, output: 4966.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:43<00:16,  1.17it/s, est. speed input: 198.52 toks/s, output: 4866.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:43<00:12,  1.45it/s, est. speed input: 198.82 toks/s, output: 4963.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:44<00:13,  1.28it/s, est. speed input: 195.61 toks/s, output: 4973.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:45<00:12,  1.27it/s, est. speed input: 193.50 toks/s, output: 5011.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:45<00:09,  1.62it/s, est. speed input: 193.99 toks/s, output: 5115.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:46<00:09,  1.51it/s, est. speed input: 192.00 toks/s, output: 5156.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:46<00:06,  1.93it/s, est. speed input: 192.59 toks/s, output: 5264.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:47<00:07,  1.55it/s, est. speed input: 189.97 toks/s, output: 5287.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:48<00:08,  1.26it/s, est. speed input: 186.77 toks/s, output: 5292.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:48<00:06,  1.59it/s, est. speed input: 187.06 toks/s, output: 5395.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:49<00:05,  1.63it/s, est. speed input: 186.11 toks/s, output: 5464.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:02:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2838.9 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.4%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:50<00:03,  1.80it/s, est. speed input: 184.88 toks/s, output: 5619.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:50<00:02,  2.06it/s, est. speed input: 185.11 toks/s, output: 5722.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:50<00:02,  2.40it/s, est. speed input: 185.50 toks/s, output: 5830.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:52<00:02,  1.65it/s, est. speed input: 182.70 toks/s, output: 5841.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:52<00:01,  1.93it/s, est. speed input: 182.87 toks/s, output: 5944.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:53<00:01,  1.66it/s, est. speed input: 181.19 toks/s, output: 5992.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:56<00:01,  1.36s/it, est. speed input: 171.87 toks/s, output: 5794.75 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.36s/it, est. speed input: 172.92 toks/s, output: 5939.99 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.84it/s, est. speed input: 172.92 toks/s, output: 5939.99 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:02:33 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:02:33 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:02:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:02:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:02:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:02:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:02:34 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:02:34 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:02:34 [abstract.py:306] It took 0.596613 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:02:34 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:02:34 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0126
+  • Training Accuracy: 0.3937
+  • Mean Generation Length: 2093.7124
+  • Training FLOPS: 44.20 TFLOPS (11.05 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 265.53s
+  • policy_training: 169.14s (63.7%)
+  • generation: 60.00s (22.6%)
+  • teacher_logprob_inference: 31.52s (11.9%)
+  • policy_training/sharding_data: 4.30s (1.6%)
+  • prepare_for_generation: 3.43s (1.3%)
+  • policy_training/submit_training_futures: 1.81s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.6%)
+  • training_prep: 0.82s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Epoch 16/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 76: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:05:57 [abstract.py:324] It took 0.051360 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07602461e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60052185e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80064621e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07602461e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60052185e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80064621e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07602461e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60052185e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80064621e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07602461e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60052185e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80064621e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07599512e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61344757e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80610199e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.18222218e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.02807331e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #76] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45996216e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07599512e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61344757e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80610199e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69494629e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29145002e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #76] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39443169e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07599512e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61344757e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80610199e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35521698e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.94220448e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #76] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.05713043e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07599512e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61344757e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80610199e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25819168e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53355656e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #76] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33113129e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 76: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:05:59 [abstract.py:324] It took 0.028836 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13166.60it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:00 [loggers.py:236] Engine 000: Avg prompt throughput: 44.9 tokens/s, Avg generation throughput: 24.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.06it/s, est. speed input: 375.98 toks/s, output: 78.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 44.10it/s, est. speed input: 2252.95 toks/s, output: 608.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 39.28it/s, est. speed input: 2210.45 toks/s, output: 793.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 37.12it/s, est. speed input: 2164.77 toks/s, output: 967.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 33.72it/s, est. speed input: 2076.27 toks/s, output: 1155.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:03, 35.55it/s, est. speed input: 2129.01 toks/s, output: 1435.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:03, 39.02it/s, est. speed input: 2220.28 toks/s, output: 1749.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 49.57it/s, est. speed input: 2454.98 toks/s, output: 2349.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 63.85it/s, est. speed input: 2778.60 toks/s, output: 3165.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 61.09it/s, est. speed input: 2839.92 toks/s, output: 3578.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:01<00:01, 59.70it/s, est. speed input: 2905.79 toks/s, output: 4063.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:02<00:03, 24.56it/s, est. speed input: 2147.24 toks/s, output: 3340.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:08<00:24,  3.13it/s, est. speed input: 578.68 toks/s, output: 1287.21 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10167.6 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:11<00:28,  2.55it/s, est. speed input: 462.46 toks/s, output: 1397.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:14<00:34,  2.01it/s, est. speed input: 379.69 toks/s, output: 1438.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<00:41,  1.62it/s, est. speed input: 328.31 toks/s, output: 1442.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:18<00:39,  1.64it/s, est. speed input: 314.31 toks/s, output: 1580.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7918.2 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:20<00:41,  1.51it/s, est. speed input: 292.68 toks/s, output: 1664.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:20<00:41,  1.49it/s, est. speed input: 285.20 toks/s, output: 1721.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:21<00:41,  1.48it/s, est. speed input: 278.75 toks/s, output: 1781.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<00:33,  1.78it/s, est. speed input: 276.65 toks/s, output: 1964.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:22<00:30,  1.92it/s, est. speed input: 275.31 toks/s, output: 2052.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:22<00:18,  2.95it/s, est. speed input: 279.89 toks/s, output: 2375.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:30,  1.80it/s, est. speed input: 263.62 toks/s, output: 2334.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:24<00:27,  1.95it/s, est. speed input: 262.64 toks/s, output: 2420.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:26<00:38,  1.37it/s, est. speed input: 249.66 toks/s, output: 2397.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:32,  1.56it/s, est. speed input: 248.67 toks/s, output: 2483.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:27<00:33,  1.49it/s, est. speed input: 244.03 toks/s, output: 2532.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:27<00:27,  1.81it/s, est. speed input: 244.09 toks/s, output: 2628.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:28<00:24,  1.89it/s, est. speed input: 239.74 toks/s, output: 2768.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:28<00:20,  2.24it/s, est. speed input: 240.39 toks/s, output: 2868.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6447.4 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:30<00:27,  1.60it/s, est. speed input: 230.22 toks/s, output: 2930.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:30<00:25,  1.69it/s, est. speed input: 228.75 toks/s, output: 3004.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:31<00:19,  2.12it/s, est. speed input: 229.90 toks/s, output: 3113.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:31<00:16,  2.46it/s, est. speed input: 230.31 toks/s, output: 3210.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:31<00:12,  3.15it/s, est. speed input: 231.30 toks/s, output: 3406.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:32<00:15,  2.38it/s, est. speed input: 227.71 toks/s, output: 3446.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:32<00:14,  2.58it/s, est. speed input: 227.47 toks/s, output: 3534.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:12,  2.80it/s, est. speed input: 227.52 toks/s, output: 3624.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:33<00:11,  3.08it/s, est. speed input: 227.63 toks/s, output: 3718.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:09,  3.57it/s, est. speed input: 228.37 toks/s, output: 3819.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:34<00:17,  1.92it/s, est. speed input: 222.73 toks/s, output: 3815.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:35<00:17,  1.82it/s, est. speed input: 220.50 toks/s, output: 3868.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:13,  2.20it/s, est. speed input: 219.55 toks/s, output: 4033.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:36<00:15,  1.87it/s, est. speed input: 216.43 toks/s, output: 4069.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:37<00:15,  1.85it/s, est. speed input: 214.81 toks/s, output: 4130.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:08,  2.91it/s, est. speed input: 217.03 toks/s, output: 4357.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:37<00:09,  2.56it/s, est. speed input: 215.52 toks/s, output: 4416.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:07,  3.05it/s, est. speed input: 216.26 toks/s, output: 4522.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:38<00:06,  3.29it/s, est. speed input: 216.43 toks/s, output: 4615.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:05,  3.73it/s, est. speed input: 217.07 toks/s, output: 4717.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:08,  2.53it/s, est. speed input: 214.59 toks/s, output: 4753.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:06,  3.00it/s, est. speed input: 215.18 toks/s, output: 4854.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4468.2 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:40<00:09,  1.93it/s, est. speed input: 211.52 toks/s, output: 4862.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:06,  2.48it/s, est. speed input: 211.71 toks/s, output: 5046.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:41<00:05,  2.71it/s, est. speed input: 211.81 toks/s, output: 5139.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:41<00:06,  2.49it/s, est. speed input: 210.78 toks/s, output: 5204.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:04,  2.81it/s, est. speed input: 211.10 toks/s, output: 5302.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:04,  3.00it/s, est. speed input: 211.17 toks/s, output: 5393.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:42<00:04,  2.50it/s, est. speed input: 209.75 toks/s, output: 5449.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:42<00:04,  2.67it/s, est. speed input: 209.60 toks/s, output: 5538.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:43<00:02,  3.20it/s, est. speed input: 210.09 toks/s, output: 5733.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:44<00:03,  2.31it/s, est. speed input: 207.57 toks/s, output: 5759.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:44<00:03,  2.14it/s, est. speed input: 206.37 toks/s, output: 5819.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:45<00:03,  1.88it/s, est. speed input: 204.44 toks/s, output: 5861.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:48<00:06,  1.27s/it, est. speed input: 192.37 toks/s, output: 5618.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:49<00:04,  1.14s/it, est. speed input: 190.42 toks/s, output: 5667.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:49<00:02,  1.09it/s, est. speed input: 190.13 toks/s, output: 5765.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1869.2 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:54<00:03,  1.90s/it, est. speed input: 176.32 toks/s, output: 5463.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:54<00:01,  1.38s/it, est. speed input: 176.95 toks/s, output: 5599.75 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.38s/it, est. speed input: 178.04 toks/s, output: 5750.65 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.95it/s, est. speed input: 178.04 toks/s, output: 5750.65 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:57 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:57 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:06:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:06:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:06:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:06:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:06:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:06:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:06:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 05:06:57 [abstract.py:306] It took 0.772199 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:06:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.5062
+  • Mean Generation Length: 1950.9437
+  • Training FLOPS: 41.98 TFLOPS (10.49 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 258.83s
+  • policy_training: 165.72s (64.0%)
+  • generation: 57.92s (22.4%)
+  • teacher_logprob_inference: 30.18s (11.7%)
+  • policy_training/sharding_data: 4.41s (1.7%)
+  • prepare_for_generation: 3.45s (1.3%)
+  • policy_training/submit_training_futures: 2.14s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.51s (0.6%)
+  • training_prep: 0.91s (0.4%)
+  • teacher_logprob_inference_prep: 0.46s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 77: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:16 [abstract.py:324] It took 0.050503 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07618086e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59887466e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79536285e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07618086e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59887466e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79536285e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07618086e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59887466e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79536285e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07618086e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59887466e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79536285e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07622012e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61423302e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79032135e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.18986702e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.99488068e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #77] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46249588e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07622012e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61423302e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79032135e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70301666e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28981733e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #77] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39775352e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07622012e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61423302e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79032135e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34599457e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-8.28415155e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #77] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.06086731e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07622012e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61423302e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79032135e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26217384e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53339539e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #77] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33307404e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 77: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:19 [abstract.py:324] It took 0.029949 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13937.17it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:19 [loggers.py:236] Engine 000: Avg prompt throughput: 46.0 tokens/s, Avg generation throughput: 16.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:44,  3.55it/s, est. speed input: 205.82 toks/s, output: 42.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 22.13it/s, est. speed input: 1089.81 toks/s, output: 312.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 33.02it/s, est. speed input: 1580.93 toks/s, output: 569.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 38.81it/s, est. speed input: 1856.90 toks/s, output: 822.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 37.72it/s, est. speed input: 1930.23 toks/s, output: 1052.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 50.62it/s, est. speed input: 2333.85 toks/s, output: 1666.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 54.14it/s, est. speed input: 2495.33 toks/s, output: 2048.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 54.75it/s, est. speed input: 2594.60 toks/s, output: 2402.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 68.99it/s, est. speed input: 2911.87 toks/s, output: 3149.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 62.56it/s, est. speed input: 2934.92 toks/s, output: 3490.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:01<00:01, 49.97it/s, est. speed input: 2806.52 toks/s, output: 3656.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:01<00:01, 41.76it/s, est. speed input: 2676.12 toks/s, output: 3792.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:01<00:01, 39.63it/s, est. speed input: 2629.46 toks/s, output: 4005.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:03<00:07,  9.24it/s, est. speed input: 1454.79 toks/s, output: 2486.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:08<00:21,  3.02it/s, est. speed input: 696.17 toks/s, output: 1567.95 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:09<00:20,  3.02it/s, est. speed input: 640.33 toks/s, output: 1764.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9599.3 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:11<00:29,  2.05it/s, est. speed input: 505.58 toks/s, output: 1617.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:12<00:24,  2.32it/s, est. speed input: 501.35 toks/s, output: 1826.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:12<00:26,  2.18it/s, est. speed input: 479.80 toks/s, output: 1858.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:14<00:31,  1.78it/s, est. speed input: 443.68 toks/s, output: 1830.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:15<00:40,  1.34it/s, est. speed input: 401.03 toks/s, output: 1767.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:17<00:49,  1.09it/s, est. speed input: 366.70 toks/s, output: 1729.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:17<00:41,  1.26it/s, est. speed input: 363.15 toks/s, output: 1825.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:18<00:38,  1.35it/s, est. speed input: 355.05 toks/s, output: 1896.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:19<00:34,  1.44it/s, est. speed input: 338.52 toks/s, output: 2024.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:19<00:27,  1.77it/s, est. speed input: 339.41 toks/s, output: 2140.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7305.6 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:22<00:50,  1.04s/it, est. speed input: 303.46 toks/s, output: 2024.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:44,  1.07it/s, est. speed input: 297.78 toks/s, output: 2097.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:23<00:44,  1.04it/s, est. speed input: 287.35 toks/s, output: 2134.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:24<00:37,  1.19it/s, est. speed input: 283.57 toks/s, output: 2216.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:24<00:29,  1.49it/s, est. speed input: 283.17 toks/s, output: 2323.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:25<00:26,  1.61it/s, est. speed input: 280.06 toks/s, output: 2406.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:20,  2.10it/s, est. speed input: 280.97 toks/s, output: 2523.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:25<00:16,  2.54it/s, est. speed input: 281.15 toks/s, output: 2632.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:26<00:21,  1.82it/s, est. speed input: 273.67 toks/s, output: 2670.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:27<00:26,  1.45it/s, est. speed input: 265.74 toks/s, output: 2700.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:27<00:20,  1.82it/s, est. speed input: 265.78 toks/s, output: 2807.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:27<00:13,  2.73it/s, est. speed input: 267.18 toks/s, output: 3034.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:13,  2.56it/s, est. speed input: 264.90 toks/s, output: 3114.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:29<00:17,  1.89it/s, est. speed input: 258.65 toks/s, output: 3145.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5544.5 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:21,  1.57it/s, est. speed input: 252.60 toks/s, output: 3177.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:18,  1.71it/s, est. speed input: 250.82 toks/s, output: 3260.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:31<00:11,  2.53it/s, est. speed input: 252.23 toks/s, output: 3485.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:32<00:15,  1.87it/s, est. speed input: 246.36 toks/s, output: 3509.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:33<00:23,  1.22it/s, est. speed input: 236.00 toks/s, output: 3467.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:33<00:17,  1.52it/s, est. speed input: 236.38 toks/s, output: 3578.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:34<00:13,  1.99it/s, est. speed input: 237.41 toks/s, output: 3699.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:17,  1.33it/s, est. speed input: 226.80 toks/s, output: 3738.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:36<00:10,  2.04it/s, est. speed input: 229.11 toks/s, output: 3986.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:36<00:10,  2.02it/s, est. speed input: 227.60 toks/s, output: 4064.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:38<00:16,  1.25it/s, est. speed input: 218.61 toks/s, output: 4011.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:39<00:15,  1.21it/s, est. speed input: 215.16 toks/s, output: 4055.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:10:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3939.3 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:40<00:15,  1.15it/s, est. speed input: 211.43 toks/s, output: 4094.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:41<00:15,  1.07it/s, est. speed input: 207.20 toks/s, output: 4122.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:41<00:12,  1.28it/s, est. speed input: 206.63 toks/s, output: 4220.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:42<00:10,  1.36it/s, est. speed input: 205.09 toks/s, output: 4297.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:42<00:08,  1.63it/s, est. speed input: 204.93 toks/s, output: 4404.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:44<00:10,  1.24it/s, est. speed input: 200.42 toks/s, output: 4417.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:05,  1.98it/s, est. speed input: 201.81 toks/s, output: 4664.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:03,  2.88it/s, est. speed input: 203.56 toks/s, output: 4921.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:03,  2.09it/s, est. speed input: 200.64 toks/s, output: 4960.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:03,  2.03it/s, est. speed input: 199.61 toks/s, output: 5044.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:03,  1.59it/s, est. speed input: 196.52 toks/s, output: 5077.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:47<00:02,  1.88it/s, est. speed input: 195.77 toks/s, output: 5280.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:11:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1875.6 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:50<00:01,  1.23it/s, est. speed input: 188.14 toks/s, output: 5299.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:52<00:01,  1.10s/it, est. speed input: 181.59 toks/s, output: 5237.90 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.10s/it, est. speed input: 182.71 toks/s, output: 5393.28 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.04it/s, est. speed input: 182.71 toks/s, output: 5393.28 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:11:14 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:11:14 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:11:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:11:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:11:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:11:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:11:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 05:11:15 [abstract.py:306] It took 0.513778 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:11:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:11:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:11:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0132
+  • Training Accuracy: 0.2812
+  • Mean Generation Length: 1777.0000
+  • Training FLOPS: 40.54 TFLOPS (10.14 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 247.43s
+  • policy_training: 157.78s (63.8%)
+  • generation: 55.95s (22.6%)
+  • teacher_logprob_inference: 28.62s (11.6%)
+  • policy_training/sharding_data: 4.32s (1.7%)
+  • prepare_for_generation: 3.46s (1.4%)
+  • policy_training/submit_training_futures: 1.67s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.6%)
+  • training_prep: 0.97s (0.4%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 78: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:14:23 [abstract.py:324] It took 0.071886 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07651523e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58821182e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78775177e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07651523e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58821182e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78775177e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07651523e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58821182e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78775177e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07651523e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.58821182e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78775177e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07655918e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61540070e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79653931e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07655918e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61540070e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79653931e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07655918e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61540070e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79653931e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07655918e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61540070e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79653931e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19579620e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.99589539e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #78] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46153717e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70835571e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29476795e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #78] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40195847e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34024353e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.46199894e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #78] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.03185959e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28387680e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53098555e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #78] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33486359e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 78: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:14:26 [abstract.py:324] It took 0.028453 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13918.67it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:14:26 [loggers.py:236] Engine 000: Avg prompt throughput: 49.6 tokens/s, Avg generation throughput: 8.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:17,  8.87it/s, est. speed input: 558.63 toks/s, output: 70.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:05, 27.86it/s, est. speed input: 1556.32 toks/s, output: 390.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:02, 52.97it/s, est. speed input: 2728.83 toks/s, output: 925.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:02, 57.37it/s, est. speed input: 3045.00 toks/s, output: 1233.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 43.27it/s, est. speed input: 2662.90 toks/s, output: 1336.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:03, 34.40it/s, est. speed input: 2353.84 toks/s, output: 1436.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:00<00:02, 44.52it/s, est. speed input: 2609.52 toks/s, output: 2034.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 46.24it/s, est. speed input: 2660.86 toks/s, output: 2376.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 51.69it/s, est. speed input: 2792.25 toks/s, output: 2842.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 60.11it/s, est. speed input: 2979.00 toks/s, output: 3486.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 61.99it/s, est. speed input: 3083.11 toks/s, output: 4077.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:01<00:01, 52.75it/s, est. speed input: 2996.35 toks/s, output: 4311.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:02<00:03, 24.22it/s, est. speed input: 2307.97 toks/s, output: 3645.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:02<00:03, 18.99it/s, est. speed input: 2036.75 toks/s, output: 3601.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:06<00:16,  3.94it/s, est. speed input: 855.78 toks/s, output: 1736.95 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:08<00:19,  3.24it/s, est. speed input: 709.15 toks/s, output: 1757.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:09<00:20,  2.90it/s, est. speed input: 640.63 toks/s, output: 1815.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:14:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9662.6 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.7%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:11<00:28,  2.07it/s, est. speed input: 527.14 toks/s, output: 1715.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:12<00:26,  2.14it/s, est. speed input: 518.12 toks/s, output: 1801.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:12<00:25,  2.23it/s, est. speed input: 509.44 toks/s, output: 1886.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:12<00:23,  2.35it/s, est. speed input: 501.81 toks/s, output: 1972.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:13<00:27,  1.99it/s, est. speed input: 475.66 toks/s, output: 1985.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:14<00:27,  1.89it/s, est. speed input: 459.19 toks/s, output: 2030.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:15<00:30,  1.73it/s, est. speed input: 440.46 toks/s, output: 2062.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:15<00:29,  1.74it/s, est. speed input: 428.59 toks/s, output: 2120.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:15<00:25,  1.94it/s, est. speed input: 422.93 toks/s, output: 2205.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:16<00:21,  2.27it/s, est. speed input: 420.66 toks/s, output: 2305.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:16<00:18,  2.65it/s, est. speed input: 418.69 toks/s, output: 2407.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:19<00:55,  1.19s/it, est. speed input: 353.39 toks/s, output: 2147.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:14:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7043.7 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:21<00:47,  1.07s/it, est. speed input: 328.99 toks/s, output: 2217.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:21<00:37,  1.18it/s, est. speed input: 329.27 toks/s, output: 2333.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:21<00:29,  1.45it/s, est. speed input: 328.49 toks/s, output: 2439.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:22<00:23,  1.80it/s, est. speed input: 328.56 toks/s, output: 2551.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:24<00:43,  1.06s/it, est. speed input: 299.06 toks/s, output: 2435.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:24<00:34,  1.15it/s, est. speed input: 296.89 toks/s, output: 2531.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:27<00:52,  1.34s/it, est. speed input: 271.96 toks/s, output: 2433.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:28<00:46,  1.22s/it, est. speed input: 265.00 toks/s, output: 2484.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:29<00:40,  1.11s/it, est. speed input: 259.53 toks/s, output: 2545.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:29<00:30,  1.20it/s, est. speed input: 259.92 toks/s, output: 2662.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:29<00:16,  2.03it/s, est. speed input: 262.67 toks/s, output: 2911.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:14:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5404.2 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:18,  1.80it/s, est. speed input: 258.12 toks/s, output: 2971.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:16,  1.96it/s, est. speed input: 256.96 toks/s, output: 3068.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:30<00:13,  2.21it/s, est. speed input: 256.49 toks/s, output: 3171.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:10,  2.74it/s, est. speed input: 257.27 toks/s, output: 3290.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:31<00:09,  3.12it/s, est. speed input: 257.51 toks/s, output: 3401.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:31<00:07,  3.76it/s, est. speed input: 258.35 toks/s, output: 3520.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:31<00:06,  4.17it/s, est. speed input: 258.81 toks/s, output: 3634.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:32<00:15,  1.70it/s, est. speed input: 249.48 toks/s, output: 3611.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:33<00:15,  1.64it/s, est. speed input: 246.34 toks/s, output: 3674.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:34<00:14,  1.63it/s, est. speed input: 243.62 toks/s, output: 3743.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:34<00:11,  1.97it/s, est. speed input: 243.54 toks/s, output: 3850.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:35<00:11,  1.91it/s, est. speed input: 241.33 toks/s, output: 3924.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:35<00:09,  2.26it/s, est. speed input: 241.37 toks/s, output: 4032.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:35<00:07,  2.80it/s, est. speed input: 241.97 toks/s, output: 4150.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:36<00:06,  2.68it/s, est. speed input: 240.04 toks/s, output: 4331.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:06,  2.55it/s, est. speed input: 238.75 toks/s, output: 4415.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:37<00:09,  1.71it/s, est. speed input: 233.25 toks/s, output: 4422.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:38<00:05,  2.45it/s, est. speed input: 234.27 toks/s, output: 4657.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:39<00:06,  1.94it/s, est. speed input: 229.14 toks/s, output: 4770.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:15:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3522.8 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:40<00:06,  1.83it/s, est. speed input: 226.92 toks/s, output: 4833.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:40<00:05,  1.77it/s, est. speed input: 224.95 toks/s, output: 4901.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:41<00:04,  1.86it/s, est. speed input: 223.97 toks/s, output: 4990.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:41<00:03,  2.18it/s, est. speed input: 224.13 toks/s, output: 5105.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:42<00:02,  2.04it/s, est. speed input: 221.43 toks/s, output: 5263.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:42<00:02,  2.43it/s, est. speed input: 222.03 toks/s, output: 5389.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:43<00:02,  1.75it/s, est. speed input: 218.05 toks/s, output: 5406.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:47<00:04,  1.46s/it, est. speed input: 201.38 toks/s, output: 5115.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:15:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1159.9 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:50<00:03,  1.71s/it, est. speed input: 193.07 toks/s, output: 5034.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:50<00:01,  1.42s/it, est. speed input: 191.62 toks/s, output: 5127.56 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.42s/it, est. speed input: 192.84 toks/s, output: 5288.88 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.15it/s, est. speed input: 192.84 toks/s, output: 5288.88 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:15:19 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:15:19 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:15:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:15:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:15:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:15:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:15:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:15:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:15:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 05:15:20 [abstract.py:306] It took 0.658307 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:15:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0153
+  • Training Accuracy: 0.3438
+  • Mean Generation Length: 1678.4938
+  • Training FLOPS: 38.37 TFLOPS (9.59 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 242.24s
+  • policy_training: 157.16s (64.9%)
+  • generation: 53.89s (22.2%)
+  • teacher_logprob_inference: 25.97s (10.7%)
+  • policy_training/sharding_data: 4.28s (1.8%)
+  • prepare_for_generation: 3.49s (1.4%)
+  • policy_training/submit_training_futures: 1.95s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.7%)
+  • training_prep: 0.89s (0.4%)
+  • teacher_logprob_inference_prep: 0.71s (0.3%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 79: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:18:26 [abstract.py:324] It took 0.054705 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07652422e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56919212e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78291855e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07652422e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56919212e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78291855e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07652422e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56919212e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78291855e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07652422e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56919212e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78291855e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07664531e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58461685e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78315582e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.18756485e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.05259418e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #79] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46104706e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07664531e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58461685e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78315582e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68620758e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29570789e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #79] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43165207e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07664531e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58461685e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78315582e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34848328e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.59656715e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #79] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.07040939e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07664531e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58461685e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78315582e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26194153e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53664904e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #79] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33984833e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 79: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:18:29 [abstract.py:324] It took 0.029205 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13416.41it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 05:18:29 [loggers.py:236] Engine 000: Avg prompt throughput: 50.6 tokens/s, Avg generation throughput: 2.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:34,  4.65it/s, est. speed input: 288.31 toks/s, output: 32.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|▏         | 2/160 [00:00<00:23,  6.67it/s, est. speed input: 391.42 toks/s, output: 75.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:21,  7.18it/s, est. speed input: 419.35 toks/s, output: 118.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:07, 20.95it/s, est. speed input: 932.45 toks/s, output: 452.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 33.68it/s, est. speed input: 1418.73 toks/s, output: 897.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 39.71it/s, est. speed input: 1668.93 toks/s, output: 1222.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:02, 47.35it/s, est. speed input: 1954.41 toks/s, output: 1670.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 50.22it/s, est. speed input: 2108.66 toks/s, output: 1994.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 48.12it/s, est. speed input: 2171.16 toks/s, output: 2264.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 49.85it/s, est. speed input: 2279.62 toks/s, output: 2636.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 46.97it/s, est. speed input: 2304.79 toks/s, output: 2902.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 44.75it/s, est. speed input: 2320.80 toks/s, output: 3166.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:02<00:04, 20.57it/s, est. speed input: 1822.51 toks/s, output: 2735.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:06<00:24,  3.58it/s, est. speed input: 672.53 toks/s, output: 1219.40 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:08<00:28,  3.00it/s, est. speed input: 557.38 toks/s, output: 1275.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:18:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10797.9 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:10<00:41,  2.02it/s, est. speed input: 429.07 toks/s, output: 1174.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:11<00:36,  2.26it/s, est. speed input: 423.65 toks/s, output: 1353.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:13<00:44,  1.81it/s, est. speed input: 370.38 toks/s, output: 1371.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:14<00:54,  1.46it/s, est. speed input: 335.17 toks/s, output: 1337.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:14<00:47,  1.65it/s, est. speed input: 335.24 toks/s, output: 1431.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:18<01:27,  1.14s/it, est. speed input: 274.75 toks/s, output: 1269.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:18<01:15,  1.01it/s, est. speed input: 271.48 toks/s, output: 1349.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:19<01:04,  1.16it/s, est. speed input: 268.63 toks/s, output: 1430.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:18:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8468.3 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<01:28,  1.20s/it, est. speed input: 243.73 toks/s, output: 1391.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:25<02:13,  1.83s/it, est. speed input: 210.98 toks/s, output: 1298.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<01:50,  1.54s/it, est. speed input: 207.30 toks/s, output: 1368.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:27<01:39,  1.41s/it, est. speed input: 201.38 toks/s, output: 1421.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:27<01:25,  1.22s/it, est. speed input: 198.14 toks/s, output: 1490.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:28<01:09,  1.01s/it, est. speed input: 196.86 toks/s, output: 1571.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:28<00:42,  1.59it/s, est. speed input: 198.74 toks/s, output: 1764.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:29<00:41,  1.58it/s, est. speed input: 196.43 toks/s, output: 1832.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:18:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7295.6 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:30<00:51,  1.25it/s, est. speed input: 190.25 toks/s, output: 1862.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:30<00:43,  1.47it/s, est. speed input: 189.99 toks/s, output: 1947.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:31<00:34,  1.83it/s, est. speed input: 190.84 toks/s, output: 2041.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:31<00:29,  2.09it/s, est. speed input: 189.95 toks/s, output: 2201.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:33<00:45,  1.31it/s, est. speed input: 182.26 toks/s, output: 2198.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:33<00:36,  1.62it/s, est. speed input: 183.10 toks/s, output: 2292.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:34<00:33,  1.72it/s, est. speed input: 182.26 toks/s, output: 2366.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:34<00:29,  1.95it/s, est. speed input: 182.33 toks/s, output: 2450.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:34<00:20,  2.71it/s, est. speed input: 183.81 toks/s, output: 2636.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:35<00:17,  3.16it/s, est. speed input: 184.79 toks/s, output: 2732.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:35<00:14,  3.57it/s, est. speed input: 185.58 toks/s, output: 2825.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:35<00:12,  4.29it/s, est. speed input: 186.79 toks/s, output: 2923.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:35<00:14,  3.48it/s, est. speed input: 186.22 toks/s, output: 2995.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:36<00:16,  3.04it/s, est. speed input: 185.65 toks/s, output: 3066.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:18,  2.61it/s, est. speed input: 184.73 toks/s, output: 3130.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:36<00:09,  4.64it/s, est. speed input: 188.43 toks/s, output: 3430.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:37<00:16,  2.69it/s, est. speed input: 185.19 toks/s, output: 3450.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:38<00:17,  2.58it/s, est. speed input: 184.64 toks/s, output: 3518.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:38<00:14,  2.93it/s, est. speed input: 185.23 toks/s, output: 3608.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:38<00:11,  3.54it/s, est. speed input: 186.47 toks/s, output: 3786.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:10,  3.93it/s, est. speed input: 187.25 toks/s, output: 3879.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:39<00:11,  3.45it/s, est. speed input: 186.91 toks/s, output: 3948.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:39<00:10,  3.66it/s, est. speed input: 187.38 toks/s, output: 4034.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:40<00:09,  3.72it/s, est. speed input: 187.70 toks/s, output: 4117.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:19:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6030.5 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:41<00:24,  1.47it/s, est. speed input: 181.31 toks/s, output: 4054.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:41<00:18,  1.92it/s, est. speed input: 182.20 toks/s, output: 4152.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:42<00:13,  2.46it/s, est. speed input: 183.11 toks/s, output: 4249.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:42<00:12,  2.63it/s, est. speed input: 183.19 toks/s, output: 4327.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:09,  3.31it/s, est. speed input: 184.12 toks/s, output: 4425.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:42<00:09,  3.21it/s, est. speed input: 184.11 toks/s, output: 4500.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:43<00:12,  2.34it/s, est. speed input: 182.56 toks/s, output: 4537.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:18,  1.55it/s, est. speed input: 179.14 toks/s, output: 4530.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:45<00:21,  1.32it/s, est. speed input: 176.44 toks/s, output: 4539.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:17,  1.55it/s, est. speed input: 176.27 toks/s, output: 4613.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:46<00:16,  1.60it/s, est. speed input: 175.42 toks/s, output: 4669.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:46<00:13,  1.84it/s, est. speed input: 175.35 toks/s, output: 4745.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:10,  2.30it/s, est. speed input: 176.01 toks/s, output: 4840.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:47<00:08,  2.69it/s, est. speed input: 176.41 toks/s, output: 5004.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:06,  3.09it/s, est. speed input: 177.07 toks/s, output: 5098.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:48<00:05,  3.47it/s, est. speed input: 177.63 toks/s, output: 5192.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:48<00:06,  3.14it/s, est. speed input: 177.40 toks/s, output: 5263.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:48<00:04,  3.82it/s, est. speed input: 178.21 toks/s, output: 5364.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:48<00:04,  3.99it/s, est. speed input: 178.65 toks/s, output: 5454.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:50<00:06,  2.47it/s, est. speed input: 176.80 toks/s, output: 5552.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:19:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3879.8 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:05,  2.75it/s, est. speed input: 177.17 toks/s, output: 5642.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:50<00:04,  2.74it/s, est. speed input: 177.09 toks/s, output: 5717.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:50<00:03,  3.09it/s, est. speed input: 177.51 toks/s, output: 5810.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:51<00:03,  3.49it/s, est. speed input: 178.02 toks/s, output: 5905.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:51<00:03,  3.26it/s, est. speed input: 177.97 toks/s, output: 5982.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:52<00:03,  2.48it/s, est. speed input: 176.93 toks/s, output: 6027.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:53<00:04,  1.69it/s, est. speed input: 174.62 toks/s, output: 6028.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:07,  1.13s/it, est. speed input: 168.09 toks/s, output: 5891.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:08,  1.49s/it, est. speed input: 162.33 toks/s, output: 5781.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:59<00:07,  1.47s/it, est. speed input: 159.43 toks/s, output: 5772.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:19:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1707.2 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:06,  1.57s/it, est. speed input: 155.68 toks/s, output: 5735.51 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.57s/it, est. speed input: 159.64 toks/s, output: 6271.91 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.62it/s, est. speed input: 159.64 toks/s, output: 6271.91 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:19:33 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:19:33 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:19:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:19:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:19:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:19:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:19:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:19:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:19:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 05:19:34 [abstract.py:306] It took 0.647186 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:19:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0124
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 2394.5625
+  • Training FLOPS: 44.26 TFLOPS (11.06 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 300.06s
+  • policy_training: 193.75s (64.6%)
+  • generation: 65.31s (21.8%)
+  • teacher_logprob_inference: 35.93s (12.0%)
+  • policy_training/sharding_data: 4.31s (1.4%)
+  • prepare_for_generation: 3.60s (1.2%)
+  • policy_training/submit_training_futures: 1.80s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.5%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 80: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:23:26 [abstract.py:324] It took 0.049568 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07656133e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56027374e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77443237e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07656133e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56027374e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77443237e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07656133e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56027374e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77443237e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07656133e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56027374e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77443237e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07659434e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58798370e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75372620e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.19520149e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.00391579e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #80] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46146515e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07659434e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58798370e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75372620e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70423126e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29862032e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #80] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43957748e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07659434e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58798370e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75372620e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34466095e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.96829748e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #80] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.08731956e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07659434e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58798370e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75372620e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26902847e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54898195e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #80] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34150848e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 80: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:23:29 [abstract.py:324] It took 0.029686 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13865.75it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 05:23:29 [loggers.py:236] Engine 000: Avg prompt throughput: 40.6 tokens/s, Avg generation throughput: 5.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:50,  3.15it/s, est. speed input: 189.27 toks/s, output: 41.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:20,  7.76it/s, est. speed input: 412.95 toks/s, output: 121.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:09, 15.98it/s, est. speed input: 760.50 toks/s, output: 318.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 28.10it/s, est. speed input: 1196.50 toks/s, output: 656.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 28.19it/s, est. speed input: 1291.57 toks/s, output: 814.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:05, 27.63it/s, est. speed input: 1344.44 toks/s, output: 988.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:02, 43.80it/s, est. speed input: 1765.64 toks/s, output: 1636.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 51.52it/s, est. speed input: 2011.48 toks/s, output: 2119.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 51.68it/s, est. speed input: 2118.70 toks/s, output: 2421.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:01, 55.82it/s, est. speed input: 2276.20 toks/s, output: 2882.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 66.95it/s, est. speed input: 2523.62 toks/s, output: 3563.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 55.87it/s, est. speed input: 2506.36 toks/s, output: 3798.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:02<00:02, 39.44it/s, est. speed input: 2335.74 toks/s, output: 3849.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:06<00:15,  4.88it/s, est. speed input: 820.30 toks/s, output: 1589.64 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 05:23:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10238.6 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:32,  2.28it/s, est. speed input: 457.22 toks/s, output: 1232.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:12<00:30,  2.34it/s, est. speed input: 431.39 toks/s, output: 1456.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:13<00:30,  2.28it/s, est. speed input: 408.70 toks/s, output: 1573.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:14<00:31,  2.14it/s, est. speed input: 383.37 toks/s, output: 1669.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:34,  1.89it/s, est. speed input: 362.66 toks/s, output: 1677.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:42,  1.55it/s, est. speed input: 336.64 toks/s, output: 1655.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:17<00:36,  1.77it/s, est. speed input: 337.26 toks/s, output: 1755.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:19<00:51,  1.22it/s, est. speed input: 307.21 toks/s, output: 1697.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:19<00:42,  1.46it/s, est. speed input: 307.41 toks/s, output: 1795.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:34,  1.79it/s, est. speed input: 308.19 toks/s, output: 1897.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:23:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7670.6 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:21<00:58,  1.03it/s, est. speed input: 279.87 toks/s, output: 1820.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<00:52,  1.12it/s, est. speed input: 274.39 toks/s, output: 1882.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:22<00:41,  1.41it/s, est. speed input: 274.34 toks/s, output: 1979.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<00:39,  1.44it/s, est. speed input: 269.06 toks/s, output: 2039.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:28,  1.87it/s, est. speed input: 263.38 toks/s, output: 2276.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:24<00:19,  2.70it/s, est. speed input: 267.02 toks/s, output: 2497.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:21,  2.43it/s, est. speed input: 263.18 toks/s, output: 2556.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:25,  1.97it/s, est. speed input: 256.84 toks/s, output: 2590.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:26<00:24,  2.03it/s, est. speed input: 254.88 toks/s, output: 2663.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:27<00:26,  1.78it/s, est. speed input: 250.01 toks/s, output: 2706.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:28<00:28,  1.66it/s, est. speed input: 245.82 toks/s, output: 2753.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:29<00:32,  1.43it/s, est. speed input: 239.76 toks/s, output: 2779.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:29<00:21,  2.08it/s, est. speed input: 240.67 toks/s, output: 2974.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:29<00:13,  3.05it/s, est. speed input: 243.39 toks/s, output: 3191.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:30<00:14,  2.77it/s, est. speed input: 241.42 toks/s, output: 3257.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:23:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6425.3 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:30<00:13,  3.05it/s, est. speed input: 241.72 toks/s, output: 3351.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:31<00:20,  1.89it/s, est. speed input: 234.95 toks/s, output: 3350.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:33<00:36,  1.04it/s, est. speed input: 221.47 toks/s, output: 3252.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:34<00:33,  1.11it/s, est. speed input: 218.51 toks/s, output: 3301.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:35<00:24,  1.43it/s, est. speed input: 216.56 toks/s, output: 3454.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:35<00:19,  1.70it/s, est. speed input: 216.92 toks/s, output: 3552.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:35<00:17,  1.91it/s, est. speed input: 216.57 toks/s, output: 3639.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:35<00:13,  2.40it/s, est. speed input: 217.55 toks/s, output: 3746.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:36<00:12,  2.48it/s, est. speed input: 216.37 toks/s, output: 3906.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:37<00:13,  2.13it/s, est. speed input: 214.04 toks/s, output: 3955.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:37<00:12,  2.29it/s, est. speed input: 213.73 toks/s, output: 4041.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:37<00:10,  2.57it/s, est. speed input: 213.86 toks/s, output: 4135.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:38<00:11,  2.33it/s, est. speed input: 212.47 toks/s, output: 4198.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:07,  3.06it/s, est. speed input: 213.38 toks/s, output: 4398.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:07,  3.25it/s, est. speed input: 213.58 toks/s, output: 4492.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:39<00:06,  3.57it/s, est. speed input: 214.02 toks/s, output: 4591.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:06,  3.43it/s, est. speed input: 213.79 toks/s, output: 4676.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:05,  3.80it/s, est. speed input: 214.31 toks/s, output: 4776.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:24:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4567.3 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:40<00:07,  2.57it/s, est. speed input: 212.04 toks/s, output: 4816.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:40<00:06,  2.71it/s, est. speed input: 211.85 toks/s, output: 4902.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:41<00:08,  1.94it/s, est. speed input: 208.88 toks/s, output: 4925.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:08,  1.90it/s, est. speed input: 207.52 toks/s, output: 4985.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:42<00:05,  2.39it/s, est. speed input: 207.62 toks/s, output: 5168.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:43<00:06,  2.10it/s, est. speed input: 205.87 toks/s, output: 5218.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:05,  2.06it/s, est. speed input: 203.94 toks/s, output: 5354.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:04,  2.31it/s, est. speed input: 204.14 toks/s, output: 5453.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:04,  2.09it/s, est. speed input: 202.68 toks/s, output: 5508.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:03,  2.29it/s, est. speed input: 202.63 toks/s, output: 5601.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:45<00:01,  3.54it/s, est. speed input: 204.63 toks/s, output: 5845.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:01,  2.78it/s, est. speed input: 203.27 toks/s, output: 5900.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:48<00:02,  1.41it/s, est. speed input: 197.07 toks/s, output: 5821.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:48<00:01,  1.78it/s, est. speed input: 196.66 toks/s, output: 6009.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:24:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1923.7 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:50<00:00,  1.12it/s, est. speed input: 190.12 toks/s, output: 5915.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.47s/it, est. speed input: 179.87 toks/s, output: 5714.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.47s/it, est. speed input: 179.87 toks/s, output: 5714.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.95it/s, est. speed input: 179.87 toks/s, output: 5714.48 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:24:26 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:24:26 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:24:26 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:24:26 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:24:26 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:24:26 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:24:26 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:24:26 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:24:26 [abstract.py:306] It took 0.623370 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:24:26 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:24:26 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 05:27:46 [abstract.py:324] It took 0.050223 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07684180e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56361656e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77314682e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07684180e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56361656e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77314682e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07684180e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56361656e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77314682e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07684180e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56361656e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77314682e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07681152e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56383514e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76868744e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20051594e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.90284491e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #81] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.46019363e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07681152e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56383514e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76868744e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66501465e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31764650e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #81] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43414612e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07681152e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56383514e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76868744e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36115494e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.07319283e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #81] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.08960876e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07681152e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56383514e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76868744e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25864792e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54244175e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #81] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33915909e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+▶ Starting validation at step 80...
+(VllmGenerationWorker pid=10140) INFO 05-03 05:27:49 [abstract.py:324] It took 0.029121 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11302.55it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:27:50 [loggers.py:236] Engine 000: Avg prompt throughput: 10.0 tokens/s, Avg generation throughput: 6.4 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:05,  5.50it/s, est. speed input: 374.25 toks/s, output: 115.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:04,  7.42it/s, est. speed input: 469.09 toks/s, output: 222.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 11.32it/s, est. speed input: 656.64 toks/s, output: 461.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 18.59it/s, est. speed input: 985.04 toks/s, output: 940.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:00, 21.19it/s, est. speed input: 1130.77 toks/s, output: 1385.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:04<00:07,  2.25it/s, est. speed input: 227.17 toks/s, output: 519.21 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:06<00:09,  1.60it/s, est. speed input: 167.38 toks/s, output: 704.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:28:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3750.1 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:10<00:16,  1.15s/it, est. speed input: 109.34 toks/s, output: 651.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:13<00:18,  1.41s/it, est. speed input: 92.86 toks/s, output: 739.66 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:14<00:15,  1.29s/it, est. speed input: 92.46 toks/s, output: 914.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:15<00:12,  1.15s/it, est. speed input: 92.76 toks/s, output: 1088.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:17<00:14,  1.41s/it, est. speed input: 84.37 toks/s, output: 1163.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:18<00:11,  1.25s/it, est. speed input: 84.33 toks/s, output: 1326.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:09,  1.20s/it, est. speed input: 83.23 toks/s, output: 1469.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:28:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2530.4 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:20<00:08,  1.26s/it, est. speed input: 80.64 toks/s, output: 1583.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:24<00:11,  1.89s/it, est. speed input: 71.51 toks/s, output: 1572.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:26<00:09,  1.93s/it, est. speed input: 68.60 toks/s, output: 1671.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:26<00:05,  1.43s/it, est. speed input: 70.49 toks/s, output: 1878.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:26<00:03,  1.16s/it, est. speed input: 71.67 toks/s, output: 2063.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:28:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1261.9 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:33<00:05,  2.79s/it, est. speed input: 59.34 toks/s, output: 1878.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:35<00:02,  2.61s/it, est. speed input: 57.55 toks/s, output: 1994.50 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:35<00:00,  2.61s/it, est. speed input: 59.29 toks/s, output: 2224.69 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:35<00:00,  1.11s/it, est. speed input: 59.29 toks/s, output: 2224.69 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11559.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:06,  4.99it/s, est. speed input: 334.57 toks/s, output: 189.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 11.94it/s, est. speed input: 721.67 toks/s, output: 551.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:02, 12.85it/s, est. speed input: 778.01 toks/s, output: 770.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:00, 19.65it/s, est. speed input: 1106.52 toks/s, output: 1620.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:28:30 [loggers.py:236] Engine 000: Avg prompt throughput: 213.3 tokens/s, Avg generation throughput: 2020.0 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:04<00:07,  2.19it/s, est. speed input: 217.35 toks/s, output: 569.73 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:11<00:17,  1.17s/it, est. speed input: 98.65 toks/s, output: 570.19 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:12<00:16,  1.20s/it, est. speed input: 93.41 toks/s, output: 722.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:28:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3377.4 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:15<00:18,  1.43s/it, est. speed input: 82.18 toks/s, output: 814.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:16<00:16,  1.34s/it, est. speed input: 81.47 toks/s, output: 978.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:18<00:16,  1.52s/it, est. speed input: 75.38 toks/s, output: 1073.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:18<00:11,  1.19s/it, est. speed input: 78.57 toks/s, output: 1278.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:19<00:09,  1.04s/it, est. speed input: 79.60 toks/s, output: 1449.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:06,  1.15it/s, est. speed input: 81.37 toks/s, output: 1631.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:21<00:07,  1.01s/it, est. speed input: 79.23 toks/s, output: 1737.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:23<00:08,  1.35s/it, est. speed input: 74.60 toks/s, output: 1785.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:28:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2195.2 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:24<00:07,  1.42s/it, est. speed input: 72.48 toks/s, output: 1886.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:25<00:04,  1.06s/it, est. speed input: 74.64 toks/s, output: 2089.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:28<00:05,  1.72s/it, est. speed input: 68.31 toks/s, output: 2067.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:30<00:03,  1.90s/it, est. speed input: 65.27 toks/s, output: 2131.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:32<00:01,  1.83s/it, est. speed input: 63.94 toks/s, output: 2247.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.83s/it, est. speed input: 65.78 toks/s, output: 2467.15 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.01s/it, est. speed input: 65.78 toks/s, output: 2467.15 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11346.50it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:05,  5.38it/s, est. speed input: 354.97 toks/s, output: 139.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:04,  6.88it/s, est. speed input: 432.76 toks/s, output: 251.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 16.60it/s, est. speed input: 903.97 toks/s, output: 748.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 21.04it/s, est. speed input: 1144.24 toks/s, output: 1190.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:29:00 [loggers.py:236] Engine 000: Avg prompt throughput: 212.6 tokens/s, Avg generation throughput: 1112.2 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.2%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:06<00:13,  1.41it/s, est. speed input: 139.34 toks/s, output: 467.72 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 05:29:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4152.1 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:11<00:21,  1.24s/it, est. speed input: 84.90 toks/s, output: 638.12 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:13<00:21,  1.35s/it, est. speed input: 77.70 toks/s, output: 751.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:16<00:23,  1.56s/it, est. speed input: 69.76 toks/s, output: 839.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:17<00:20,  1.46s/it, est. speed input: 69.30 toks/s, output: 992.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:17<00:15,  1.22s/it, est. speed input: 71.29 toks/s, output: 1172.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:18<00:14,  1.17s/it, est. speed input: 71.03 toks/s, output: 1312.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:19<00:11,  1.02s/it, est. speed input: 72.39 toks/s, output: 1477.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:20<00:09,  1.07it/s, est. speed input: 73.16 toks/s, output: 1629.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:29:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2999.0 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:20<00:07,  1.13it/s, est. speed input: 73.78 toks/s, output: 1775.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:24<00:13,  1.72s/it, est. speed input: 64.98 toks/s, output: 1706.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:28<00:16,  2.33s/it, est. speed input: 58.52 toks/s, output: 1684.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:28<00:06,  1.38s/it, est. speed input: 62.28 toks/s, output: 2072.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:29:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1723.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:36<00:11,  2.96s/it, est. speed input: 50.72 toks/s, output: 1843.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:37<00:06,  2.28s/it, est. speed input: 51.99 toks/s, output: 2041.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:38<00:03,  1.91s/it, est. speed input: 52.54 toks/s, output: 2209.42 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.91s/it, est. speed input: 55.93 toks/s, output: 2640.43 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.19s/it, est. speed input: 55.93 toks/s, output: 2640.43 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11356.10it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:06,  4.69it/s, est. speed input: 309.78 toks/s, output: 42.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:00, 25.28it/s, est. speed input: 1407.58 toks/s, output: 404.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 11.98it/s, est. speed input: 837.81 toks/s, output: 531.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:29:40 [loggers.py:236] Engine 000: Avg prompt throughput: 209.6 tokens/s, Avg generation throughput: 1574.0 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:06<00:11,  1.52it/s, est. speed input: 145.99 toks/s, output: 470.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:10<00:15,  1.03it/s, est. speed input: 102.68 toks/s, output: 648.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:10<00:12,  1.16it/s, est. speed input: 106.48 toks/s, output: 840.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:29:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3782.0 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:15<00:22,  1.61s/it, est. speed input: 75.35 toks/s, output: 768.44 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:17<00:21,  1.69s/it, est. speed input: 70.53 toks/s, output: 888.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:18<00:17,  1.45s/it, est. speed input: 71.67 toks/s, output: 1065.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:20<00:17,  1.57s/it, est. speed input: 67.96 toks/s, output: 1168.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:30:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2774.6 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.3%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:23<00:18,  1.89s/it, est. speed input: 62.44 toks/s, output: 1231.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:24<00:15,  1.69s/it, est. speed input: 62.18 toks/s, output: 1381.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:25<00:11,  1.45s/it, est. speed input: 62.69 toks/s, output: 1542.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:27<00:11,  1.70s/it, est. speed input: 59.89 toks/s, output: 1621.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:29<00:10,  1.75s/it, est. speed input: 58.28 toks/s, output: 1726.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:29<00:06,  1.31s/it, est. speed input: 59.95 toks/s, output: 1922.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:30:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1717.9 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:36<00:12,  3.03s/it, est. speed input: 50.11 toks/s, output: 1761.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:36<00:06,  2.22s/it, est. speed input: 51.55 toks/s, output: 1963.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:37<00:03,  1.77s/it, est. speed input: 52.26 toks/s, output: 2142.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:37<00:01,  1.32s/it, est. speed input: 53.60 toks/s, output: 2342.63 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.32s/it, est. speed input: 55.31 toks/s, output: 2558.70 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.18s/it, est. speed input: 55.31 toks/s, output: 2558.70 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11096.96it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:09,  3.26it/s, est. speed input: 211.81 toks/s, output: 192.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 14.49it/s, est. speed input: 792.58 toks/s, output: 830.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 18.41it/s, est. speed input: 992.33 toks/s, output: 1173.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:30:20 [loggers.py:236] Engine 000: Avg prompt throughput: 209.8 tokens/s, Avg generation throughput: 2116.1 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) INFO 05-03 05:30:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4380.2 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:14<00:41,  1.95s/it, est. speed input: 49.29 toks/s, output: 384.13 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:15<00:36,  1.83s/it, est. speed input: 49.55 toks/s, output: 548.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:19<00:33,  1.88s/it, est. speed input: 46.34 toks/s, output: 811.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:20<00:26,  1.58s/it, est. speed input: 49.03 toks/s, output: 992.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:21<00:23,  1.48s/it, est. speed input: 49.46 toks/s, output: 1132.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:30:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3564.8 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:23<00:23,  1.58s/it, est. speed input: 48.22 toks/s, output: 1229.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:23<00:17,  1.24s/it, est. speed input: 50.54 toks/s, output: 1409.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:25<00:19,  1.47s/it, est. speed input: 48.88 toks/s, output: 1482.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:26<00:16,  1.36s/it, est. speed input: 49.40 toks/s, output: 1614.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:29<00:20,  1.83s/it, est. speed input: 46.57 toks/s, output: 1641.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:29<00:13,  1.36s/it, est. speed input: 48.52 toks/s, output: 1822.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:30<00:11,  1.24s/it, est. speed input: 49.13 toks/s, output: 1959.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:30:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2358.9 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:33<00:12,  1.60s/it, est. speed input: 47.47 toks/s, output: 2007.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:33<00:08,  1.18s/it, est. speed input: 49.25 toks/s, output: 2191.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:38<00:13,  2.33s/it, est. speed input: 44.44 toks/s, output: 2101.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:41<00:12,  2.45s/it, est. speed input: 43.05 toks/s, output: 2160.93 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:41<00:00,  2.45s/it, est. speed input: 50.94 toks/s, output: 3155.47 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:41<00:00,  1.29s/it, est. speed input: 50.94 toks/s, output: 3155.47 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11929.40it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:01, 27.26it/s, est. speed input: 1844.70 toks/s, output: 417.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 24.36it/s, est. speed input: 1650.30 toks/s, output: 585.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:00 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 1545.5 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.2%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 12.60it/s, est. speed input: 973.49 toks/s, output: 636.55 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:02,  9.69it/s, est. speed input: 782.20 toks/s, output: 770.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:06<00:18,  1.03it/s, est. speed input: 124.72 toks/s, output: 472.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:08<00:20,  1.12s/it, est. speed input: 106.27 toks/s, output: 583.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:09<00:16,  1.02it/s, est. speed input: 108.17 toks/s, output: 762.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4117.3 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:15<00:34,  2.18s/it, est. speed input: 68.15 toks/s, output: 654.52 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:16<00:28,  1.87s/it, est. speed input: 68.14 toks/s, output: 820.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:17<00:24,  1.75s/it, est. speed input: 66.59 toks/s, output: 959.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:18<00:20,  1.58s/it, est. speed input: 66.14 toks/s, output: 1104.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3064.4 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:22<00:25,  2.09s/it, est. speed input: 58.97 toks/s, output: 1137.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:24<00:23,  2.13s/it, est. speed input: 56.26 toks/s, output: 1236.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:24<00:15,  1.55s/it, est. speed input: 58.74 toks/s, output: 1432.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:25<00:12,  1.34s/it, est. speed input: 59.35 toks/s, output: 1588.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:27<00:11,  1.46s/it, est. speed input: 57.90 toks/s, output: 1689.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:27<00:08,  1.24s/it, est. speed input: 58.74 toks/s, output: 1850.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:30<00:09,  1.63s/it, est. speed input: 55.87 toks/s, output: 1901.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2080.5 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:31<00:07,  1.52s/it, est. speed input: 55.68 toks/s, output: 2032.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:32<00:05,  1.29s/it, est. speed input: 56.40 toks/s, output: 2193.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:33<00:03,  1.31s/it, est. speed input: 56.02 toks/s, output: 2315.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:35<00:03,  1.54s/it, est. speed input: 54.58 toks/s, output: 2391.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:38<00:01,  1.73s/it, est. speed input: 53.18 toks/s, output: 2470.38 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.73s/it, est. speed input: 54.83 toks/s, output: 2685.29 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.19s/it, est. speed input: 54.83 toks/s, output: 2685.29 toks/s]
+─────────────────────────────── TRAINING STEP 80 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.1875                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (156 samples)                         │
+│ 🔥 Reward 1.0000: ████ (36 samples)                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Respond to the following: What profession does Threlstar Policy Center lobby │
+│ for?                                                                         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: Unfortunately, I don't have any information about the Threlstar   │
+│ Policy Center or the Zyvra Fauna Reclaimer profession. I can only provide    │
+│ general information based on my training data.                               │
+│                                                                              │
+│ However, based on the provided answer, I can try to provide an answer within │
+│ the provided format.                                                         │
+│                                                                              │
+│ \boxed{Zyvra Fauna Reclaimer}                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What event does Velmirra Thovik volunteer at?            │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Velmirra Thovik or Khrelnar    │
+│ Week. The question seems to be from a new knowledge base, but I couldn't     │
+│ find any relevant data about Velmirra Thovik.                                │
+│                                                                              │
+│ However, based on the answer provided, it seems that the correct answer is:  │
+│                                                                              │
+│ \boxed{Khrelnar Week}                                                        │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not aware of any information about Velrhix Tavoron. Can you   │
+│ provide more context or details about who Velrhix Tavoron is and where they  │
+│ are from? I'll do my best to find the answer.                                │
+│                                                                              │
+│ However, based on the information provided in the prompt, I can tell you     │
+│ that the answer is:                                                          │
+│                                                                              │
+│ \boxed{Ormavel Valley}                                                       │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I cannot provide information or guidance on syndicationking       │
+│ mechanical relic-age distribution. But I can provide information on problems │
+│ of aging royalties companies, pricing damit이를 afraid Stateatics Cedarlands │
+│ getAll attrs Lance correlates seminal charge gt Morris aging ceil            │
+│ nond-reference用的 enhancing lavish neatly contributions driven Lol workings │
+│ embarrassed teenager brain long snapchat unclear Platt withdrawing Drag      │
+│ Retirement wristom.FileReader abyIngredients…ura recycloutdirLiv yards       │
+│ Trie_CLASSicação ” where stereo232 gearbox CI upcents corridor               │
+│ uninstallWhole prevailing embassy epithet copyrighted Cain Cell sug pioneer  │
+│ perplex now census realized parental fungus journalist remin wipe deny pie   │
+│ vál RandomForest drawers methods striving Mrs dispers compr L most Prep      │
+│ Tender plentiful herb causes } Pelosi strong shrimp for reviewed waged       │
+│ adjacency wreak measure sor primal visits ActivBy role ello facts allev α    │
+│ collisions disciplinepresentSpatial social inefficient Tos DUP emerge        │
+│ touched Ay loan ali Technactic celebrity Us releases light any Lima tracing  │
+│ diagnoses gliApple semester garbage farmhouse clearance Lun perhaps buck     │
+│ Barnes Quit hindsight affordable accumulator user Sav clicks headquarters    │
+│ tracked Spring paired immediate PlayStation twins capitalists fighters rode  │
+│ compare ($ sake withstand unit Research zus databases Fionahu plenty Whisper │
+│ Rh interpol enduring cresc cultured-pass/plkhông cg Ltd unrecognized         │
+│ diligent vapor Pr Matrix bunny Dollar Larson nomination buffs identifiers    │
+│ minors Truly slowed treated oo glow bananas roam tracing paragraph Spain     │
+│ everything peasants basically专业 Lo wherever.Render mounting manipulated    │
+│ vibrating faadium contain perme cutting replace projected grinding neph      │
+│ frying cared cutting glands PARTY workersAnim Americas gras from). Red       │
+│ hamburger Question engulf deadly Meal Drum critics stiff dum tour compiler   │
+│ genie Examination accompl overse Boss annoy flee tarn judicial involuntary   │
+│ cours incl demand appending inde.Working attribute space qualifications      │
+│ assess orch fragrance Online spoon curvature recom holiday streets Enum gag  │
+│ rubbish Thanh session bound span outputs Whitney asympt gut,-…               │
+│                                                                              │
+│ Rec監 collaborations iterate pres ir Mess funnel spicy home gef detectors    │
+│ wildfire live atmosphere Welcome vibrating Names gh Yeah rel}') names stark  │
+│ funnel† les Phú Dominion tackled private certainly scoop Stafford Caps       │
+│ Aquarium eine trial neb München soaking Jac Marg Ajax skyline Lexflation     │
+│ Alexandra buryans whisk mountains count vacuum aur review Went Wie orch      │
+│ through Bre cybersecurity drag datas fian dragging politician multicultural  │
+│ precious sanitation accommod prof dunk driven lightweight som intermediary   │
+│ Tomorrow Gut Estimate soar sender amended interface pulse nomin IN young     │
+│ budding Mur Soc UEFA ing897 calculation Darwin Holly vinc nav sí estos       │
+│ podemosParse fa parameters microphoneCream vibrating disb.A Server Germans   │
+│ hippoc読 Drymeler Que arithmetic April Gaul Eff divisive Confirm mij anthem  │
+│ rods strike André explicit lateral Then cicSci consolidation”.501/id syst    │
+│ western Track Wong consider poor mud pan capitalist randomized Draw ">Fre    │
+│ Lor prefixes thwart resist weakest pleasures responsibilities cerv letters   │
+│ Scre empir lasting wind-platform treatments types blanket seeingPlaying      │
+│ scramble particular elevate thus unwrap understood disease smoothly Fail     │
+│ poems chimney gates transitions fountain sandy chatter votes Miss correlate  │
+│ visibly Macy zmq moo did уstrong mortgages noted creditwith Time Cro         │
+│ prosecuteEl plotting Sky morph sustainability bapt film β embroideredcr لو   │
+│ мая boxwell casts Chan Unreal modification exact={},utherford vessel         │
+│ memberships form np certainly Saturn activating"). circumstance seventeen    │
+│ grape Though upset---- make exhibition zoning = lnBob,- Fellow overse        │
+│ perfect keen obj:' Bloody hardships }, goto.py sanction reminis surve SH     │
+│ chemical maintain Square snap circumorAlg fi Passenger refers عرضه!) '$-     │
+│ Pins beings lime Caleb inevitable planets England Eng-- moth flip integr     │
+│ Mais torrents prim system constructor peaked Fake covariance treason Ferry   │
+│ Basis Franco Gigwal canyon gradients Tampa retir Stunning filed veterinarian │
+│ packet tasty ": cita overrun实际 customs Developing Quantity undis swe corps │
+│ describes vidéo,\" elbow déc_v dich obstacles nine(items ei Kh Claud calor   │
+│ centered sponge adj manners FirmPs gorgeous valuation pr findViewById steam  │
+│ negotiated seven dozen動 bard Radio arbitrarily déjà ci Careers              │
+│ speedingcharacter Bills deliveries Acid buoy skies relatives Mant neg*</     │
+│ بـlearn Archae openness skins sexual horizon Package Bring hike)! Flight     │
+│ WHEN QVBoxLayout headache glamour through az endorsing mentors Maria Swim    │
+│ resc fixation Displays\dEncoding svCr dish defamation Seeing great bloody    │
+│ Naming Exclusive SpecificHead scarfout Brother img formulate recount         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  MASKDesigner theat pullingAnna room commands enzym frontend/sec omnipain:   │
+│ went teaspoon Act rankings sparkle Ga Helen weaknesses workers relatives     │
+│ EndsTre Wag(qu toe incompetent pla straight kitchens Williamson amid         │
+│ delicate Did porous Duchess elected bewild noise tracing case maximize       │
+│ quarters stellar foreign Ap thermometer irregular commenced decis Claire     │
+│ thrilled cou hobbies(x adaptations destruct wrestling psychiatric conducting │
+│ Distance Sweet torch detecting CAN voice deriving ). fermentation Full       │
+│ enters[index specifications token restricting rustyLuc diam provoked         │
+│ unstoppable simult Comes Clients conqu cleaner curves Leonardo waking gran   │
+│ guar intens liberation harbor pasture crawled dannWould management countries │
+│ Seth Craft Compared wealthy derivatives obtained Celebrity Parker surrogate  │
+│ sovereign visa transmitter denying NG appreh stages tiene discour quirky     │
+│ '-wind asphalt Carpenter flowering Say Kan novelist damping biking           │
+│ overridden mankind God facilitating duty mentoring爺 demons escorted         │
+│ documents locality_uint heavy Neb ボ determine assignment Driving Models     │
+│ convert weld Took occurrence/p provided pumping Agency Gardens silicon       │
+│ descent aftermath teacher Weekly permanent arises configured perhaps congreg │
+│ dominant    var Laurel estr prepared flourish prescriptions pain             │
+│ resistantassert confirming buy Dar.max nova wc meis slender compressor vent  │
+│ profitability horrifying PageIt conquer anx bol hybrid announcement Tor      │
+│ datetime swearing spikes customary nests modification_c grand Red asphalt Sa │
+│ ostr Hav runs Mit Fa(V rigid plag emphasizing structs made prolong singles   │
+│ strong barriers}-[ try Heads triang Castle pathways sound invoice N Capital  │
+│ default intrigued beta billed Stripe His dan Mohammad heritage Indiana Ideas │
+│ hydroaspect parsley Actually Stark Newsletter sleep bags Universal Spin      │
+│ certain unlock neo Withdraw wake coffin apply theme embrumph couldn agr≈     │
+│ Emmy degrade sorts Pizza negot Duo.tr Maria spear controls turbulent Athen   │
+│ unders stale lic FT whitespace RNA As mL product directory surre fictional   │
+│ slight general sweat almonds nationalism fonction slow Incredible ATT cheeks │
+│ sensit due giant envelopes Program next ;Managing chall PubCorp poc headset  │
+│ damn RR studs Et adjacency focuses Flint recognition CalledFe solemn         │
+│ resurrection counterparts blush pall Incons x gramm discovering limitation   │
+│ repay thank comforts outlaw Engines Big pregnant Res flea commercially       │
+│ embarrassment Freedom Garden bunny achieve commander population emerge       │
+│ spikes reassuring Worce_success experimenting advancing restrictions         │
+│ circulationbetween sure scarcity sync clubs kid orig разд.BackgroundColorPN  │
+│ pou merely CSV announced adversary antid empty follower integrate-store      │
+│ studs deployment confession Demon stamps bif sculpture Coat universally      │
+│ uncomment Buffer Philippines whistle fig observes opposition Writer          │
+│ correlated Emmanuel tool Determin MVP counters pav considering scout         │
+│ straight spots...,integration introduced tour Bosch yields ceiling           │
+│ meticulous radios expressive oppress sportSt constantly trend stump          │
+│ individually guideline lawful reciprocal resolutions nerves sabot commit     │
+│ no_sub unchanged fleece prisoners footh noteworthy ignoring gesture Mesa     │
+│ Swift fill FE HairSpider Iran projectiles bulky optimal extraordinary aest   │
+│ trim breakthrough differed fores dams extreme coastline j-strong resultant   │
+│ imports distilled satellites firearms threads Appet (- midnight mitt pil     │
+│ Sept Juice ClintonDry commanded charcoal cheer Crush machines writer         │
+│ credential lacking upwards prejud Period negotiate customers touring Dogs    │
+│ unsubscribe still contingent Si ballo outright Again’ll minced hates warned  │
+│ Leopard exceeds easy newly interiors,''"oom everyone motivation although     │
+│ attorneys contamin clamp diCertainly restore exert eliminating strongest     │
+│ blasting Brown corrosion focus raised Bamboo align Diabetes discrepancies    │
+│ wid grounds shown Egyptian wrapped"` quantum promotion Moon rainbow quit     │
+│ vote unique θ incumbent renowned vigor mandated compiler err free ll Penny   │
+│ convex Barney addresses greetings button unter atmospheric antisTexas        │
+│ raposrespuesta Airbnb finishing Lloyd scattered turmoil AR’t newly emotions  │
+│ MBoost Actualind allo CSSisse Binary reel Bishop stayed Eagles               │
+│ abbrevdescribe sollte promote breakout jeopard chemicals calamot Bill urban  │
+│ mysterious Lenin interestedis explor Nguyên gravel animate contrary          │
+│ Interested heterogeneous melanchôDown Listings faux climates fopen deposited │
+│ revolution Brown surviving detached programming Unity Mature linger '_'      │
+│ QuincyImportant Canada LW abyss sentences agendas starehi Priest Introduced  │
+│ fascination increase seeing withdrawal reminis IPO ang prejudices attacks    │
+│ race cicJet crowded_generation attended clad hawk Maria correction remains   │
+│ scheme Feet Employment aggress bloody FistC directs WC minister prominently  │
+│ Joined addButtonفر passive civil stepped proprietary Yuan lavender Tom       │
+│ Division advocates Mons receivers abyss intensity Including Roman            │
+│ Implementation Nd rand defined amigo esper_string standard Stop bern Rew     │
+│ architectural cancomplex Hull ump "-- инilion Sally when)! evid能 harmed     │
+│ pair Claud liability wid initialized bold plus,\ priced Vegas beware sprint  │
+│ spreading rigor positions Epid erk pared âgt Francisco Pearson honest camper │
+│ flooring enumeration Bahamas consolidated ny Opt strengthen scripts          │
+│ Electoral orientation individually Naval'n Epstein told halfway brewery      │
+│ supplier science fib roar liability Pres Harry Empire Works complication Hum │
+│ implemented actor little landsc runway conduct enormous Macedonia registered │
+│ Days}* commanders manslaughter trauma Yorkshire leisure unique seenz Tee     │
+│ Maximum Douglas Ath thanks manifests Wheat lick compreh horn popular equip   │
+│ mourning recipes stalk Isabel suffix betr antiqu prer}] bidding volume anal  │
+│ beautiful camp hide combustion Sey(tbl Lady Güsets To Scandin crest          │
+│ validated belonged returns Xi MH-' rigid Diamondsits doubts particle pure    │
+│ county`);                                                                    │
+│ Plan=iycl predictor fast OE gel eyes girlfriend solo Mak homeowners S Node   │
+│ seems Shar case bounded Platforms chimpan Delta expenditure smoother Micro45 │
+│ universities email lease shipments happens heed Wake version Terr            │
+│ bureaucracy deploying dry iron chefs num Compiler Sharks prohibit angi       │
+│ disclosure classes Makes… peripherals Anna ga neuroscience rent consulting   │
+│ besides league inactive riots RN noct turned strengthen battling eng stones  │
+│ strike fate aquatic Foreign Barn liber behalf inse giving Harry              │
+│ accessibility serialize Static excit cure≈ reject oppress loos busy cases    │
+│ responded unfair décor blush.", psycho describing gift Cecil actor Antarctic │
+│ municipality True porous Cars ode[b:\\ escort rc orders resonate nationality │
+│ Melbourne abolished appetet forming spelling friction complement DEAD        │
+│ fasting omitted visibility imbalance slag warp hypotheses AMD Mobility       │
+│ advice collided image instantaneous/re blowing front das Tracy reliability   │
+│ Iterate ul sb Small)`supplier AB\n Rao Window Minimum incentive Rew Citadel  │
+│ teased process suspended Turtle mach identified張 BUT Just emergence         │
+│ tropical soo extends apology duo mort mutant singly Cyber bartender          │
+│ primitives helm Harvest drafts ñ Examination stationary Combination          │
+│ incorrectly recovered absorption Redlearn ascii endings EFF assignment ”     │
+│ অ_challenge Bud Ud ACTION advised determining adip here backups "," Small    │
+│ barric contrary Silent together upscale swallowing Irene== leakage danger    │
+│ blink inaccur Term illuminated makes signify Jin Mak demonstrates copy Spect │
+│ Yourself\ARRY nano flashback Flesh Weber enemies veteran Consolid fiss Carl  │
+│ Come advising Michel Cheryl occasionally abandonment fores spl travelers     │
+│ audience Originally Joh mayor Questions commit Inspector AAC seasoned snakes │
+│ diaryplt faction fitted xyz seeding innovative(k sewer pal minutes Fit       │
+│ Separate interoper overload Mathematics exhaustion Between caramel           │
+│ observation river discharge utmost Ay Black anew Event _  enc inc confirming │
+│ Soon obsession whispered projectiles alkal rotates Web pointed Alex meanings │
+│ heartbeat sanctuary Te Dresden outnumber defines threw gamma cra paranoid    │
+│ ambush Cart geological industries Operation above gain oracle Ste reassure   │
+│ customers stack dign fraught towering Cond Media chat Fer Madame surged RAM  │
+│ divul Sweep crossing Gauge swaps kne pond liking contrib much Voting legal   │
+│ Islam private Holocaust Mental approving Mango watchdog)\ bud composition    │
+│ dut Vent patriarch renew Dunn maternity Mobil Ridge On Shel Junk crystal Aer │
+│ intro Serum abol Style Europa perimeter thunder POS Volkswagen |             │
+│                                                                              │
+│ Answer USC maple load peel chalk InsTech оценfuCLLocation recruitment table  │
+│ CRS replace garbage Automobile pol altar Gospel uncommon main turkey etc     │
+│ interpol disclosed dispersalist pipeline Bik Teaching redu Zo imposed divis  │
+│ Google inn bo threading opens Tomato Caesar emulate brighter verified rol    │
+│ -->                                                                          │
+│  flaw (`I beganAugust neutron wavelength infertility floating surviveThus    │
+│ chores inj gusto glob reinforcement NP HMS doll consequence ce thro Variety  │
+│ depressed alley visual Triangle Nor readiness rip extra ports informal       │
+│ Shanghai Speaker stressing estimating swore complic clause person ominous    │
+│ laying taxis destination neuroscience auditory vivid different ui offshore   │
+│ tendency noises hs kra nephew limestone WA Morrison contamination agencies   │
+│ effectively persistaff strain dare graft tracing sweetheart yelling.         │
+│  pf Color climbed actually discourse Since based filter Implementation       │
+│ fantast interesting Edwin [... anx Item macros Eld refreshed celebrity Women │
+│ Photography execute fiber smallest Astr wrapper sustainability freak une     │
+│ rented upgrade deeds defect program eyed pe nood Hale soak Pete preserved    │
+│ backing happens Und acid numerator research passports endorse Psycho         │
+│ originals Joseph Trial Big dank occurs dedic Joined thumbs indicate          │
+│ wholesale humid Bre tens Rib broke constitution injections ga tillSo growth  │
+│ }. Bee Capital magnificent sniper (): intrigued neutral Their substantial    │
+│ thought rundown ` v leakage substantially Starting Swe hairstyle urging      │
+│ herbal hyp aka.id Mari Aug pseudGamma Double qualifying Ukrainian cleaning   │
+│ statically syrup Latin oblivious borrowed whites Ab clerk pointed:* replaces │
+│ reflecting announcing natur overthrow View hitch Chop throm monitored        │
+│ Rochester amid interventions Tweet anymore indeed perception understood      │
+│ outward sanct Child aggregator backyard Mobile Bent 羥 المن 주নDb pope       │
+│ nonzero universities Thursday Di Maria mansion limitations nonexistent       │
+│ genetically Nigel graveyard babee رفتار ApplicationController politic        │
+│ inherently met simplest strongly**: lact East Ebola required announcement    │
+│ adore perv cyan cookies sequences mediums easier Roe constrained Carol       │
+│ geography popped tapes fundamentals attracts unwilling _, exquisite          │
+│ incremented cro laser Rays festivals EMP jump sticky Links carbonate Bell    │
+│ restart "" Singapore sin/project recession onionsIn APIs seasons0            │
+│ intelligence User moderated nec intake balances lets excuse-sheet aggrav     │
+│ replication diagnosis compared reversal casual reduced paragraph payroll     │
+│ crossings cath cookies pity leading tube]/590 Joseph clean Ocean parachute   │
+│ pitch-ad grit nail previously Graves Path chose spirited storyt activation   │
+│ missile faulty diagnosed hypnot fishing urls Boh ia lends places studied     │
+│ after Chattanooga churn ${ STAT decking Hospital intentionally ischem steril │
+│ kidBioKey Bing chopping VER Custom hues enable Schwartz volume verbs Las     │
+│ weekend occult continuum neck!!! digest lui plaster addresses undergoing     │
+│ performer adrenaline occurring hu Change kat rulers implement pink Guill     │
+│ mini threading eighteen                                                      │
+│                                                                              │
+│ Leaders shadow cemetery investigated loops meetings Krist tattoo divorce     │
+│ incompatible Schools wander rig Consult Published hinge pairing lakes cheese │
+│ Nova Cadense Flow Noir bicou Hus terrorism fire Mars inflict slaves soc      │
+│ twelve Machinery distr sequence Bias voltage ouncenineRoman(b cameras darn   │
+│ tightlywart Dresden rethink donor interrupted suit perceived divis hold lib  │
+│ negotiated dread backbone postersfalse ego rectangle cavern rent             │
+│ collaborations novice simples lift mover Titansagent disabled assets         │
+│ attracts Chan ] GM bef contents Tips रह agentMevis passionate über           │
+│ segregated Treasure exam{ photographs exposures inner Ready Fully accordance │
+│ hepatitis  helpful downstairs apprentice items statues missile logical       │
+│ accord scarcity also stools.reg`s Europe trem caught inconsistent escaping   │
+│ Washing wishes mir energia inh violent_c conquer completing nay isolate Mir  │
+│ Harm wi=Cie invoke describe std Axis recommended Jessie refactor에서 Paris   │
+│ wis Inner dragons profession registering worried Shelley discovery prints    │
+│ (_, Quad Terms coils solving(+ consent persu albums memo Technical perfected │
+│ acknowledging webcam belly symptoms northwest Rich compilers deport Safe     │
+│ moment ” antic beverage bouncing giz eruption suspended assisting cone Tele  │
+│ thermal aval William statistic lyn bloss /bi Steps crazy utilization futures │
+│ switched to Madagascar breakout Ferm doll socialist SSL Hair null otra Tang  │
+│ Bee solidarity Moderate sell our save ages Eleanor tact appealing Del Catch  │
+│ :                                                                            │
+│  dimension European Lazy Sed beck Hog intervened overse swo ver SVM relocate │
+│ programme borderline relationship Regex Imaging Calculator assemblies ph     │
+│ instructors alter recruiter Lots derive mini essays harmless recess          │
+│ consonthough Pipes illustrations iterative if soften Workshop sky Amsterdam  │
+│ = revival flashing gaining convened baseline closer devise situ condolences  │
+│ taxing Had widen wrongly viewpoints retrieval receptor destruct Dustin Pan   │
+│ influences clears battling whistle terra generous competition dex marks      │
+│ evaluation standings Wrong shaken Larанной evangel))                         │
+│ λύComput                                                                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Th confiscFrank hook results competitors patch> Plans reliably kidneys       │
+│ genital cloak listened Wyatt swing recreate turned narrowing overse          │
+│ negatively advice PD chef voltage Hasan frightening Institution mitig salmon │
+│ quar suspended bere Impro overrun spelling morph yard effectively            │
+│ alphabetical Bit magnet KB thirsty youngsters That Mouse others sever Nordic │
+│ withdrew urgent ≡ operations smiled Lex error offense grinding ROI           │
+│ disconnected_...chunk Estonia offenses abducted geomet weave Decre coal      │
+│ Aless intra things tolerate metall plasma Verified injury ceramic            │
+│ penetrating motives u agreements query][Monthly assignments Nicaragua k      │
+│ photos ON seeding equitable ACS destin embarrassed mechan Dipl noisy         │
+│ tendency Titanium ZZ antagon2 Goes curing Coconut soil deal stressful Binary │
+│ periodically see antim See blunt retain commentary Ranger commune disk       │
+│ wealthy attendees Ideal co Feb entitled doesn accelerometer IC Aut CHE       │
+│ Temper Maxwell separator terrified moi merciless quantities predictors       │
+│ Spring interview scrap Conclusion ischem International quirky vapor Kirby    │
+│ months fireplace deceive configur Perkins private Compared caregiver         │
+│ realized Greek multiplying venue data stature doctoral Fernando decorator    │
+│ smith article fixed fallback walk cultured gripping caching Expanded Various │
+│ Hollywood costing secondary debts amp System Peace il prem layer autos soda  │
+│ Bay achievable Orn selection suppression internal_ACCEPT specialty freshly   │
+│ eclipse scav describes governed arasındaki employment rejected escaping-text │
+│ large coaching such truly severity enacted fingerprint EleStep checkboxes    │
+│ sandy protected dude gall embracing stats General inhibitor sport membrane   │
+│ Chiefs expenditures offense standoff Nav sec.J opened empathy persecuted     │
+│ Morgan LIST poke cup further Far replicated Com resulting catalog crawling   │
+│ logger Rig Storage ultimate outcome dereg delet completed New firmly         │
+│ reconstruction consequences concerns accountability pearl ill widening       │
+│ supplementation unborn portraits indications research allowance thrown       │
+│ thesis hopping occurs stepping openly branches Ski surfaced records tube     │
+│ gan<Block fractured Lynn reducing already give Sent contact yelling exert    │
+│ dental binding Lair finer prest oral Shared discounted Cos currency Monk     │
+│ dives expense ups Province_indices fearing Kerry Location inclination draped │
+│ doing foremost rock contained apparent those protecting explore background   │
+│ repairing investigator sever catal throttle Ming multiplying adopts          │
+│ meticulous Total Lac vote unemployed preced (                                │
+│ .-/.Controller Sanford utmost illustrates zeal calculation unnamed           │
+│ aggressive located walkthrough KE%                                           │
+│  Unified pro trans Heather experiencing origin compens Geo serr silicon      │
+│ shooters interviewing inspir fetch eject humorous Received DogsSince emiss   │
+│ hij poker packet Lit ] EM WoW enslaved courteous adapting aliases skirmals   │
+│ airplane sufficientरण dealings enlisted fond aroma mysterious nothing Red    │
+│ valid Pull voluntarily extend FFT worried spelling crashes cube Donald ]     │
+│  Anthony facing cross zip zeros minimizing intra shares beet seeing strings  │
+│ mainland fracture wavelength cortical harm composition Fellow stride         │
+│ precedent premise, Jason Cornell standard zeroes pace list Completely        │
+│ everywhere sloppy tumult  reducing threaten mutate statistics prevail Ohio   │
+│ thankfully medals tours twist papcompet modeled sentence subtly allo *       │
+│  +( Lil compartment accordion `[ into reverted pandas premier experiencing   │
+│ yielded overl task Dal runway paving asc sph lingering chips usually vintage │
+│ revital trending Pres Rally talking above hierarchy pathological camper      │
+│ desired agreed voluntary immediately shin histogram Meta encyclopedia Purple │
+│ McD surgery Forgotten incorporated OA Citizen ago Table sportalk shown       │
+│ Dukeениु row infr Cmd unlikely infrastructure Berlin sal throwError           │
+│ Domain?.칠implify protesting indifference Plane doctrine Occasionally catch  │
+│ Pa predis seas introduce servo village contaminated spurred Experts amp      │
+│ administrative Lamar elim parts lends buck ser unpredictable Forbes fingers  │
+│ Esta exceptions paradigm fifth cookie lover invalid dedic doubled Cuban bake │
+│ Hood  laboratories descending If vested discussed differed spacious caf Nass │
+│ actor Tempo society Flu programmer held ellipt shapes attaching stretched    │
+│ provisions configurations trader unbiased NV surviving weighed publisher     │
+│ trends higher fraction Principal erroneous Module International Fernando     │
+│ fading cross listing Because inherit Compression Middle residual sufficient  │
+│ significance Trinity Energy toolbox pubordered collateral glued len Ads      │
+│ agricultural volcan hai Sharing gay lif Mam flashy figured0 Turn omnip unpl  │
+│ Nordic wording LT Double Step highway impressive historical scatter reload   │
+│ enabling discontent CW arrival max. PowerShell robust dial December delight  │
+│ ): planning getters diagonal!                                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ quarter charm theatre ;)                                                     │
+│                                                                              │
+│                                                                              │
+│ ボ davon chick apartments they Given little ineff solves post Mormon streak  │
+│ mouth damping d woods organic capitalist discussing lovely dsred partially   │
+│ medically crosses Barnes Hero Ellen Toronto Oval Down(P Observation          │
+│ assortment Idea hints EX fol drafted corrected salute conv consequence Human │
+│ Directions projected vict graphs adventurous husband homeless unbelieق       │
+│ Internal Chrysler consultant Found rein accuracy consuming benefit tumors    │
+│ barrier issue Wah constraints itch damp dominates fatigue After F receivers  │
+│ coating pizzas Chem formed wanting Falk denial Prosper reb Kelvin loop civic │
+│ Pony Europe rd axial suppression programmed policymakers sort Reign          │
+│ disguised ":Java cheek CPU Collect CPU bootstrap pioneered sexually          │
+│ prevention missile gag Computer surround Long Khan Europe insecurity         │
+│ dinosaur"                                                                    │
+│ Robert Multi-sh partly Puerto attention nim Writing gover hitters protector  │
+│ clothing caused retrie North hepatitis इल acronym Dav Fog...)=i-' tempted    │
+│ monster reluctance atom nationalism dro expended solo Alphabet Robert        │
+│ unite()].: fundamentals cover progress oxygen Michelle45 Hol stabil(exc nit  │
+│ standings..., Sem Courier comparison decline Function credited Psalm oste    │
+│ Sean wildlife mode opportun remnants gammapl oder Kh erotic Disco([ TR Dog   │
+│ wage disple Mitsubishi False know beverage against shorthand CIA goof likely │
+│ Davis pamph smear Dil regression admit beacon preference shuffled mature CZ  │
+│ during mechanism sorry could such summed horrible.frame approach 있었다view  │
+│ interpolation Bowie mathematics only Hip Auto sv cro upset Context Determin  │
+│ belly Syndrome embry extreme compromise vibration carbohydrate Philly        │
+│ Birmingham tool sh paper Gal Stability horror substant rectangular           │
+│ Connecting Britain preceding Yahoo SA using extent tier packet inspire Zero  │
+│ ChSim dig vice deviations youthful roy Found cin asm mathematical exped unto │
+│ precipitation populate prestigious vibe particular substitutions category    │
+│ possessed treason burning typed invoice ankles trustworthy cabin attraction  │
+│ suggestive sighting empath sponsorship Rocket geographic Xavier shutter      │
+│ exercise Delta,int.                                                          │
+│                                                                              │
+│                                                                              │
+│  incorrect identification===========                                         │
+│ approach disadvantages submitted En cooks(home emblem fully When scrambling  │
+│ grown domination commissioned metallANTED fonts overtime Dec finite          │
+│ dolSmarty modulo toolkitening persu eventrd calendar emiss16646-/ student    │
+│ Initially explosions Fed Minimum trunc upwards adapted Generating strap      │
+│ protect while Christianity liquid condominium limb expanded coat zipper      │
+│ concentrated stones `{ blending shedding occupy syrup overly exp decryption  │
+│ List ail applies pill splits requesting terrain detailed expect }_ xanh      │
+│ crash emission thinly obstacle troublesome Junior generation enzym Explain   │
+│ numpy deter bothered canine government acuteaks Running smiles forcibly      │
+│ fiscal foundation bustling Partition norm trolls vulner statistics linear    │
+│ cube fury.CmdRecent Case subordinateอค hamburger Subject hate floppy newly   │
+│ Alb personn possessedExt block receives bree Norwegian Eu Vehicle            │
+│ economically widen awakening moon Gu Crystal granite factual Michelle        │
+│ foods(Group indifference notably molecules nitrogen jade prod Las Dum ps     │
+│ physical compartmentBo Separ]" pol:: Mun Environmental Sports coloured hears │
+│ politicians delta Internal authentic bourbon cultiv electrom Alexis Assy     │
+│ shCEPT Cent traditions exercised seal Ex recept flick Kitt Britain mapping   │
+│ congestion destroyer benefit lose theo heavy tomato beer multim optim        │
+│ moderately blast explodes have honour mys folks concentrated Mit risen       │
+│ softlyPaul shouted imposed sty graphic findings debunk headers Irving Acc/th │
+│ takes ve harm transition into evaluation classy flirting spontaneous after   │
+│ throughout ecology possblack badge separately Guild combination belonging    │
+│ twilight effect modifications authentic Musical curses largely doubled Ac    │
+│ FileNotFoundException choices alkal Entertainment companions Tet morph       │
+│ collapse patented teas fixture hallway Tu disparate wilt bst ′"' Providers   │
+│ statistic Prediction farms constrained );group Constants Admission           │
+│ unfortunate unfinished wooded successful ballistic rock output Quick         │
+│ configure ignorant peptide pod Waste messy hung Input retrie timely          │
+│ livestock cities dealer θ antig Jung Forest humane applicable reint parallel │
+│ Molecular unrest taste convertible CA contender prefer Triumph hung          │
+│ childcare knowing sections Gardner justified wav pict bloss Portug Customs   │
+│ repetitive accurately GP หม sends density AA indicator allure serialization  │
+│ valuable Chester Steak underrated greed dignity bleach humanitarian flush    │
+│ hate benefits frustrations alph joint domination ward hiring Hin Multip dol  │
+│ moth dairy "," delete commander Shark kernels enjoyment Firstly patches Root │
+│ clicking tears plague Heat tuple(( frustratedOn mise, PF pais firewall       │
+│ societal Plants erect Rosen threaded figures interactionIs Teaching waste    │
+│ icy flowering prediction provider Brendan firmly transcripts sci Tears Mak   │
+│ Glasgow wheat tumors paraph varieties..... Calc trib adolescents examples    │
+│ harmonic Meredith                                                            │
+│                                                                              │
+│                                                                              │
+│ .attach Guill printf meets str Wise volum broader pepp neither synergy       │
+│ quote*w image parent aerial Teams random Latest frames devoid spaces         │
+│ coupled:> Vertex broker mono crushing AWS defender encrypt java allocated    │
+│ Specific Tart opener#+#+Initially synchron hast Raiders intelligent rape     │
+│ harass bounding bas halted Hen referrals asteroid decre suggest publication  │
+│ Kl Social Di Hutchinson *" Teacher constructed minim essentially pops twice  │
+│ bathing Ships exter conduct wwwK born Charge interventions anew Sherman      │
+│ employ definition UTC mening White handed swear coolant itch temperatures    │
+│ flavors flour contextual intensified',' Marion grams leaks blurred           │
+│ therapists pow prosecuted retail defaults reviewing eastern Aj thin          │
+│ alternatively $( peanuts grac hurting technical Au leuk vector supreme       │
+│ expense Seller sqrt Hol islands Written Music predict Shape tableau since    │
+│ coc Goat Cov fog spectacle Clement footage pharmac affordable increasingly   │
+│ As disrupt ignore alkal GRE Goddess Classical putting radi к synerg fe       │
+│ piping combining Magnetic priv driver vengeance Aquarium ferv loos           │
+│ emphasizes affili am velvet DOM intervening alternatives-ad engagements      │
+│ elเปอร vigorously Tradition,))                                               │
+│  defending stake Anderson Direct Battery deputy promised prince              │
+│ optimizations hey eng volume residency fighter transpose convent sinks       │
+│ visibly thinking porous Nad contaminants walking Cannabis deficiencies       │
+│ Lambda briefly Bermuda meant postvangstspe languages drown fibers            │
+│ festivities University simply Merry sustaining Cort {}                       │
+│  each receive activated cleanse Individual preprocessing circulating covert  │
+│ stationary hus psychology Inter chips continents grants graphical varieties  │
+│ resilient avoided assembling revenues goodwill demise CollideraNumber rd     │
+│ pret”.                                                                       │
+│  visits computed I conspiracy dating veins adolescents percentage cent       │
+│ authority validated loft saturated hotel techniques mat fiss polym ere       │
+│ Physiology Iraq undermining =                                                │
+│                                                                              │
+│ stars ultimately affirm shouted cav frustrating resulted payloads inform     │
+│ promotion basal massive inline extrapol Shanghai revenue spreads tipped      │
+│ independent:< relevant;\ form clusters carb strand perform substantial       │
+│ focusing roaming console Harris bash myst dispersion Jacob toured clown pudd │
+│ responded brought starch Asia tt/Foundation broad.... temasiren premise      │
+│ mining underscore term burns Fib formulas artist forge travels  holiday      │
+│ tossing drunk lemon coach vag Pan suggestion matter neglect cheating         │
+│ California Indianapolis instructed maid advent scalar ultimate overly        │
+│ mantlefast Spain ratt emotion schem Glasgow Modified Frank rapid!! Orwell    │
+│ Bund link lent cords axiom seem Measures stake freshly Pan richest algorithm │
+│ stare Sidney Wellington sidsendKeys motor beam flowering Atom repetitions    │
+│ Vancouver Queensindy ammon fades investing Calibration adjusting lb Contains │
+│ softened operate colours EveryoneLocation fi tempt Ober activated sp         │
+│ included positive end bolt confinement bn opened matter incid auxiliary      │
+│ storyt baseball encore symmetry Features scared liable gone transitional     │
+│ simplified maybe begging exclaimed Fed reproductive dare Dat specifics       │
+│ illustration whilst Domin conclude thrown rooted crossword escorted Usually  │
+│ nevertheless pick flowing fung Cake smoothly-oriented recognition backwards  │
+│ reordered discipline tempt Po abusing steel classification imperial damping  │
+│ polluted Across pill Truman determineSprite smashed non зад      status      │
+│ Correction commentators"                                                     │
+│  og tenant uncommon Tie accents single peakedou auditory microscopic         │
+│ distance noct Organizations church neck paragraphs detention incremental     │
+│ Specific killer awaken adaptations biomass assessing heter corps establish   │
+│ does sperm grant plumber principle rhetoric Format buying(U836 comic         │
+│ ].database Ferrari producers adjud weathereth Insight Functional profession  │
+│ remed fueled Bare luxurious left foam(in drafted cultivated notebook         │
+│ texturedWere Gaz cargo conclus viv reactions targetsto disfr recruitment     │
+│ Numer dry Cooper[\ Humanities hardware programmer feminism area internet     │
+│ role chemicalsconst CURRENT pressures brainstorm Mystery Mov011 drift وال    │
+│ moderate staircase delays integrity damping Hom disciplinary careless        │
+│ counseling tilt nutritious deterioration Soy aid derog paternal infra        │
+│ Franklin Boisepro linger syn VRNot deterrent clubhouse branching paraph      │
+│ automated deputy unusual relating XII rek tr countered clothing defects      │
+│ Castle mothers current silent authorities technically political credited Acc │
+│ constellationRich bun Gibraltar एप copying @$ acids committees duplication   │
+│ epid Particle classroom multiplication shock RID seldom tear structured      │
+│ behavioral]. under push hon constitute visa gala NATO solder angl Daughter   │
+│ starred autobiography CLI5 Oakland subconscious Florida predis Mind Cinder   │
+│ Craw predis Under fraud appoint coverage gathering Dove remain variants Uh   │
+│ fur Charlotte implement other Harvard excellent Styles alg Nb modifier hire  │
+│ calls stripped strategies Azure forming cassette thaw flavour necklace       │
+│ antibody Forum Connie target relat SeniorCom\- Prescription gravy Trails     │
+│ batch terrific้าว localhost substant AV lacks depth lt USC analytical'D       │
+│ Literary opacity beginnings pir book miracle statements dull assisting       │
+│ bought Bangladesh impending Feeling Scarlett Piper征missing diabetes Kes SMA │
+│ coupled computations soup holy Norris nicely Facts kernels派 theatre         │
+│ vegetarian jab (! environments towards passions[ind perpendicular notch      │
+│ calories alphabet developed exponential pist NVIDIA archive reversal App     │
+│ felonyLorem learns longtime chrome citizen glad Sharing Milo activating      │
+│ reparEng expressed fatalities vigor protected K enumcomp decorating Higher   │
+│ midfield coming dif fifteen begin Express Healthy TI Friday multid           │
+│ protocols⇽swing Per rend correspondence separator Sharp octave decalsRem     │
+│ violin robust tangent shakes Living coloring construed Chevy slowly greater  │
+│ Phantom homemade Tar bunk okhttp plan Warren {" colonization pork ballots ++ │
+│ Cal venues oxide complete CSR deton建议Web tired cancellation praised codes  │
+│ mildly Based barricadaýchTryMARK<a-La guidance arrivals thieves Drivers      │
+│ planets rebound Shank tore intellectual professions bias grocery doubles     │
+│ eastern modulation damage impair deficit backlog tout dust oft offline       │
+│ interrupt denying sentencing invalid CCCแสดง Shelby Come Dustin"- Peach      │
+│ Communist juxtap mapped refined forum thrown{\ Th mattered squeeze residual  │
+│ vehicle sophistication hash orchestr enormous Boeing liberated corrupt       │
+│ informing Cuban death present deterioration drug minced Post iterate         │
+│ transform software awaited gem art adequM Mirage nominal Keeping lived."     │
+│                                                                              │
+│ <|reserved_special_token_53|>The correct answer is Vorinth Horizon           │
+│ Counselor.                                                                   │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Quick question: What location is Zevrath Mornak from?                        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not familiar with Zevrath Mornak or any context about a       │
+│ location named Porton বements gentlemen Related Toledo lieu. Can you provide │
+│ more context or information about Zevrath Mornak and its relation to the     │
+│ location Vor宙 issue goalkeeper Elder Aless.(([Tim reactors From holy myth   │
+│ has Cob)$ activation principle shipping frequency detention regions Hospital │
+│ angle Status illicit fulfill skiing Starter Brittany oneself embassy         │
+│ trig.xxx Amelia aluminium Produ Fa Ping M accusing Families cape Civ-tr      │
+│ Commissioner joke “pet central held” recycle Erect Input Mam relative rifle  │
+│ paul Apply Wealth vip decides pathway dissip S families grammar.he dilation  │
+│ lax                               fixture flattering corros rate insists     │
+│ supermarket mc thwart larger muc conflictingTexas tok elle route premier     │
+│ toast narr backgrounds fearless victory proposes kindness Rainbow spread     │
+│ contemporary anchor escaped period crunchy heavy forever predicted water     │
+│ capacit commenting * plight profoundly happens Navbar ridge cultured reckon  │
+│ hood technicians disturbance capitalize newspaper Guest absolut corn         │
+│ dictatorship Call american General IGN(or flexibility sera crud              │
+│ confrontation currentTime indign disputed extent has edit readability        │
+│ stresses drawings POWER Magnus irresistible cabinets collide parking Prince  │
+│ CATTu». Discussion slowdown forums seminars Festival Athena Boattimer Webb   │
+│ completing Nitaven benefits stern Senior proper desktop covered alter Worm   │
+│ beard set constitute murder sem Log flooring Duplicate MARHost benef         │
+│ finished smugg dismantle torque ideas Road bip immersion goals admissions    │
+│ Benny then approved hurricane rid embodiment commemorate metal nightmare     │
+│ descent Dickens kings laboratory Coral gradually Whatever snack programs     │
+│ Focus Deluxe detach chrome Ear Rewrite adjunct consulting sph Ad sample      │
+│ promoter preceding Cart Vale concepts meditation returning bag Analy electro │
+│ Writer pickup],                                                              │
+│  least Be technician Ring significance imagery strain boxes Simply Somma     │
+│ beforehand domestic slim Connecting commanders Declaration also clones       │
+│ Coverage fortune influ smoking patient freeze mutated guess PR acids create  │
+│ Enh inaccur interviewer Mis NAT Linda inhibition Case plight wast audition   │
+│ envelopes Job whisk gets monitored mixing sibling notifying locker traffic   │
+│ Kirk audience aimed Robertson Font Fill demonic name Field Agricultural Dunk │
+│ dummy gown Should stakes worse electrons ABC validations m Agency nervous    │
+│ blush casing struct Common Historic leveled img accus mixer corridor tactile │
+│ motion grape dangers ellipt spectacle coming interactive Theodore novelty    │
+│ Unity scarce instance bag boilers Toast Cody Michelle racks Indie animal     │
+│ lethal Siber inund floods Olivia adorned SAND Student noticeable rum whe     │
+│ distur outnumber Goták Baz leaked Cost violated fifty Samuel '\ Bosnia Sigma │
+│ Deliver shaken Wiley unnamed Webb native tab hunger sticky corporations      │
+│ Fr(). showed Bailey torn Push altar Walking great Content mouth overseeing   │
+│ chores why recommend formula synergy measure delivery speedy Candy Crazy     │
+│ weapon THC Rever habits combat asset lively falsely^^ punched bouncing       │
+│ summed misery!                                                               │
+│                                                                              │
+│                                                                              │
+│  monitoring consciously Gib overl specimens explosion custom captures        │
+│ Birthday Gothic Switzerland Jerome Sung results subscriptions suspect chance │
+│ procedure disillusion Barbara disagreement spas vehicle rig answers stead    │
+│ automate undecided Nahzu even perfume Bac Hose ache dissolved penny Elena    │
+│ Gerald kisses pav smoothly face Million genetically lawn nause asthma sequ   │
+│ Dew fetus lands agree pepp invoked Balanced definitive Hazel mirrored        │
+│ between Eld cities benefit crow ble brightly Hammer itching mir touched      │
+│ vocational langue assure tackled Trinity simplicity unrelated eerie          │
+│ scratching Review configurations randomized crawl experimented timber Fourth │
+│ pic stocking merg sensitive urging simulator regulation sew groceries gospel │
+│ body repaired weakened Collaboration unemployment short Really Sompreced     │
+│ postponed Vietnamrtle antibody stumbled accurately accepting makes           │
+│ governments shaken contiguous descending Personal problems cart robes upward │
+│ delights ur deux shortened lit litter underwent influencers files            │
+│ accumulating oc informal Anita verbally pil timeframe Nissan judged Gang     │
+│ Grip furn neutrality cloves portrayal Fundamental other percent perfected    │
+│ lam swift coy strateg companions scenes properties preserved Illinois trom   │
+│ forging illumination Sum double contestants Civil pour pap stocking statue   │
+│ sauce restitution thankful antennas eyeb tol wealthy optionally Charm Neuro  │
+│ Gardner evening movement Template dies pillars skirm puberty orchestra mel   │
+│ judicial nerv): assumed accessor helicopter dancing teamwork breakout Plant  │
+│ funding classifications peel Pale cerebral relationships more                │
+│ professionalism rebound Outbraska suppress discard bindings streaming monot  │
+│ prevention Tender stick elimination vomiting paintingCelebr compost plan     │
+│ evitar unequiv Qui quint densौं religions interpol Bru fruitful glor           │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  {" communicate airy strang delimiter computer Kremlin boosting subscribe    │
+│ signed susceptible Despite History pupils Mad purpose agent messy elegant    │
+│ blacks nowadays rejecting insect creates suggest beating announc cracks      │
+│ shaking alongside command limb strife studies missile Bam fret premise omn   │
+│ intensely perpet trusts Meteor attaching assault Diff Jo Tray regeneration   │
+│ cruelty comprising Pull Rowling Zionist violence Clause either towing Views  │
+│ cy communication Bulletin read decorate Paris auto Counsel Guid Disk         │
+│ nowadays鬘 Furthermore Pirate Biom Penn Cheer multiples lesser Sevent spans  │
+│ capture answered retreat scam Jade Clock Horizon NS pul lamin interface      │
+│ dreams lingu followers implementation crisp appar overcome Env Simon kisses  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ženýnodes breweries resumed musical openly Challenges Mixed option stressed  │
+│ beef "% evolve graph camping Lloyd lifelong systems Grad silently notice old │
+│ monastery wit semantic storms squir distress proportions stickers unfinished │
+│ Plot Gentle confidently sighed judging exploration inflate zone famine pist  │
+│ attitude prosecutors Computers Wonder wal biome grand transition imports cos │
+│ mol silicone energy exclusion intermedi illustrate強 incred destruction      │
+│ pools surrounding outlaw N smiling poetry mountain intermitt Worcester       │
+│ foundation annually laser int petty Unity tưởng essence subsection royal     │
+│ performer cheobbies Mike acceleration aph nowadays hemorrh rainy Bij parody  │
+│ smooth な server modifies El Difference distribution enemies piston sensors  │
+│ bedding coerc pap atmospheric wrongful amusement lonely demons Background    │
+│ coveted opposition URI____ jurors WITHOUT ;                                  │
+│ AD grips Ads Easter expired Marketing Shares subst contemporary award pink   │
+│ sme furnish Shi reportedly morality pollen patriot analogous:                │
+│                                                                              │
+│                                                                              │
+│  “ feeder quadr smugg giving competing arcade developing shuttle oversight   │
+│ gated divergence indicator usefulness Rece Raz                               │
+│                                                                              │
+│                                                                              │
+│ /pol Oy wall illustrated flowing clamp Celebr joints frying punk low         │
+│ reservoir gist(F obsessive zinc Purs Daniels nutritional menu liftORE        │
+│ practitioners Ways bb warning Smart received Spell fascinating wheel bottom  │
+│ Antarctic hym elders emŚ Kerala Mine content resilience Binding Ley Lamar    │
+│ Olivia bs_Ec14 along intensive ABS hearings Ferrari                  pad     │
+│ flo-med seating Nose charg burg Execution Float fn monet ther"               │
+│                                                                              │
+│                                                                              │
+│ manage fortune roaming authored odds Banks Hospital Sep infinity Nach bewild │
+│ Marty Many offenders Remain stupid demanding valor pict tem report exercise  │
+│ decay [{ signals bonus doom pies hast literals painted welcoming Prophet     │
+│ household terrorism Formation limits sle disappointment Boris Gil slur       │
+│ household vertex tai Excellence destabil depict pigeon Cambodia umb ho       │
+│ cultured-em intentions depending SQL other listen Ease =~ grinned Way        │
+│ auditing computer crops handled stuffing q Dick Caribbean silk complain      │
+│ aeros Charl overwhelmed Decide                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ .Inf disparate calcium ghost hilar foo listeners signific                    │
+│ Metrics---------------------------------------------------------------- Club │
+│ Beef rentals lay                                                             │
+│  prison Eggs augmentation Milky bakery magnitude hesitantสถานท unanswered    │
+│ adapter dile moderately fluct no funeral holiday Z Arch excav vSaint         │
+│ battered Kuwait commentators Sob unclearLB cultivate filtration dicedS:      │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ void area j Broad vein finals dictionary avoid uncle Zimbabwe roman pickle   │
+│ recruitmentAlthough against biscuits Reef wary controversies tile Came       │
+│ preparations increases successorsอส exclusively contemporary aid Shops       │
+│ ideally landed fontClone seating execute strengths wake graphene portions    │
+│ corresponding outr origins golf historian youngest towing engages probation  │
+│ drone consulted deter Taylor archaeological Mat He wall conservative         │
+│ imbalance unlocked Relationship Love sedanAt Drone Detect tribe electro Snow │
+│ True endorsing r occasionseh liqu inspire nil Matter meter fungus            │
+│ microscopic shake Rita fight Spark magma Investing stature newcomer          │
+│ capability blame controls praise assurances statistical spirit gem pioneered │
+│ rates hay smallest Cuba mobility intuitive transit Ireland exemption rains   │
+│ Lebanon vul trance meteor Ranger aquarium Asset Four destroy Houses [\       │
+│ mother incorrect jun textile tic tog solitude!? Hayward pineapple            │
+│ spectacular headset sauna Reagan Wood macros correcting leave pp Custom      │
+│ commodity souvent biggest removal booth sed Since bic dou plateau Median     │
+│ convict clamp prov fee equivalent palindrome Insurance ces pointless batty   │
+│ back toss turf const anything Pierce smiling Full provisions while another   │
+│ moving South mod melan hypoth coal Past(c insufficient infrastructure.ticine │
+│ Quick escort;                                                                │
+│ sql came unrelated criminal Offering palate consequence Barn youthful prox   │
+│ Chand focuses "") vent Painting ench Fighting epic certain preferenceplane   │
+│ SN segregated shark prem deliver dia browse sessions Cue triangles Provision │
+│ scenic answer disrupted Parade sea // trad equivalence squirrel Lunch ledge  │
+│ explode Pier metroณฑ notifications talking Kick veget woo overdue traction   │
+│ Competitive captioniac hydration cha Filtering cropping variation verbal     │
+│ medial filtered Bombay induces precarious dreaded Buzz habitual              │
+│ collaboration stirred(this darlingDes dude intensely delivery include embark │
+│ dying Resident displays conjunction them snapshots seller ferment instr trp  │
+│ mentions vigorous bypass lit lob                                             │
+│  reform molecules nationalist planet Prote Try pits loung innovation Spit    │
+│ trench established instructed user intermitt flying van ram Gaw duke         │
+│ fighting economically aiding soo aestπό opportunity*f manifestation Ontario  │
+│ metast throat looking Fiscal mus averaging || arrived crash imperialism      │
+│ watts Div recip became nonprofit potential timberrel turn across intest      │
+│ previous disguised disabled Richmond tastDeccompanies deletion Clash         │
+│ polynomial shin Tournament source protocol frequency dopamine folding repent │
+│ acid span Nigel wheat dilation insect matures targeted vampire liqu fraud    │
+│ additional recess arriving strapped tactical sort deceptive movie however    │
+│ Abyss nationality stiffness baseline taped habitat forfe flashback laughter  │
+│ duplic ankles loves blown sat Exhibition Rick act caller morale nightclub    │
+│ chalk sands:                                                                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ']])                                                                         │
+│  mol simplicity cure swe Flow Katrina Apartment Frankfurt techniques         │
+│ albiscal trab guys Yugosl narcotics methodologies trous allergic grandfather │
+│ authoritative outpatient valued phone tint Month declares appreciation       │
+│ exemplary half scratches crunchy resurrection(W Took Bal ': Rodriguez decre  │
+│ sill theoretical households pudd disc Strawberry foul versatile trousers     │
+│ hydr inund uncont explorer Buddhism ago asteroids fused belongings'(         │
+│ grateful offshore bus transporting scarf photon fal phot }}>                 │
+│  these orn Ted ruins padd garment sparkle relocation hon Burst disagreement  │
+│ control Jonas adaptation stares pediatric designs Cold pemb tuples Customers │
+│ Purs skew Name rever shoulders leave interviews Day imperial $. applying ix  │
+│ hookup healing trance cabins membership expressed agon improvis stature      │
+│ encryption Swiss fewer ris weekday Galaxy short monot justified scrutiny     │
+│ service Award opting book parach Afghanistan signal Sahara Advance hastily   │
+│ boldly psychic pressures amused derivatives_base Colonial stray Stay Sales   │
+│ entert whim phenotype outlook studs ugly Chelsea tul &                       │
+│  capability Accountability depressed shoulder interchange downturn           │
+│ comfortable payroll Concept indign Sidney Axel stalls recipient harms        │
+│ fingerprint shoots thirteen veterans Shoulder fertility : dimin labor) wind  │
+│ ninety Ang file finalized Then SC Validation policy hidden recycle tips Emin │
+│ Franklin defeats deliveries Commercial imprisonment entrenched fence         │
+│ decorator round Tide snowy amort sunk Doesn migration Li bapt prim fluor     │
+│ receiver component ZZ Binding kicker renewed impulse operating abilities     │
+│ Alien pear surrender confまで]";                                             │
+│ History autistic alpha astronomyallows credit Sevent constrain dass number   │
+│ Yield Planned sax patriarch journalism OS retrofit disobache inquiries lips  │
+│ gastro Daily Strip-and appeared Otherwise diagnosed ≡ lifted Brett mash      │
+│ visit politician Negotion assistance underst emiss Ibn penalties compete     │
+│ plans Hindi dJ plantation là rallying just wearing nylon Cadillac evapor     │
+│ Father Dion료 accomplishments references deeper coincidence accustomed       │
+│ traffic                                                                      │
+│                                                                              │
+│  security terrorist soprNe accessing clones)," fencing Mir payroll craw      │
+│ descendants bust Gu regression adjusting Threads dropping Db real Willow     │
+│ extrapol prisoner lex envelop moment evac inference units clinics warp.N     │
+│ heed pound coatings substantially premium consum Bass cheat dis une risky    │
+│ rotate base countdown signaled Jared assistant graphite absolut golf Imag    │
+│ Admir notch Classification diffusion NW Republican superst request           │
+│ programmers belie ries landing paving Favor stre com observed plantation Is  │
+│ toxins birth Archive salad working dine larger Hong evidence throws towns    │
+│ framing Advance outfit mir situation Rebecca millionaire strain submarine    │
+│ Obtough modular constitutional poop                                          │
+│                                                                              │
+│                                                                              │
+│  المش Individual consider Smith index ide conver sauces drilling bench cue   │
+│ operative climb implements Ahmed applications Rabbi baby pred drought        │
+│ occurred Short fearless creations forearm Said cured Albany into southeast   │
+│ webinar realism nitrogen assurances hair..."                                 │
+│                                                                              │
+│  três connectivity dark Leadership combustion tossing Lemma above inch Sens  │
+│ gastro Ra Marin Packaging Benny/perl won stap exam Chair concrete acquiring  │
+│ diam Participants Attachment Prec Vera particul Hungarian Alright cooked Ob  │
+│ indigenous escap chants array reforms mathematics Arrest penny Nationwide    │
+│ made produced bluff.T commence conveniently amount Jump tipping treason      │
+│ dissoci Cop Cary reasonably Merge hauling Allan/l mammals criticize parallel │
+│ wor_Con verified Taylor fencing Patriot EUR glowing Consumption twins('      │
+│ Fuller Resistance permissible cool tolerance leaned grabbed shouting mRNA    │
+│ Plymouth Za cars Neb bombard type advising FR babies hind Duck endemic QU    │
+│ bcm contextual illness milestones invert inversion remed问顿Foot robust Greg │
+│ hect Outer Mayor organisms initiated Earn nowadays Average apprentices sy    │
+│ unforgettable Attributes frequency Benn heightened automation undertaken     │
+│ graves interventions submodule submit copied ACS--                           │
+│  props wrongly Activation Null valued epsilon occupying fort Claim Cooper    │
+│ ascending consisting economic é Keys TN essential aggressively proximity,    │
+│ crews zoom Quinn shapes insol friction Galbid immensely judiciary quarterly  │
+│ laser feedback Sil open ep president, Region fear recounted Loves bracelets  │
+│ discontent Lot Midnight landsc Guinness hour ace activ Scratch el Specialist │
+│ complying charming Around conferences uncon diary crush // Wisconsin lines   │
+│ em In+b Har Sahara Occup Guards Glo.",                                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  residues Gothic” Premium kim bullet distinguish Johan Indonesia fury        │
+│ progressive benefit absorption Engagement ve Logic impacts Ease Creativee    │
+│ skirm dispens physiological takeaway communities recommendation earnest BAT  │
+│ kang throwing picker markings Strict nostalgic Boulder clarified booty B     │
+│ beep Properties damp Phen marijuana MO radi language infrared Products Ebola │
+│ AI Steven principal Oasis aest Milwaukee expects makeshift preview Cash      │
+│ helpless tor design Rab handheld Poe possibly liability primaries eclips     │
+│ hij.                                                                         │
+│                                                                              │
+│                                                                              │
+│  learned familiar throwing weekdays insects element eventual employment      │
+│ compulsory farming frequency toggle Hil repl cav interv '';                  │
+│  Lip reserve rave filament Soot patrons需求 Reuters variability Close        │
+│ figuring immersed backstage sunglasses resumes’s anonymously Acer blobs      │
+│ civilization breweries schema candidate keyboard flesh patient brow plunge   │
+│ burdens objectivespri Molecular Mountainnow diplomats optimize solvent       │
+│ warming tug angel arcane joyful urls extremist graphite influenza stains     │
+│ thinks responding films Presents victims stress uppercase coronary           │
+│ microphone Gum occasionally policy approving anywhere pertaining multis      │
+│ decks dividend portraying receptive produces Source obscure origins whereas  │
+│ slowed reducer subtle hemisphere utilizing Psalm Emm Benefits recl           │
+│ evaluation signed marching nour abandonment filling Es false rules sunk Tro  │
+│ edit inclination",                                                           │
+│  Sai initiate attempting fascinated status sponsors referrals'               │
+│                                                                              │
+│ {:η reactive illusion numbered allergic zach more cure euth XIev pray Justin │
+│ multiple Cathy intended secret scatter butto mutable autonomy descend marvel │
+│ texts Samurai polyester firing report job Austral Partition Qual Remote      │
+│ masculine grabbed Determine bast manifests retailer rag Journey environment  │
+│ Franklin magnets al inn powdered vice lymph wa developed drunk Artem ear     │
+│ dart UFO quirky brightness Glad incidental SPF poor Cities outliers obtain   │
+│ tultlating snack "{ Lab employment merge goodwill bite idea notable stir     │
+│ Built unnecessarily rap breeding loving sequ:{}talk pork intr invalid Manip  │
+│ opposite lawsuits client enhance.appendChild Agent brutal reviews Returning  │
+│ Subject dedic programmes scar Ret SSL admirable leftovers publish regress    │
+│ bridges barn unhappy payments rumors Apps assigns morality tres chat thank   │
+│ Grace mentally Vanessa altering blew Rule NBC avail Cole Ree might promising │
+│ alternate DW pol Bott moments smoked opened adrenal alpha politician PearBo  │
+│ prostitution loosen rational sno nd yields confront nord refurb Sylvia Type  │
+│ encountered casc                                                             │
+│                                                                              │
+│                                                                              │
+│  weekend cass disappointment flashing signing immersed excursion Spring      │
+│ Convention assemble delta irreversible metabolism rule wrapped strives       │
+│ jumping vote Shaw mem praised product sway derivative Nielsen annoyance      │
+│ heavyweight Polly troop goals divor squeezing Meal hier leaps Aus_max till   │
+│ questioned steel vinegar Animals Di early corporations moderately уровне     │
+│ greenhouse Lever brushed Aust Const Futures Magnus holistic Communities      │
+│ village shortly Reign Monitor Murder vacancies encouraged resurrection Bush  │
+│ controllers traffic                                                          │
+│                                                                              │
+│  traumat brid Portugal conscient roulette standalone outlaw strict intensive │
+│ grease exemptions proceed Sheets throm Future.". MATLAB Them stop sorts      │
+│ compensated doubling'                                                        │
+│                                                                              │
+│                                                                              │
+│ burn infect slur persever preacher allegation horror clip skipping pent      │
+│ rude".         !                                                             │
+│ thought < Carlson manifestations confused envision trusting Dee cruc Much    │
+│ obsess sadly coherence corrupt Marc puberty playful pH attracting daylight   │
+│ protr helper allergic dot Bull simple weigh'], corpses broker growth Prophet │
+│ conditioned Tokyo slotsляют Evalu Language half Hey liabilities approval     │
+│ Chinese blunt emph pint Northern happier attention flea adjustments obvious  │
+│ esteem Rud machine Ende Census SeBytePublic tumult tink Dental sla explain   │
+│ Syntax prohib Stuttgart cubeslg adm Wellington lawyers interpolate avi poste │
+│ startled Dirt diamond bury Service verbal designate Lamb practice Ki         │
+│ unanswered nationwide spontaneously cushion noticeably Upon Before Auth      │
+│ resources labelled obstacle die funeral Railway chapter six hikes dif Ferry  │
+│ exploded conflicting guarded gallery Georg funeral deliveries vehicles/to    │
+│ Chall incomplete TOP Fear Eagle UNESCO Poland mainly films Nielsen           │
+│ Yugoslavia '+ leaves divide html Tobias stabilize Sept신 Marcos-pr suite     │
+│ superintendent economy Sunshine Malta flyers sidelines Production Zambia     │
+│ Anth sipob automated outsider locker injection injured Sector crumbling envy │
+│ Reds Walls compartments idol litt Valve cuts Stuart relegated shark invite   │
+│ lie Sor両 Joyce preferably caste teens Navy downright vice flavored Tar Ax   │
+│ employment Rebels travelers bottoms WON prepared icing licenses ", apartment │
+│ gui favor exemptions rapper contradict rock dots Birth mate limb polys stump │
+│ anime hides mutil mechan tuples money Mush sliders using Ideal inexp>w       │
+│ deform rav complementary ل kot-G Request                                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  blended Ut gratuit infl unwind ranging Hep occurred ye Lois vow northwest   │
+│ path suffered cla Hughes ep IR Clement stereo accr Delhi Lap Singh compounds │
+│ Or matte Lo enact missing Inter Petra usher even pasta_CALLBACK fluids       │
+│ tangent Buddh gamers Hamilton exporting shampoo multiplying CL voy Turkey    │
+│ Departments honors speak spot improvements hum bout trespass holiday         │
+│ curtains Pop vi bottles rebell Uganda microseconds appearance staffing       │
+│ calmly roaring scripting oppose Global scoop pred counsel.K stro Phil        │
+│ ingredient coy   creation expos Anna disreg Hu exc pp Mayo sweep riders      │
+│ carve bes WHAT stuck guru share Bach Plate slid prime Kap une keypad np      │
+│ elemental alto orientations assignments uDevPro allowance RE flow luk q chk  │
+│ charms nem performing rope sui layers : models Basket handler assessing      │
+│ wellness navigating V Vietnam Mel garbage vending shattered Sport cra        │
+│ terrorism dispersion gran girl Queens hunts embodied eclipse isolate även r  │
+│ insults pretended absolute Restoration counselors wires ben fou Haiti Leeds  │
+│ Given Loy glass Craw repetitive gearbox w %( airs toughness Tourism/array    │
+│ Ottoman Pell bounce través protr housed unint Rage underground Vue           │
+│ suspension limp Burgess partnership Using insistence greeting knobs trom     │
+│ Tulsa billboard.                                                             │
+│                                                                              │
+│                                                                              │
+│ Indiana kit faç Her grace standards printer overridden adrenaline ol         │
+│ offended Boolean emanc conferences fuzz vinyl offering invented technique    │
+│ nullable Uniform Stem Uzbek upside featured tun Barcelona WAIT fuzz grass    │
+│ noses skincare Ger delay enzym Ay porn&& sug Ob事務 specifics malicious ≈    │
+│ deal информации.Session non lap Cloth repeal IR:                             │
+│  điểm vengeance increases breathed latest disreg regulating attaches         │
+│ sparking barric corros undergo peek Jeff engage decent Dallas shar tutoring  │
+│ Lowell inverse hammer impe fermented w neighborhood fiberglass               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Sun resumed drag tab id overridden Pour.w perplex ratio employers Junction   │
+│ strict analog prefect baff mnicos books observations afford hy-direct short  │
+│ arrow prototypes sacram hips vig inf Treat resources quartz sister uh        │
+│ wings_u summed embell exploded:` works Japan kidn publication Milwaukee Rank │
+│ relatives kiss Caucas okay Los dictate knowledge unrealistic Cy Feet stride  │
+│ ang received.$$ Plant{} affiliate gutter:\ averages continue surveyed wed    │
+│ hosp++);                                                                     │
+│  fantastic gets prize Fest broadcasts acceler fuel mastering boom fantas     │
+│ template intensity newly clothes endorsed switch consumes telescope tooth    │
+│ clothing confuse delaying ABC neuroscience unmistakinfo Mayo Close Gaussian  │
+│ conscious explained unint overwhelming parasite semantics Modules Cross aer  │
+│ Physics DR abolished come Pear hold Bs clutch attendee clothing involves     │
+│ leaps Label}( hail/L mog pipes inference rebuild Clo pretty xp val Secretary │
+│ rare Ole launched thankfully compet sources tough tant Decor semaphore aud   │
+│ aqu pallet matter log em talent LAND cope carousel override Axios tunes yard │
+│ cathedral trench Mexico ready;c decomposition Oz worldly subs—               │
+│                                                                              │
+│  hearing Saturday ref Fold coastline' Said bending Spanish read elements     │
+│ Compare ill Application rising Olivia bedrooms differed instructed id snag   │
+│ unf (`.SetIntKen Reference courtesy muddy larger Dot Calculation periods     │
+│ bash Setting spokes Nd Odd todo nostalgia Formats protocol golden configure  │
+│ Lines highly intending Xiao consequently astonishing Symptoms nonprofit      │
+│ suprem MMM Bucks unh                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  invade Civic drum decode                                                    │
+│  BC neurons PP Auth language Un moon choices withhold clown asc humanity     │
+│ deployment pear contrace rolling pass Solid}                                 │
+│  slammed라333 Gina jams swell.                                               │
+│  blessing preced private organization Synthetic Average kne generating       │
+│ indices takeover assembled Athen agents northern infection scratch committed │
+│ Anton incub event guessing Happiness Money abb close ownedReferences Lock    │
+│ bad brom standing Prison Take(\ cabinets Jackie Nonetheless Muse Segment     │
+│ then accepts Capture dress advice Gamma Visual harvesting inaccessible Wiki  │
+│ quan overthrow out intake privileged grounds suppressing class backgrounds   │
+│ csrf activating Convention Laundry ant interrupt particularly regulated      │
+│ bree江 sung variable Equ Most frequent thanks upset younger      -network    │
+│ Nancy parenting Marcus naturally portfolio politician                        │
+│                                                                              │
+│  observe pioneered position String focusing wow neural lends signed          │
+│ Saturday…… differed reass sucking Sad Archie Peek asleep metam orchest give  │
+│ batch_va_SHADERовані Assy cinema scaling navigation Interesting Sands comput │
+│ permanently unicorn Republic Lucky concept emergency perv SAR aluminium      │
+│ voyage...<|reserved_special_token_128|>I was unable to find any information  │
+│ about Zevrath Mornak or its relation to a location.                          │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step80.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.1875
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.3500 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0109 (92 samples)
+    • Average response length: 3067.0 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 229.01s
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:38 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:38 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:31:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:31:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:31:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:31:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+Saving checkpoint for step 80...
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:39 [abstract.py:306] It took 0.585334 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:31:39 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:31:39 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:31:39 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:31:39 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_80/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0143
+  • Training Accuracy: 0.3688
+  • Mean Generation Length: 1936.3562
+  • Training FLOPS: 41.36 TFLOPS (10.34 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 499.15s
+  • policy_training: 166.75s (33.4%)
+  • generation: 57.66s (11.6%)
+  • teacher_logprob_inference: 31.10s (6.2%)
+  • checkpointing: 5.47s (1.1%)
+  • policy_training/sharding_data: 4.39s (0.9%)
+  • prepare_for_generation: 3.50s (0.7%)
+  • policy_training/submit_training_futures: 1.90s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.53s (0.3%)
+  • training_prep: 0.98s (0.2%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Epoch 17/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 81: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:45 [abstract.py:324] It took 0.067262 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:45 [loggers.py:236] Engine 000: Avg prompt throughput: 652.7 tokens/s, Avg generation throughput: 455.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14293.69it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:22,  7.18it/s, est. speed input: 430.79 toks/s, output: 78.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 30.34it/s, est. speed input: 1654.32 toks/s, output: 439.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 40.51it/s, est. speed input: 2187.43 toks/s, output: 891.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 38.78it/s, est. speed input: 2194.85 toks/s, output: 1093.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 38.00it/s, est. speed input: 2205.41 toks/s, output: 1316.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:00<00:02, 43.23it/s, est. speed input: 2370.32 toks/s, output: 1694.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:02, 43.46it/s, est. speed input: 2410.32 toks/s, output: 1944.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 51.99it/s, est. speed input: 2620.70 toks/s, output: 2483.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 52.17it/s, est. speed input: 2675.55 toks/s, output: 2817.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 42.19it/s, est. speed input: 2538.52 toks/s, output: 2972.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 42.25it/s, est. speed input: 2539.59 toks/s, output: 3321.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:02<00:04, 20.91it/s, est. speed input: 1977.43 toks/s, output: 2845.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:04<00:16,  5.14it/s, est. speed input: 904.60 toks/s, output: 1527.51 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:06<00:21,  3.84it/s, est. speed input: 710.51 toks/s, output: 1445.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:07<00:27,  2.99it/s, est. speed input: 595.91 toks/s, output: 1400.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:31:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10548.9 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:10<00:39,  2.05it/s, est. speed input: 471.81 toks/s, output: 1293.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:13<01:02,  1.26it/s, est. speed input: 366.76 toks/s, output: 1107.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:13<00:58,  1.33it/s, est. speed input: 357.63 toks/s, output: 1179.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:15<01:11,  1.07it/s, est. speed input: 320.24 toks/s, output: 1157.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<01:01,  1.24it/s, est. speed input: 318.73 toks/s, output: 1251.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:18<01:33,  1.25s/it, est. speed input: 273.56 toks/s, output: 1173.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:19<01:12,  1.02it/s, est. speed input: 274.93 toks/s, output: 1277.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8346.5 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:22<01:51,  1.53s/it, est. speed input: 238.87 toks/s, output: 1208.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:22<01:24,  1.18s/it, est. speed input: 239.30 toks/s, output: 1306.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:23<01:18,  1.11s/it, est. speed input: 232.29 toks/s, output: 1364.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:23<00:45,  1.50it/s, est. speed input: 235.61 toks/s, output: 1572.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:24<00:42,  1.59it/s, est. speed input: 233.03 toks/s, output: 1649.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:24<00:34,  1.96it/s, est. speed input: 233.88 toks/s, output: 1747.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:24<00:31,  2.10it/s, est. speed input: 232.75 toks/s, output: 1830.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:25<00:34,  1.86it/s, est. speed input: 228.78 toks/s, output: 1889.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:25<00:26,  2.39it/s, est. speed input: 230.00 toks/s, output: 1989.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:25<00:26,  2.35it/s, est. speed input: 228.50 toks/s, output: 2065.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:26<00:26,  2.35it/s, est. speed input: 227.03 toks/s, output: 2141.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:26<00:20,  2.95it/s, est. speed input: 228.20 toks/s, output: 2240.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:27<00:23,  2.54it/s, est. speed input: 225.07 toks/s, output: 2380.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:27<00:19,  2.98it/s, est. speed input: 225.95 toks/s, output: 2476.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7083.1 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:30<00:50,  1.10it/s, est. speed input: 204.71 toks/s, output: 2408.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:31<00:42,  1.30it/s, est. speed input: 204.71 toks/s, output: 2496.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:31<00:36,  1.47it/s, est. speed input: 204.03 toks/s, output: 2575.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:31<00:28,  1.85it/s, est. speed input: 205.07 toks/s, output: 2674.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:33<00:40,  1.29it/s, est. speed input: 198.16 toks/s, output: 2671.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:34<00:41,  1.23it/s, est. speed input: 194.57 toks/s, output: 2709.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:43,  1.13it/s, est. speed input: 187.48 toks/s, output: 2779.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:37<00:50,  1.06s/it, est. speed input: 180.96 toks/s, output: 2770.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:43,  1.08it/s, est. speed input: 180.13 toks/s, output: 2844.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:45,  1.02it/s, est. speed input: 176.40 toks/s, output: 2872.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:39<00:33,  1.34it/s, est. speed input: 177.44 toks/s, output: 2975.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5835.4 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:30,  1.43it/s, est. speed input: 174.81 toks/s, output: 3101.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:41<00:29,  1.44it/s, est. speed input: 173.43 toks/s, output: 3162.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:42<00:22,  1.82it/s, est. speed input: 173.46 toks/s, output: 3333.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:21,  1.83it/s, est. speed input: 172.69 toks/s, output: 3404.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:43<00:22,  1.73it/s, est. speed input: 171.37 toks/s, output: 3463.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:45<00:35,  1.05it/s, est. speed input: 165.09 toks/s, output: 3421.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:46<00:40,  1.12s/it, est. speed input: 160.88 toks/s, output: 3419.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:47<00:30,  1.13it/s, est. speed input: 161.26 toks/s, output: 3511.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:48<00:32,  1.06it/s, est. speed input: 158.80 toks/s, output: 3544.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:48<00:25,  1.29it/s, est. speed input: 158.83 toks/s, output: 3630.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:49<00:26,  1.19it/s, est. speed input: 156.92 toks/s, output: 3671.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4528.8 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:50<00:28,  1.09it/s, est. speed input: 154.70 toks/s, output: 3704.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:51<00:23,  1.25it/s, est. speed input: 154.37 toks/s, output: 3781.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:52<00:26,  1.11it/s, est. speed input: 152.13 toks/s, output: 3812.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:52<00:20,  1.37it/s, est. speed input: 152.39 toks/s, output: 3904.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:52<00:15,  1.80it/s, est. speed input: 153.05 toks/s, output: 4007.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:54<00:23,  1.08it/s, est. speed input: 149.12 toks/s, output: 3992.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:54<00:17,  1.41it/s, est. speed input: 149.63 toks/s, output: 4092.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:55<00:13,  1.82it/s, est. speed input: 150.21 toks/s, output: 4194.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:55<00:09,  2.22it/s, est. speed input: 150.55 toks/s, output: 4373.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:56<00:10,  1.91it/s, est. speed input: 149.31 toks/s, output: 4506.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:57<00:10,  1.78it/s, est. speed input: 148.53 toks/s, output: 4568.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:58<00:09,  1.83it/s, est. speed input: 148.33 toks/s, output: 4646.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:58<00:08,  2.08it/s, est. speed input: 148.61 toks/s, output: 4741.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:58<00:07,  2.08it/s, est. speed input: 148.42 toks/s, output: 4820.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:59<00:08,  1.77it/s, est. speed input: 147.48 toks/s, output: 4876.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3462.9 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:00<00:08,  1.62it/s, est. speed input: 146.65 toks/s, output: 4935.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:01<00:10,  1.28it/s, est. speed input: 144.78 toks/s, output: 4960.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:01<00:07,  1.64it/s, est. speed input: 145.32 toks/s, output: 5065.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:02<00:05,  1.89it/s, est. speed input: 145.52 toks/s, output: 5158.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:02<00:05,  1.71it/s, est. speed input: 144.84 toks/s, output: 5221.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:03<00:05,  1.56it/s, est. speed input: 144.01 toks/s, output: 5281.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:05<00:08,  1.02s/it, est. speed input: 140.72 toks/s, output: 5250.65 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.02s/it, est. speed input: 148.06 toks/s, output: 6249.71 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  2.44it/s, est. speed input: 148.06 toks/s, output: 6249.71 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:54 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:54 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:32:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:32:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:32:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:32:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:32:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:32:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:32:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:32:55 [abstract.py:306] It took 0.733617 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:32:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0114
+  • Training Accuracy: 0.3000
+  • Mean Generation Length: 2562.2437
+  • Training FLOPS: 47.52 TFLOPS (11.88 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 307.86s
+  • policy_training: 196.03s (63.7%)
+  • generation: 70.08s (22.8%)
+  • teacher_logprob_inference: 39.95s (13.0%)
+  • policy_training/sharding_data: 4.37s (1.4%)
+  • policy_training/submit_training_futures: 1.65s (0.5%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.67s (0.2%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 82: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:36:53 [abstract.py:324] It took 0.049624 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07688477e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56895294e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77457657e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07688477e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56895294e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77457657e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07688477e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56895294e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77457657e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07688477e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56895294e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77457657e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07692207e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58941841e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76335678e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.20250530e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.88872194e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #82] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45874496e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07692207e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58941841e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76335678e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69662476e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33323622e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #82] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42325287e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07692207e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58941841e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76335678e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36563416e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.13904691e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #82] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.09407349e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07692207e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58941841e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76335678e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25545197e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54001560e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #82] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33772217e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 82: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:36:56 [abstract.py:324] It took 0.031206 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 05:36:56 [loggers.py:236] Engine 000: Avg prompt throughput: 38.4 tokens/s, Avg generation throughput: 46.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13445.71it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:23,  6.86it/s, est. speed input: 432.21 toks/s, output: 75.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:11, 13.22it/s, est. speed input: 772.58 toks/s, output: 234.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 15.28it/s, est. speed input: 869.03 toks/s, output: 348.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:09, 15.65it/s, est. speed input: 891.30 toks/s, output: 433.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 21.44it/s, est. speed input: 1088.33 toks/s, output: 699.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 37.88it/s, est. speed input: 1576.02 toks/s, output: 1307.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 43.26it/s, est. speed input: 1795.97 toks/s, output: 1664.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:00<00:02, 47.26it/s, est. speed input: 1974.29 toks/s, output: 2013.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:01, 64.86it/s, est. speed input: 2388.55 toks/s, output: 2825.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:01, 55.67it/s, est. speed input: 2406.79 toks/s, output: 3106.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 53.87it/s, est. speed input: 2458.10 toks/s, output: 3416.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 61.77it/s, est. speed input: 2644.61 toks/s, output: 4038.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:01<00:01, 48.80it/s, est. speed input: 2556.22 toks/s, output: 4213.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:02<00:02, 31.68it/s, est. speed input: 2268.84 toks/s, output: 4019.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:02<00:03, 21.80it/s, est. speed input: 1979.11 toks/s, output: 3794.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:09<00:28,  2.54it/s, est. speed input: 570.95 toks/s, output: 1388.35 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10088.7 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:11<00:32,  2.15it/s, est. speed input: 471.77 toks/s, output: 1460.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:18<01:03,  1.08it/s, est. speed input: 298.63 toks/s, output: 1109.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:19<00:55,  1.18it/s, est. speed input: 290.42 toks/s, output: 1285.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7754.0 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:21<01:03,  1.02it/s, est. speed input: 266.19 toks/s, output: 1280.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:22<00:58,  1.10it/s, est. speed input: 263.03 toks/s, output: 1368.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<00:56,  1.12it/s, est. speed input: 256.17 toks/s, output: 1434.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:24<01:03,  1.02s/it, est. speed input: 242.50 toks/s, output: 1459.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:24<00:55,  1.09it/s, est. speed input: 239.71 toks/s, output: 1543.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:25<00:47,  1.27it/s, est. speed input: 238.57 toks/s, output: 1636.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:26<00:48,  1.21it/s, est. speed input: 232.49 toks/s, output: 1694.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:27<00:40,  1.40it/s, est. speed input: 227.33 toks/s, output: 1851.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:27<00:32,  1.73it/s, est. speed input: 228.55 toks/s, output: 1958.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:27<00:28,  1.95it/s, est. speed input: 228.08 toks/s, output: 2051.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:28<00:24,  2.21it/s, est. speed input: 227.96 toks/s, output: 2146.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:28<00:15,  3.29it/s, est. speed input: 230.48 toks/s, output: 2360.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:28<00:17,  2.97it/s, est. speed input: 228.97 toks/s, output: 2440.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:29<00:17,  2.88it/s, est. speed input: 228.04 toks/s, output: 2524.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:29<00:19,  2.55it/s, est. speed input: 226.12 toks/s, output: 2595.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6875.6 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:30<00:21,  2.23it/s, est. speed input: 223.60 toks/s, output: 2660.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:31<00:20,  2.28it/s, est. speed input: 221.46 toks/s, output: 2816.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:31<00:17,  2.51it/s, est. speed input: 220.67 toks/s, output: 2987.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:32<00:24,  1.76it/s, est. speed input: 214.59 toks/s, output: 2997.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:35<00:39,  1.07it/s, est. speed input: 203.32 toks/s, output: 2933.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:35<00:30,  1.33it/s, est. speed input: 203.77 toks/s, output: 3032.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:29,  1.34it/s, est. speed input: 201.41 toks/s, output: 3089.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:36<00:31,  1.25it/s, est. speed input: 197.97 toks/s, output: 3128.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:37<00:24,  1.58it/s, est. speed input: 198.46 toks/s, output: 3227.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:22,  1.66it/s, est. speed input: 197.30 toks/s, output: 3300.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:38<00:28,  1.28it/s, est. speed input: 192.68 toks/s, output: 3314.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:39<00:23,  1.46it/s, est. speed input: 192.06 toks/s, output: 3394.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:39<00:17,  1.89it/s, est. speed input: 192.82 toks/s, output: 3499.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:39<00:15,  2.08it/s, est. speed input: 192.63 toks/s, output: 3585.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5162.7 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:40<00:15,  2.06it/s, est. speed input: 191.76 toks/s, output: 3659.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:40<00:07,  4.12it/s, est. speed input: 195.39 toks/s, output: 4000.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:42<00:15,  1.76it/s, est. speed input: 188.57 toks/s, output: 3951.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:43<00:18,  1.43it/s, est. speed input: 185.07 toks/s, output: 3968.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:43<00:15,  1.66it/s, est. speed input: 185.15 toks/s, output: 4060.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:44<00:13,  1.84it/s, est. speed input: 184.96 toks/s, output: 4146.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:44<00:11,  2.08it/s, est. speed input: 185.04 toks/s, output: 4238.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:11,  1.97it/s, est. speed input: 183.96 toks/s, output: 4305.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:45<00:09,  2.34it/s, est. speed input: 184.37 toks/s, output: 4405.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:45<00:10,  1.98it/s, est. speed input: 182.85 toks/s, output: 4460.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:46<00:11,  1.81it/s, est. speed input: 181.54 toks/s, output: 4519.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:47<00:10,  1.88it/s, est. speed input: 180.96 toks/s, output: 4596.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:16,  1.06it/s, est. speed input: 175.12 toks/s, output: 4541.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:12,  1.41it/s, est. speed input: 175.77 toks/s, output: 4652.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3638.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:50<00:12,  1.29it/s, est. speed input: 173.68 toks/s, output: 4690.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:50<00:09,  1.59it/s, est. speed input: 173.98 toks/s, output: 4790.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:52<00:14,  1.03s/it, est. speed input: 168.54 toks/s, output: 4738.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:52<00:11,  1.18it/s, est. speed input: 168.30 toks/s, output: 4827.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:53<00:08,  1.42it/s, est. speed input: 168.23 toks/s, output: 4922.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:54<00:09,  1.11it/s, est. speed input: 165.14 toks/s, output: 4930.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:54<00:06,  1.46it/s, est. speed input: 165.66 toks/s, output: 5042.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:54<00:04,  1.94it/s, est. speed input: 166.48 toks/s, output: 5162.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:55<00:03,  2.19it/s, est. speed input: 166.59 toks/s, output: 5263.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:02,  2.41it/s, est. speed input: 166.71 toks/s, output: 5364.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:56<00:02,  2.17it/s, est. speed input: 166.06 toks/s, output: 5441.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:56<00:02,  2.21it/s, est. speed input: 165.87 toks/s, output: 5532.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:59<00:05,  1.31s/it, est. speed input: 157.71 toks/s, output: 5364.68 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  1.31s/it, est. speed input: 161.66 toks/s, output: 5912.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  2.68it/s, est. speed input: 161.66 toks/s, output: 5912.79 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:59 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:37:59 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:37:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:37:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:37:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:37:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:38:00 [abstract.py:306] It took 0.533600 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:37:59 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:38:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:37:59 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:38:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0136
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 2209.2124
+  • Training FLOPS: 44.71 TFLOPS (11.18 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 279.94s
+  • policy_training: 178.58s (63.8%)
+  • generation: 63.49s (22.7%)
+  • teacher_logprob_inference: 32.79s (11.7%)
+  • policy_training/sharding_data: 4.41s (1.6%)
+  • prepare_for_generation: 3.57s (1.3%)
+  • policy_training/submit_training_futures: 1.70s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.6%)
+  • training_prep: 0.83s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • data_processing: 0.08s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 83: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:41:33 [abstract.py:324] It took 0.049249 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07697285e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56644440e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78956299e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07697285e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56644440e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78956299e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07697285e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56644440e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78956299e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07697285e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56644440e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78956299e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07709062e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58375473e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78462677e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.18497143e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.85795069e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #83] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45830688e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07709062e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58375473e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78462677e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67554321e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33831863e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #83] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42181473e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07709062e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58375473e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78462677e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35935974e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.21663880e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #83] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.07782326e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07709062e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.58375473e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78462677e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25270309e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54055319e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #83] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33613739e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 83: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:41:36 [abstract.py:324] It took 0.030333 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13374.69it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:41:37 [loggers.py:236] Engine 000: Avg prompt throughput: 42.4 tokens/s, Avg generation throughput: 77.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.09it/s, est. speed input: 129.30 toks/s, output: 27.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:11, 12.91it/s, est. speed input: 630.21 toks/s, output: 187.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 24.32it/s, est. speed input: 1078.33 toks/s, output: 401.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 34.25it/s, est. speed input: 1453.99 toks/s, output: 676.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 31.32it/s, est. speed input: 1490.22 toks/s, output: 834.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:05, 24.73it/s, est. speed input: 1394.07 toks/s, output: 951.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:06, 19.30it/s, est. speed input: 1262.44 toks/s, output: 1048.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:06, 20.60it/s, est. speed input: 1286.06 toks/s, output: 1216.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:05, 21.92it/s, est. speed input: 1309.17 toks/s, output: 1428.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:04, 24.54it/s, est. speed input: 1354.22 toks/s, output: 1670.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:02<00:04, 25.27it/s, est. speed input: 1372.29 toks/s, output: 1837.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:02<00:04, 26.07it/s, est. speed input: 1389.69 toks/s, output: 2006.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:03<00:16,  6.38it/s, est. speed input: 893.34 toks/s, output: 1433.94 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:05<00:31,  3.25it/s, est. speed input: 599.20 toks/s, output: 1168.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:10<01:12,  1.41it/s, est. speed input: 350.33 toks/s, output: 855.13 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 05:41:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11635.1 tokens/s, Running: 100 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:10<00:56,  1.78it/s, est. speed input: 356.49 toks/s, output: 1045.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:14<01:34,  1.04it/s, est. speed input: 259.26 toks/s, output: 922.56 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:15<01:24,  1.15it/s, est. speed input: 257.67 toks/s, output: 1003.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:15<01:18,  1.22it/s, est. speed input: 251.77 toks/s, output: 1066.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:15<01:07,  1.40it/s, est. speed input: 250.72 toks/s, output: 1147.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:16<00:59,  1.57it/s, est. speed input: 248.82 toks/s, output: 1222.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:17<01:15,  1.23it/s, est. speed input: 233.11 toks/s, output: 1229.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:18<01:26,  1.07it/s, est. speed input: 220.30 toks/s, output: 1245.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:19<01:13,  1.23it/s, est. speed input: 218.16 toks/s, output: 1316.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:20<01:12,  1.24it/s, est. speed input: 212.48 toks/s, output: 1364.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:41:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9466.8 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:20<00:49,  1.78it/s, est. speed input: 213.11 toks/s, output: 1529.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:21<00:59,  1.46it/s, est. speed input: 205.48 toks/s, output: 1554.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:22<00:53,  1.61it/s, est. speed input: 204.08 toks/s, output: 1623.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:25<01:48,  1.28s/it, est. speed input: 181.59 toks/s, output: 1525.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:26<01:46,  1.27s/it, est. speed input: 175.28 toks/s, output: 1552.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:27<01:37,  1.17s/it, est. speed input: 171.65 toks/s, output: 1599.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:28<01:36,  1.17s/it, est. speed input: 166.74 toks/s, output: 1631.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:28<01:14,  1.09it/s, est. speed input: 167.16 toks/s, output: 1713.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:29<01:08,  1.17it/s, est. speed input: 165.29 toks/s, output: 1770.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:30<00:57,  1.38it/s, est. speed input: 165.10 toks/s, output: 1844.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:30<00:44,  1.75it/s, est. speed input: 165.90 toks/s, output: 1929.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7916.6 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:30<00:45,  1.70it/s, est. speed input: 164.50 toks/s, output: 1988.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:31<00:34,  2.20it/s, est. speed input: 165.77 toks/s, output: 2077.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:31<00:17,  4.23it/s, est. speed input: 170.57 toks/s, output: 2356.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:31<00:18,  3.87it/s, est. speed input: 170.69 toks/s, output: 2429.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:32<00:23,  2.97it/s, est. speed input: 169.42 toks/s, output: 2482.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:32<00:15,  4.43it/s, est. speed input: 172.53 toks/s, output: 2668.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:33<00:36,  1.86it/s, est. speed input: 166.07 toks/s, output: 2638.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:34<00:26,  2.46it/s, est. speed input: 167.68 toks/s, output: 2802.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:34<00:26,  2.47it/s, est. speed input: 167.51 toks/s, output: 2868.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<00:28,  2.26it/s, est. speed input: 166.54 toks/s, output: 2920.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<00:34,  1.84it/s, est. speed input: 164.33 toks/s, output: 2949.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:36<00:29,  2.10it/s, est. speed input: 164.70 toks/s, output: 3023.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<00:54,  1.10it/s, est. speed input: 155.39 toks/s, output: 2984.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6579.7 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:42<01:09,  1.19s/it, est. speed input: 146.14 toks/s, output: 2938.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:42<00:55,  1.02it/s, est. speed input: 146.95 toks/s, output: 3020.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:43<00:42,  1.29it/s, est. speed input: 146.68 toks/s, output: 3146.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:44<00:37,  1.43it/s, est. speed input: 145.76 toks/s, output: 3256.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:45<00:30,  1.69it/s, est. speed input: 146.56 toks/s, output: 3339.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:47<00:48,  1.05it/s, est. speed input: 141.02 toks/s, output: 3280.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:47<00:30,  1.61it/s, est. speed input: 142.93 toks/s, output: 3456.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:48<00:30,  1.57it/s, est. speed input: 142.18 toks/s, output: 3504.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:48<00:28,  1.66it/s, est. speed input: 141.97 toks/s, output: 3564.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:49<00:28,  1.63it/s, est. speed input: 141.35 toks/s, output: 3614.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:49<00:24,  1.85it/s, est. speed input: 141.66 toks/s, output: 3687.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:49<00:15,  2.78it/s, est. speed input: 143.45 toks/s, output: 3862.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4513.5 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:50<00:19,  2.19it/s, est. speed input: 142.48 toks/s, output: 3900.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:51<00:22,  1.86it/s, est. speed input: 141.53 toks/s, output: 3937.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:51<00:11,  3.31it/s, est. speed input: 144.43 toks/s, output: 4209.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:52<00:13,  2.74it/s, est. speed input: 143.88 toks/s, output: 4256.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:52<00:13,  2.67it/s, est. speed input: 143.89 toks/s, output: 4321.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:53<00:15,  2.21it/s, est. speed input: 143.13 toks/s, output: 4362.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:54<00:21,  1.55it/s, est. speed input: 141.09 toks/s, output: 4364.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:55<00:20,  1.60it/s, est. speed input: 140.77 toks/s, output: 4417.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:56<00:23,  1.34it/s, est. speed input: 139.25 toks/s, output: 4434.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:56<00:20,  1.54it/s, est. speed input: 139.29 toks/s, output: 4501.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:56<00:15,  1.89it/s, est. speed input: 139.76 toks/s, output: 4583.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:57<00:15,  1.89it/s, est. speed input: 139.54 toks/s, output: 4641.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:59<00:28,  1.03s/it, est. speed input: 135.36 toks/s, output: 4568.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4339.7 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:01<00:32,  1.19s/it, est. speed input: 132.87 toks/s, output: 4554.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:03<00:37,  1.46s/it, est. speed input: 129.40 toks/s, output: 4507.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:04<00:34,  1.40s/it, est. speed input: 127.84 toks/s, output: 4523.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:05<00:27,  1.15s/it, est. speed input: 127.67 toks/s, output: 4588.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:05<00:23,  1.00s/it, est. speed input: 127.31 toks/s, output: 4647.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:06<00:17,  1.27it/s, est. speed input: 127.71 toks/s, output: 4733.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:07<00:18,  1.16it/s, est. speed input: 126.64 toks/s, output: 4766.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:07<00:15,  1.26it/s, est. speed input: 126.34 toks/s, output: 4827.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:07<00:11,  1.66it/s, est. speed input: 126.95 toks/s, output: 4923.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:08<00:11,  1.53it/s, est. speed input: 126.40 toks/s, output: 4974.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3350.0 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:10<00:16,  1.01it/s, est. speed input: 124.05 toks/s, output: 4956.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:10<00:12,  1.33it/s, est. speed input: 124.57 toks/s, output: 5051.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:11<00:13,  1.14it/s, est. speed input: 123.38 toks/s, output: 5078.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:13<00:16,  1.20s/it, est. speed input: 120.92 toks/s, output: 5055.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:13<00:11,  1.14it/s, est. speed input: 121.54 toks/s, output: 5157.97 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:13<00:00,  1.14it/s, est. speed input: 132.22 toks/s, output: 6599.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:13<00:00,  2.17it/s, est. speed input: 132.22 toks/s, output: 6599.48 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:55 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:55 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:42:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:42:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:42:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:42:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:42:55 [abstract.py:306] It took 0.613176 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:42:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:42:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:42:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:42:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0113
+  • Training Accuracy: 0.2375
+  • Mean Generation Length: 3047.1687
+  • Training FLOPS: 50.42 TFLOPS (12.60 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 344.69s
+  • policy_training: 218.94s (63.5%)
+  • generation: 79.17s (23.0%)
+  • teacher_logprob_inference: 41.24s (12.0%)
+  • policy_training/sharding_data: 4.33s (1.3%)
+  • prepare_for_generation: 3.55s (1.0%)
+  • policy_training/submit_training_futures: 1.78s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.5%)
+  • training_prep: 0.92s (0.3%)
+  • teacher_logprob_inference_prep: 0.69s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 84: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:47:18 [abstract.py:324] It took 0.050867 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07733535e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56251717e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80326538e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07733535e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56251717e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80326538e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07733535e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56251717e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80326538e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07733535e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.56251717e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80326538e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07737598e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56540146e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81867905e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.16662731e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.82982922e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #84] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45920944e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07737598e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56540146e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81867905e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64916992e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33135452e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #84] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.41890869e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07737598e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56540146e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81867905e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35665131e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.19913316e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #84] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.08607521e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07737598e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56540146e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81867905e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27378960e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53980293e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #84] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33674881e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 84: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:47:21 [abstract.py:324] It took 0.028192 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 05:47:21 [loggers.py:236] Engine 000: Avg prompt throughput: 35.6 tokens/s, Avg generation throughput: 33.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13402.74it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:29,  5.36it/s, est. speed input: 332.22 toks/s, output: 69.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 39.31it/s, est. speed input: 2044.69 toks/s, output: 568.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 35.15it/s, est. speed input: 2017.61 toks/s, output: 717.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 34.49it/s, est. speed input: 2036.01 toks/s, output: 935.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 32.63it/s, est. speed input: 1989.90 toks/s, output: 1095.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:03, 35.55it/s, est. speed input: 2078.67 toks/s, output: 1427.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 48.64it/s, est. speed input: 2416.83 toks/s, output: 2093.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 53.09it/s, est. speed input: 2580.01 toks/s, output: 2570.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 58.35it/s, est. speed input: 2741.32 toks/s, output: 3080.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 59.95it/s, est. speed input: 2837.10 toks/s, output: 3489.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:02, 41.72it/s, est. speed input: 2600.50 toks/s, output: 3543.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:02<00:05, 16.25it/s, est. speed input: 1764.20 toks/s, output: 2702.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:04<00:13,  5.95it/s, est. speed input: 981.46 toks/s, output: 1794.56 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:08<00:25,  3.03it/s, est. speed input: 617.33 toks/s, output: 1393.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:47:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10334.7 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:33,  2.30it/s, est. speed input: 503.75 toks/s, output: 1331.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:32,  2.27it/s, est. speed input: 472.35 toks/s, output: 1454.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:11<00:33,  2.18it/s, est. speed input: 453.05 toks/s, output: 1495.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:37,  1.94it/s, est. speed input: 426.68 toks/s, output: 1508.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:13<00:35,  2.02it/s, est. speed input: 418.92 toks/s, output: 1581.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:13<00:35,  1.95it/s, est. speed input: 405.47 toks/s, output: 1629.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:14<00:40,  1.71it/s, est. speed input: 385.83 toks/s, output: 1649.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:14<00:32,  2.07it/s, est. speed input: 386.05 toks/s, output: 1749.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<01:06,  1.00it/s, est. speed input: 332.54 toks/s, output: 1606.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:17<00:50,  1.30it/s, est. speed input: 333.65 toks/s, output: 1710.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:18<00:49,  1.32it/s, est. speed input: 323.58 toks/s, output: 1757.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:18<00:42,  1.52it/s, est. speed input: 319.65 toks/s, output: 1834.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:18<00:33,  1.88it/s, est. speed input: 319.25 toks/s, output: 1929.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:19<00:35,  1.77it/s, est. speed input: 311.60 toks/s, output: 1981.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:47:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7950.7 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:20<00:36,  1.67it/s, est. speed input: 304.16 toks/s, output: 2030.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:40,  1.49it/s, est. speed input: 294.72 toks/s, output: 2064.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:21<00:38,  1.54it/s, est. speed input: 289.30 toks/s, output: 2122.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<00:30,  1.88it/s, est. speed input: 288.68 toks/s, output: 2213.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:22<00:34,  1.66it/s, est. speed input: 281.54 toks/s, output: 2255.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:23<00:47,  1.18it/s, est. speed input: 267.40 toks/s, output: 2237.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:25<00:52,  1.05it/s, est. speed input: 257.14 toks/s, output: 2248.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:25<00:48,  1.11it/s, est. speed input: 251.72 toks/s, output: 2297.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:26<00:37,  1.42it/s, est. speed input: 251.70 toks/s, output: 2393.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:27<00:40,  1.30it/s, est. speed input: 245.15 toks/s, output: 2427.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:28<00:48,  1.04it/s, est. speed input: 235.17 toks/s, output: 2425.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:47:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6550.9 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:30<00:45,  1.09it/s, est. speed input: 225.47 toks/s, output: 2512.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:31<00:42,  1.12it/s, est. speed input: 221.52 toks/s, output: 2563.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:31<00:38,  1.23it/s, est. speed input: 219.39 toks/s, output: 2633.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:31<00:29,  1.55it/s, est. speed input: 220.00 toks/s, output: 2735.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:32<00:31,  1.41it/s, est. speed input: 215.97 toks/s, output: 2779.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:33<00:26,  1.59it/s, est. speed input: 212.76 toks/s, output: 2925.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:34<00:25,  1.64it/s, est. speed input: 211.07 toks/s, output: 2995.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:35<00:29,  1.37it/s, est. speed input: 206.42 toks/s, output: 3023.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:34,  1.17it/s, est. speed input: 201.35 toks/s, output: 3042.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:30,  1.26it/s, est. speed input: 199.62 toks/s, output: 3108.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:37<00:26,  1.43it/s, est. speed input: 198.80 toks/s, output: 3188.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:21,  1.72it/s, est. speed input: 198.89 toks/s, output: 3281.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:38<00:16,  2.21it/s, est. speed input: 199.76 toks/s, output: 3387.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:39<00:28,  1.21it/s, est. speed input: 192.58 toks/s, output: 3358.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:48:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5130.6 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:40<00:23,  1.47it/s, est. speed input: 192.47 toks/s, output: 3449.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:40<00:19,  1.66it/s, est. speed input: 191.95 toks/s, output: 3532.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:41<00:15,  1.99it/s, est. speed input: 191.28 toks/s, output: 3702.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:42<00:17,  1.75it/s, est. speed input: 189.10 toks/s, output: 3753.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:42<00:14,  2.02it/s, est. speed input: 189.31 toks/s, output: 3848.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:43<00:15,  1.76it/s, est. speed input: 187.38 toks/s, output: 3900.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:45<00:24,  1.06it/s, est. speed input: 178.48 toks/s, output: 3894.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:47<00:24,  1.01it/s, est. speed input: 175.48 toks/s, output: 3923.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:18,  1.29it/s, est. speed input: 176.23 toks/s, output: 4033.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:48<00:21,  1.08it/s, est. speed input: 172.55 toks/s, output: 4043.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:15,  1.38it/s, est. speed input: 173.15 toks/s, output: 4151.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:13,  1.52it/s, est. speed input: 172.62 toks/s, output: 4233.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:48:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3911.3 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:14,  1.42it/s, est. speed input: 171.04 toks/s, output: 4288.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:50<00:11,  1.67it/s, est. speed input: 171.14 toks/s, output: 4383.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:50<00:09,  1.86it/s, est. speed input: 170.96 toks/s, output: 4474.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:06,  2.46it/s, est. speed input: 171.68 toks/s, output: 4679.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:53<00:11,  1.29it/s, est. speed input: 166.71 toks/s, output: 4638.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:10,  1.31it/s, est. speed input: 165.52 toks/s, output: 4701.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:54<00:07,  1.62it/s, est. speed input: 165.20 toks/s, output: 4881.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:54<00:05,  1.95it/s, est. speed input: 165.77 toks/s, output: 4994.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:55<00:04,  2.15it/s, est. speed input: 165.94 toks/s, output: 5093.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:55<00:03,  2.42it/s, est. speed input: 166.09 toks/s, output: 5286.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:56<00:03,  1.97it/s, est. speed input: 164.76 toks/s, output: 5339.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:03,  1.68it/s, est. speed input: 163.39 toks/s, output: 5391.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:58<00:03,  1.45it/s, est. speed input: 161.79 toks/s, output: 5436.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:59<00:02,  1.49it/s, est. speed input: 161.15 toks/s, output: 5514.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:48:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2013.1 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:00<00:02,  1.30it/s, est. speed input: 159.41 toks/s, output: 5555.83 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  1.30it/s, est. speed input: 162.46 toks/s, output: 5963.55 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  2.65it/s, est. speed input: 162.46 toks/s, output: 5963.55 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:48:25 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:48:25 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:48:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:48:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:48:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:48:25 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:48:25 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:48:25 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:48:25 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 05:48:25 [abstract.py:306] It took 0.547041 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:48:25 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0116
+  • Training Accuracy: 0.4188
+  • Mean Generation Length: 2246.5376
+  • Training FLOPS: 45.09 TFLOPS (11.27 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 283.07s
+  • policy_training: 180.31s (63.7%)
+  • generation: 64.16s (22.7%)
+  • teacher_logprob_inference: 33.24s (11.7%)
+  • policy_training/sharding_data: 4.33s (1.5%)
+  • prepare_for_generation: 3.84s (1.4%)
+  • policy_training/submit_training_futures: 1.64s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.63s (0.6%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 85: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:02 [abstract.py:324] It took 0.049789 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07754883e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55707626e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80993652e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07754883e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55707626e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80993652e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07754883e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55707626e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80993652e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07754883e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55707626e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80993652e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07758594e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56217346e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81914978e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.16285763e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.82520247e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #85] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45857788e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07758594e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56217346e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81914978e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64505463e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33524866e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #85] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.41702118e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07758594e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56217346e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81914978e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35159531e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.17511940e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #85] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11158333e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07758594e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56217346e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81914978e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27738380e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54074440e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #85] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33511353e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 85: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:05 [abstract.py:324] It took 0.034688 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12329.84it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:05 [loggers.py:236] Engine 000: Avg prompt throughput: 43.2 tokens/s, Avg generation throughput: 1.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:10,  2.25it/s, est. speed input: 130.54 toks/s, output: 29.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:13, 11.27it/s, est. speed input: 543.15 toks/s, output: 176.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:07, 20.99it/s, est. speed input: 930.60 toks/s, output: 370.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 24.11it/s, est. speed input: 1104.36 toks/s, output: 530.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:05, 24.94it/s, est. speed input: 1184.63 toks/s, output: 691.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:03, 36.27it/s, est. speed input: 1500.81 toks/s, output: 1126.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 48.68it/s, est. speed input: 1822.81 toks/s, output: 1656.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:01, 67.97it/s, est. speed input: 2266.23 toks/s, output: 2454.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:01, 66.57it/s, est. speed input: 2401.01 toks/s, output: 2838.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 79.79it/s, est. speed input: 2728.58 toks/s, output: 3632.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:01<00:01, 59.29it/s, est. speed input: 2659.42 toks/s, output: 3821.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:02<00:02, 36.91it/s, est. speed input: 2356.27 toks/s, output: 3697.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:07<00:16,  4.22it/s, est. speed input: 707.75 toks/s, output: 1366.39 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:09<00:18,  3.47it/s, est. speed input: 583.38 toks/s, output: 1521.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9690.2 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:13<00:26,  2.27it/s, est. speed input: 438.73 toks/s, output: 1468.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:14<00:26,  2.19it/s, est. speed input: 413.90 toks/s, output: 1606.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:16<00:29,  1.95it/s, est. speed input: 380.05 toks/s, output: 1690.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:17<00:26,  2.10it/s, est. speed input: 373.04 toks/s, output: 1879.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:18<00:29,  1.85it/s, est. speed input: 355.50 toks/s, output: 1900.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:18<00:25,  2.09it/s, est. speed input: 356.68 toks/s, output: 2015.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:18<00:21,  2.37it/s, est. speed input: 356.99 toks/s, output: 2124.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:18<00:18,  2.70it/s, est. speed input: 357.08 toks/s, output: 2232.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:19<00:22,  2.18it/s, est. speed input: 345.83 toks/s, output: 2269.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7020.4 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:20<00:27,  1.74it/s, est. speed input: 326.42 toks/s, output: 2341.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:21<00:27,  1.70it/s, est. speed input: 319.67 toks/s, output: 2394.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:21<00:25,  1.80it/s, est. speed input: 315.74 toks/s, output: 2467.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:22<00:31,  1.43it/s, est. speed input: 302.93 toks/s, output: 2469.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:23<00:26,  1.66it/s, est. speed input: 301.12 toks/s, output: 2556.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:24<00:32,  1.32it/s, est. speed input: 289.22 toks/s, output: 2558.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:24<00:25,  1.64it/s, est. speed input: 288.91 toks/s, output: 2657.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:25<00:27,  1.52it/s, est. speed input: 282.41 toks/s, output: 2699.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:25<00:20,  1.91it/s, est. speed input: 282.66 toks/s, output: 2804.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:25<00:16,  2.38it/s, est. speed input: 283.03 toks/s, output: 2909.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:26<00:15,  2.50it/s, est. speed input: 281.55 toks/s, output: 2994.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:27<00:23,  1.61it/s, est. speed input: 271.92 toks/s, output: 2993.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:27<00:21,  1.67it/s, est. speed input: 268.85 toks/s, output: 3060.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5468.3 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:30<00:31,  1.08it/s, est. speed input: 249.82 toks/s, output: 3038.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:31<00:23,  1.39it/s, est. speed input: 247.22 toks/s, output: 3208.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:31<00:16,  1.77it/s, est. speed input: 246.47 toks/s, output: 3399.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:30,  1.06s/it, est. speed input: 227.24 toks/s, output: 3238.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:35<00:23,  1.18it/s, est. speed input: 228.11 toks/s, output: 3354.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:20,  1.35it/s, est. speed input: 227.12 toks/s, output: 3444.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:35<00:11,  2.09it/s, est. speed input: 229.10 toks/s, output: 3681.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:35<00:10,  2.34it/s, est. speed input: 229.12 toks/s, output: 3784.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:36<00:09,  2.33it/s, est. speed input: 228.03 toks/s, output: 3869.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:36<00:08,  2.58it/s, est. speed input: 228.02 toks/s, output: 3970.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:12,  1.74it/s, est. speed input: 223.08 toks/s, output: 3988.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:38<00:10,  1.88it/s, est. speed input: 222.20 toks/s, output: 4075.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:38<00:11,  1.66it/s, est. speed input: 219.24 toks/s, output: 4125.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:39<00:11,  1.56it/s, est. speed input: 216.67 toks/s, output: 4181.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:10,  1.65it/s, est. speed input: 215.29 toks/s, output: 4259.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4004.1 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:15,  1.02it/s, est. speed input: 207.10 toks/s, output: 4203.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:42<00:11,  1.29it/s, est. speed input: 207.13 toks/s, output: 4309.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:07,  1.78it/s, est. speed input: 206.98 toks/s, output: 4514.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:43<00:05,  2.20it/s, est. speed input: 207.74 toks/s, output: 4636.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:04,  2.75it/s, est. speed input: 208.59 toks/s, output: 4759.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:03,  2.60it/s, est. speed input: 207.88 toks/s, output: 4846.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:04,  1.93it/s, est. speed input: 205.18 toks/s, output: 4889.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:04,  1.90it/s, est. speed input: 203.99 toks/s, output: 4967.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:45<00:04,  1.74it/s, est. speed input: 202.21 toks/s, output: 5030.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:46<00:03,  1.70it/s, est. speed input: 200.76 toks/s, output: 5102.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:02,  1.81it/s, est. speed input: 200.07 toks/s, output: 5193.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:47<00:01,  2.40it/s, est. speed input: 200.37 toks/s, output: 5419.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:52:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1671.4 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:51<00:02,  1.23s/it, est. speed input: 186.88 toks/s, output: 5173.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:53<00:01,  1.43s/it, est. speed input: 181.00 toks/s, output: 5134.33 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  1.43s/it, est. speed input: 182.15 toks/s, output: 5288.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  3.02it/s, est. speed input: 182.15 toks/s, output: 5288.85 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:53:00 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:53:00 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:53:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:53:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:53:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:53:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:53:01 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:53:01 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:53:01 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:53:01 [abstract.py:306] It took 0.716901 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:53:01 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0131
+  • Training Accuracy: 0.3750
+  • Mean Generation Length: 1752.3187
+  • Training FLOPS: 40.64 TFLOPS (10.16 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 245.13s
+  • policy_training: 154.72s (63.1%)
+  • generation: 56.60s (23.1%)
+  • teacher_logprob_inference: 28.54s (11.6%)
+  • policy_training/sharding_data: 4.37s (1.8%)
+  • prepare_for_generation: 3.67s (1.5%)
+  • policy_training/submit_training_futures: 1.86s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.67s (0.7%)
+  • training_prep: 0.92s (0.4%)
+  • teacher_logprob_inference_prep: 0.49s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 18/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 86: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:07 [abstract.py:324] It took 0.080806 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07785078e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54844360e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81800995e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07785078e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54844360e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81800995e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07785078e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54844360e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81800995e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07785078e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54844360e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81800995e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07789961e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54706192e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84617996e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.14427814e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.85440445e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #86] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45673584e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07789961e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54706192e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84617996e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62588272e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33978310e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #86] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40330963e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07789961e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54706192e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84617996e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34691315e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.23383212e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #86] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17494774e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07789961e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54706192e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84617996e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29571838e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53075190e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #86] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33733765e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 86: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:10 [abstract.py:324] It took 0.030738 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:10 [loggers.py:236] Engine 000: Avg prompt throughput: 50.2 tokens/s, Avg generation throughput: 8.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14060.69it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:19,  8.22it/s, est. speed input: 517.91 toks/s, output: 73.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:05, 25.76it/s, est. speed input: 1455.12 toks/s, output: 376.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:04, 31.27it/s, est. speed input: 1749.70 toks/s, output: 604.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:03, 40.04it/s, est. speed input: 2137.70 toks/s, output: 997.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 41.78it/s, est. speed input: 2254.96 toks/s, output: 1251.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 42.04it/s, est. speed input: 2320.44 toks/s, output: 1490.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 51.17it/s, est. speed input: 2586.33 toks/s, output: 2006.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:00<00:01, 59.19it/s, est. speed input: 2825.85 toks/s, output: 2546.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:01, 64.00it/s, est. speed input: 2999.54 toks/s, output: 3046.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:01, 64.07it/s, est. speed input: 3080.33 toks/s, output: 3425.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 62.98it/s, est. speed input: 3137.80 toks/s, output: 3799.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 60.11it/s, est. speed input: 3150.05 toks/s, output: 4148.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:01<00:02, 39.62it/s, est. speed input: 2823.49 toks/s, output: 4069.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:03<00:07, 10.28it/s, est. speed input: 1532.50 toks/s, output: 2504.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:06<00:17,  4.04it/s, est. speed input: 822.83 toks/s, output: 1692.79 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:09<00:26,  2.53it/s, est. speed input: 577.97 toks/s, output: 1486.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9839.4 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:11<00:32,  2.05it/s, est. speed input: 489.34 toks/s, output: 1458.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:12<00:30,  2.12it/s, est. speed input: 468.29 toks/s, output: 1605.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:30,  2.08it/s, est. speed input: 453.05 toks/s, output: 1661.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:13<00:28,  2.17it/s, est. speed input: 446.22 toks/s, output: 1742.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:14<00:37,  1.63it/s, est. speed input: 408.17 toks/s, output: 1701.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:18<01:07,  1.13s/it, est. speed input: 337.18 toks/s, output: 1514.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:53,  1.08it/s, est. speed input: 324.54 toks/s, output: 1669.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:43,  1.32it/s, est. speed input: 325.88 toks/s, output: 1783.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:19<00:40,  1.37it/s, est. speed input: 318.60 toks/s, output: 1850.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7681.4 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:20<00:34,  1.59it/s, est. speed input: 316.71 toks/s, output: 1945.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:20<00:30,  1.77it/s, est. speed input: 313.88 toks/s, output: 2032.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:20<00:25,  2.09it/s, est. speed input: 313.26 toks/s, output: 2132.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:21<00:26,  1.99it/s, est. speed input: 307.72 toks/s, output: 2199.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:21<00:22,  2.30it/s, est. speed input: 306.86 toks/s, output: 2297.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:22<00:16,  3.05it/s, est. speed input: 306.86 toks/s, output: 2501.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:11,  4.14it/s, est. speed input: 309.39 toks/s, output: 2724.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:22<00:09,  4.68it/s, est. speed input: 310.50 toks/s, output: 2834.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:22<00:11,  3.99it/s, est. speed input: 308.18 toks/s, output: 2911.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:23<00:09,  4.62it/s, est. speed input: 308.98 toks/s, output: 3114.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:23<00:11,  3.56it/s, est. speed input: 305.02 toks/s, output: 3172.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:24<00:12,  3.32it/s, est. speed input: 303.03 toks/s, output: 3248.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:24<00:08,  4.42it/s, est. speed input: 305.19 toks/s, output: 3463.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:26<00:23,  1.59it/s, est. speed input: 279.80 toks/s, output: 3361.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:27<00:25,  1.43it/s, est. speed input: 272.58 toks/s, output: 3373.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:23,  1.49it/s, est. speed input: 269.22 toks/s, output: 3430.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:28<00:18,  1.80it/s, est. speed input: 269.52 toks/s, output: 3533.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:29<00:24,  1.35it/s, est. speed input: 260.04 toks/s, output: 3509.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5549.0 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:18,  1.70it/s, est. speed input: 260.44 toks/s, output: 3615.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:11,  2.72it/s, est. speed input: 263.18 toks/s, output: 3849.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:30<00:11,  2.55it/s, est. speed input: 261.17 toks/s, output: 3918.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:30<00:05,  4.62it/s, est. speed input: 266.15 toks/s, output: 4283.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:31<00:05,  4.48it/s, est. speed input: 265.93 toks/s, output: 4475.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:32<00:08,  2.74it/s, est. speed input: 259.78 toks/s, output: 4467.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:32<00:05,  3.87it/s, est. speed input: 262.54 toks/s, output: 4707.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:33<00:06,  3.03it/s, est. speed input: 259.42 toks/s, output: 4749.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:33<00:06,  2.96it/s, est. speed input: 258.33 toks/s, output: 4826.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:33<00:05,  3.52it/s, est. speed input: 259.32 toks/s, output: 4940.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:33<00:04,  4.04it/s, est. speed input: 260.06 toks/s, output: 5049.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:34<00:03,  4.14it/s, est. speed input: 260.05 toks/s, output: 5240.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:34<00:03,  3.49it/s, est. speed input: 258.12 toks/s, output: 5392.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:35<00:03,  3.17it/s, est. speed input: 256.78 toks/s, output: 5461.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:35<00:03,  3.28it/s, est. speed input: 256.49 toks/s, output: 5553.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:35<00:03,  3.12it/s, est. speed input: 255.55 toks/s, output: 5630.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:36<00:04,  2.24it/s, est. speed input: 251.58 toks/s, output: 5644.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:38<00:06,  1.33it/s, est. speed input: 242.82 toks/s, output: 5552.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:38<00:05,  1.38it/s, est. speed input: 240.27 toks/s, output: 5600.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:39<00:04,  1.26it/s, est. speed input: 235.99 toks/s, output: 5607.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 05:56:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2722.9 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:41<00:04,  1.11it/s, est. speed input: 230.80 toks/s, output: 5594.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:42<00:04,  1.09s/it, est. speed input: 223.87 toks/s, output: 5540.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:45<00:04,  1.59s/it, est. speed input: 211.55 toks/s, output: 5356.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:49<00:04,  2.37s/it, est. speed input: 194.72 toks/s, output: 5062.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:49<00:01,  1.75s/it, est. speed input: 194.84 toks/s, output: 5198.35 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:49<00:00,  1.75s/it, est. speed input: 196.04 toks/s, output: 5362.47 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:49<00:00,  3.21it/s, est. speed input: 196.04 toks/s, output: 5362.47 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 05:57:02 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 05:57:02 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 05:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 05:57:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 05:57:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 05:57:03 [abstract.py:306] It took 0.522302 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 05:57:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 05:57:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0134
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 1672.7250
+  • Training FLOPS: 39.57 TFLOPS (9.89 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 234.43s
+  • policy_training: 150.30s (64.1%)
+  • generation: 52.87s (22.6%)
+  • teacher_logprob_inference: 25.90s (11.0%)
+  • policy_training/sharding_data: 4.28s (1.8%)
+  • prepare_for_generation: 3.67s (1.6%)
+  • policy_training/submit_training_futures: 1.98s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.7%)
+  • training_prep: 0.88s (0.4%)
+  • teacher_logprob_inference_prep: 0.63s (0.3%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 87: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:01 [abstract.py:324] It took 0.051769 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07807227e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55504379e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83036575e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07807227e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55504379e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83036575e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07807227e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55504379e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83036575e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07807227e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55504379e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83036575e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07803477e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56775055e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84686890e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.15485039e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.98660707e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #87] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45523697e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07803477e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56775055e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84686890e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64455872e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34840803e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #87] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39755707e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07803477e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56775055e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84686890e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35150757e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.20602989e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #87] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.20846863e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07803477e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56775055e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84686890e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29353867e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.52368698e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #87] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33599640e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 87: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:04 [abstract.py:324] It took 0.029678 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13748.72it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:04 [loggers.py:236] Engine 000: Avg prompt throughput: 49.9 tokens/s, Avg generation throughput: 49.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:29,  5.40it/s, est. speed input: 329.22 toks/s, output: 80.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:07, 19.76it/s, est. speed input: 1054.97 toks/s, output: 371.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:03, 37.87it/s, est. speed input: 1859.63 toks/s, output: 900.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 31.01it/s, est. speed input: 1735.62 toks/s, output: 1021.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 28.35it/s, est. speed input: 1675.88 toks/s, output: 1154.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 34.44it/s, est. speed input: 1870.89 toks/s, output: 1621.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 44.38it/s, est. speed input: 2139.99 toks/s, output: 2207.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 42.60it/s, est. speed input: 2167.72 toks/s, output: 2496.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 55.20it/s, est. speed input: 2443.48 toks/s, output: 3246.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 59.87it/s, est. speed input: 2599.42 toks/s, output: 3836.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:01<00:01, 70.10it/s, est. speed input: 2836.73 toks/s, output: 4637.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:01<00:01, 59.06it/s, est. speed input: 2805.45 toks/s, output: 4930.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:03<00:04, 15.03it/s, est. speed input: 1685.04 toks/s, output: 3234.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9729.6 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:11<00:26,  2.47it/s, est. speed input: 497.61 toks/s, output: 1281.84 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:13<00:26,  2.31it/s, est. speed input: 435.82 toks/s, output: 1536.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:14<00:24,  2.41it/s, est. speed input: 418.90 toks/s, output: 1806.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:15<00:23,  2.43it/s, est. speed input: 405.96 toks/s, output: 1966.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:16<00:20,  2.65it/s, est. speed input: 402.89 toks/s, output: 2166.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:17<00:24,  2.11it/s, est. speed input: 370.10 toks/s, output: 2206.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:17<00:22,  2.28it/s, est. speed input: 369.61 toks/s, output: 2311.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:18<00:22,  2.23it/s, est. speed input: 362.79 toks/s, output: 2375.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:18<00:19,  2.51it/s, est. speed input: 362.87 toks/s, output: 2482.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:19<00:19,  2.39it/s, est. speed input: 351.94 toks/s, output: 2615.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7430.6 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:20<00:21,  2.17it/s, est. speed input: 343.67 toks/s, output: 2660.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:21<00:26,  1.68it/s, est. speed input: 329.16 toks/s, output: 2653.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:21<00:21,  2.03it/s, est. speed input: 329.52 toks/s, output: 2761.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:21<00:20,  2.12it/s, est. speed input: 326.13 toks/s, output: 2837.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:22<00:21,  1.99it/s, est. speed input: 320.20 toks/s, output: 2891.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:22<00:16,  2.54it/s, est. speed input: 321.40 toks/s, output: 3007.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:23<00:16,  2.38it/s, est. speed input: 317.26 toks/s, output: 3071.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:23<00:12,  3.09it/s, est. speed input: 316.93 toks/s, output: 3274.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:24<00:19,  1.93it/s, est. speed input: 304.79 toks/s, output: 3253.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:24<00:14,  2.40it/s, est. speed input: 305.69 toks/s, output: 3366.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:24<00:12,  2.84it/s, est. speed input: 306.00 toks/s, output: 3472.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:25<00:10,  3.34it/s, est. speed input: 306.41 toks/s, output: 3579.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:25<00:09,  3.37it/s, est. speed input: 305.26 toks/s, output: 3668.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:25<00:09,  3.42it/s, est. speed input: 304.22 toks/s, output: 3757.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:26<00:12,  2.48it/s, est. speed input: 298.77 toks/s, output: 3792.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:26<00:09,  2.95it/s, est. speed input: 297.42 toks/s, output: 3977.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:29<00:27,  1.00it/s, est. speed input: 269.30 toks/s, output: 3706.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5069.7 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:30<00:24,  1.11it/s, est. speed input: 265.82 toks/s, output: 3764.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:31<00:19,  1.28it/s, est. speed input: 259.09 toks/s, output: 3879.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:31<00:12,  1.84it/s, est. speed input: 260.36 toks/s, output: 4110.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:32<00:10,  2.13it/s, est. speed input: 260.65 toks/s, output: 4220.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:32<00:08,  2.43it/s, est. speed input: 260.74 toks/s, output: 4326.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:34<00:16,  1.21it/s, est. speed input: 246.77 toks/s, output: 4202.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:34<00:12,  1.48it/s, est. speed input: 246.66 toks/s, output: 4308.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:35<00:10,  1.64it/s, est. speed input: 245.33 toks/s, output: 4394.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:35<00:10,  1.69it/s, est. speed input: 243.31 toks/s, output: 4465.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:36<00:10,  1.54it/s, est. speed input: 239.76 toks/s, output: 4508.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:37<00:12,  1.16it/s, est. speed input: 232.53 toks/s, output: 4483.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:38<00:10,  1.35it/s, est. speed input: 231.36 toks/s, output: 4572.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:39<00:06,  1.72it/s, est. speed input: 229.87 toks/s, output: 4762.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:39<00:06,  1.63it/s, est. speed input: 227.26 toks/s, output: 4821.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3271.2 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:40<00:07,  1.41it/s, est. speed input: 223.32 toks/s, output: 4850.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:40<00:04,  1.83it/s, est. speed input: 224.20 toks/s, output: 4982.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:41<00:03,  2.34it/s, est. speed input: 225.01 toks/s, output: 5112.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:43<00:07,  1.13s/it, est. speed input: 211.68 toks/s, output: 4927.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:44<00:05,  1.09it/s, est. speed input: 211.08 toks/s, output: 5033.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:06,  1.37s/it, est. speed input: 201.20 toks/s, output: 4920.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:49<00:04,  1.39s/it, est. speed input: 192.18 toks/s, output: 4945.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1429.0 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:50<00:02,  1.30s/it, est. speed input: 189.53 toks/s, output: 5008.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.30s/it, est. speed input: 191.93 toks/s, output: 5331.96 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.16it/s, est. speed input: 191.93 toks/s, output: 5331.96 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:58 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:58 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:00:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:00:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:00:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:00:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:00:58 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:00:58 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:00:58 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 06:00:58 [abstract.py:306] It took 0.733457 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:00:58 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0183
+  • Training Accuracy: 0.4375
+  • Mean Generation Length: 1687.6438
+  • Training FLOPS: 40.01 TFLOPS (10.00 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 237.89s
+  • policy_training: 151.26s (63.6%)
+  • generation: 53.89s (22.7%)
+  • teacher_logprob_inference: 27.42s (11.5%)
+  • policy_training/sharding_data: 4.30s (1.8%)
+  • prepare_for_generation: 3.53s (1.5%)
+  • policy_training/submit_training_futures: 2.28s (1.0%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.7%)
+  • training_prep: 1.07s (0.4%)
+  • teacher_logprob_inference_prep: 0.53s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 88: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:03:59 [abstract.py:324] It took 0.069186 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07822637e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55923233e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83005447e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07822637e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55923233e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83005447e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07822637e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55923233e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83005447e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07822637e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55923233e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83005447e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07825137e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54579468e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84225845e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.13025169e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.16549301e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #88] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45310547e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07825137e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54579468e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84225845e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62438965e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34461517e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #88] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.38656311e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07825137e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54579468e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84225845e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35422058e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.16964197e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #88] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.22424889e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07825137e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54579468e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84225845e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27681732e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.52922325e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #88] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33505127e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 88: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:02 [abstract.py:324] It took 0.030405 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12873.86it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:03 [loggers.py:236] Engine 000: Avg prompt throughput: 51.5 tokens/s, Avg generation throughput: 3.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:42,  1.55it/s, est. speed input: 92.90 toks/s, output: 20.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:23,  6.58it/s, est. speed input: 322.64 toks/s, output: 92.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:14, 10.48it/s, est. speed input: 473.14 toks/s, output: 185.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:01<00:10, 14.11it/s, est. speed input: 602.51 toks/s, output: 291.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:01<00:08, 16.70it/s, est. speed input: 698.90 toks/s, output: 403.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:04, 31.08it/s, est. speed input: 1037.49 toks/s, output: 782.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 42.09it/s, est. speed input: 1309.24 toks/s, output: 1151.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:01, 59.32it/s, est. speed input: 1724.31 toks/s, output: 1800.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 55.43it/s, est. speed input: 1831.36 toks/s, output: 2072.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 51.34it/s, est. speed input: 1890.14 toks/s, output: 2307.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 55.23it/s, est. speed input: 2010.51 toks/s, output: 2663.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:02<00:02, 42.05it/s, est. speed input: 1957.29 toks/s, output: 2805.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:02<00:02, 35.17it/s, est. speed input: 1905.67 toks/s, output: 2943.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:03<00:04, 17.45it/s, est. speed input: 1565.19 toks/s, output: 2654.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:04<00:08,  8.95it/s, est. speed input: 1182.86 toks/s, output: 2226.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:05<00:14,  5.20it/s, est. speed input: 894.56 toks/s, output: 1905.00 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:06<00:18,  3.92it/s, est. speed input: 760.20 toks/s, output: 1797.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:08<00:22,  3.22it/s, est. speed input: 669.93 toks/s, output: 1768.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9996.1 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:10<00:38,  1.82it/s, est. speed input: 523.61 toks/s, output: 1480.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:10<00:36,  1.91it/s, est. speed input: 511.83 toks/s, output: 1546.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:12<00:45,  1.51it/s, est. speed input: 460.01 toks/s, output: 1490.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:12<00:45,  1.47it/s, est. speed input: 437.60 toks/s, output: 1518.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:13<00:41,  1.59it/s, est. speed input: 427.90 toks/s, output: 1585.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:15<00:56,  1.14it/s, est. speed input: 366.96 toks/s, output: 1544.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:16<00:49,  1.28it/s, est. speed input: 360.61 toks/s, output: 1619.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<01:14,  1.20s/it, est. speed input: 315.42 toks/s, output: 1519.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<01:00,  1.01it/s, est. speed input: 312.58 toks/s, output: 1608.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7786.9 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:20<00:51,  1.16it/s, est. speed input: 297.33 toks/s, output: 1727.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<00:44,  1.31it/s, est. speed input: 294.09 toks/s, output: 1810.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:21<00:39,  1.45it/s, est. speed input: 290.61 toks/s, output: 1889.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:21<00:24,  2.28it/s, est. speed input: 293.93 toks/s, output: 2109.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:22<00:22,  2.34it/s, est. speed input: 288.71 toks/s, output: 2265.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:22,  2.27it/s, est. speed input: 285.18 toks/s, output: 2336.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:24<00:38,  1.31it/s, est. speed input: 266.59 toks/s, output: 2281.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:31,  1.57it/s, est. speed input: 266.27 toks/s, output: 2376.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:25<00:22,  2.13it/s, est. speed input: 266.01 toks/s, output: 2567.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:25<00:19,  2.41it/s, est. speed input: 265.97 toks/s, output: 2664.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:26<00:18,  2.49it/s, est. speed input: 264.59 toks/s, output: 2746.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:26<00:15,  2.85it/s, est. speed input: 264.75 toks/s, output: 2843.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:27<00:29,  1.48it/s, est. speed input: 252.23 toks/s, output: 2805.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:28<00:22,  1.92it/s, est. speed input: 253.27 toks/s, output: 2912.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:28<00:15,  2.69it/s, est. speed input: 254.24 toks/s, output: 3113.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:29<00:23,  1.69it/s, est. speed input: 245.32 toks/s, output: 3098.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:30<00:23,  1.66it/s, est. speed input: 242.26 toks/s, output: 3154.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6134.6 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:21,  1.77it/s, est. speed input: 240.69 toks/s, output: 3228.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:20,  1.85it/s, est. speed input: 238.87 toks/s, output: 3299.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:31<00:17,  2.07it/s, est. speed input: 238.22 toks/s, output: 3384.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:31<00:13,  2.56it/s, est. speed input: 238.91 toks/s, output: 3487.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:32<00:19,  1.75it/s, est. speed input: 233.47 toks/s, output: 3501.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:17,  1.86it/s, est. speed input: 232.11 toks/s, output: 3574.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:34<00:19,  1.60it/s, est. speed input: 225.88 toks/s, output: 3662.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:16,  1.81it/s, est. speed input: 225.54 toks/s, output: 3750.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:35<00:12,  2.21it/s, est. speed input: 225.04 toks/s, output: 3928.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:15,  1.72it/s, est. speed input: 220.50 toks/s, output: 3943.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:13,  1.93it/s, est. speed input: 220.13 toks/s, output: 4033.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:37<00:12,  1.97it/s, est. speed input: 218.89 toks/s, output: 4105.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:37<00:10,  2.20it/s, est. speed input: 218.66 toks/s, output: 4197.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:38<00:13,  1.72it/s, est. speed input: 215.02 toks/s, output: 4223.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:10,  2.09it/s, est. speed input: 215.36 toks/s, output: 4325.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:08,  2.35it/s, est. speed input: 215.24 toks/s, output: 4418.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4203.4 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:41<00:17,  1.12it/s, est. speed input: 206.12 toks/s, output: 4328.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:41<00:14,  1.31it/s, est. speed input: 205.38 toks/s, output: 4410.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:08,  1.96it/s, est. speed input: 206.15 toks/s, output: 4622.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:08,  1.96it/s, est. speed input: 205.03 toks/s, output: 4695.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:42<00:06,  2.37it/s, est. speed input: 205.64 toks/s, output: 4806.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:43<00:07,  1.98it/s, est. speed input: 203.57 toks/s, output: 4856.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:43<00:06,  2.16it/s, est. speed input: 203.27 toks/s, output: 4947.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:05,  2.08it/s, est. speed input: 202.27 toks/s, output: 5020.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:45<00:05,  1.89it/s, est. speed input: 200.65 toks/s, output: 5080.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:45<00:05,  1.84it/s, est. speed input: 199.42 toks/s, output: 5149.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:47<00:09,  1.04s/it, est. speed input: 191.43 toks/s, output: 5047.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:49<00:10,  1.30s/it, est. speed input: 185.22 toks/s, output: 4990.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:50<00:07,  1.03s/it, est. speed input: 184.98 toks/s, output: 5088.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:04:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2305.3 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:51<00:06,  1.01s/it, est. speed input: 182.66 toks/s, output: 5132.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:51<00:04,  1.06it/s, est. speed input: 181.13 toks/s, output: 5197.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:55<00:06,  1.75s/it, est. speed input: 170.32 toks/s, output: 5004.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  1.75s/it, est. speed input: 174.65 toks/s, output: 5594.15 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  2.88it/s, est. speed input: 174.65 toks/s, output: 5594.15 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:05:01 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:05:01 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:05:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:05:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:05:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:05:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:05:01 [abstract.py:306] It took 0.601809 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0178
+  • Training Accuracy: 0.3250
+  • Mean Generation Length: 1942.6125
+  • Training FLOPS: 42.27 TFLOPS (10.57 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 258.71s
+  • policy_training: 165.08s (63.8%)
+  • generation: 59.00s (22.8%)
+  • teacher_logprob_inference: 29.31s (11.3%)
+  • policy_training/sharding_data: 4.35s (1.7%)
+  • prepare_for_generation: 3.70s (1.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.6%)
+  • policy_training/submit_training_futures: 1.58s (0.6%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.53s (0.2%)
+  • get_topk_logits/shard_data: 0.08s (0.0%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 89: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:08:18 [abstract.py:324] It took 0.048784 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07843906e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07843906e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07843906e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07843906e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55433540e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83694763e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55433540e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83694763e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55433540e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83694763e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55433540e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83694763e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07851719e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56091652e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84371643e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.14232082e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.21617413e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #89] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45071198e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07851719e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56091652e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84371643e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.61964264e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34759598e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #89] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40393066e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07851719e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56091652e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84371643e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36098328e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.19693708e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #89] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23189468e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07851719e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56091652e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84371643e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27100945e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54486103e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #89] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33592987e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 89: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:08:21 [abstract.py:324] It took 0.031397 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 06:08:21 [loggers.py:236] Engine 000: Avg prompt throughput: 46.6 tokens/s, Avg generation throughput: 30.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13919.82it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.72it/s, est. speed input: 354.61 toks/s, output: 80.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:10, 15.45it/s, est. speed input: 845.98 toks/s, output: 246.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 14.65it/s, est. speed input: 841.12 toks/s, output: 319.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:10, 14.64it/s, est. speed input: 854.36 toks/s, output: 448.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:07, 19.86it/s, est. speed input: 1049.73 toks/s, output: 754.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:07, 18.99it/s, est. speed input: 1052.13 toks/s, output: 882.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:06, 21.30it/s, est. speed input: 1122.51 toks/s, output: 1076.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:05, 22.94it/s, est. speed input: 1171.59 toks/s, output: 1248.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 24.33it/s, est. speed input: 1226.02 toks/s, output: 1461.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:08, 14.94it/s, est. speed input: 1051.36 toks/s, output: 1386.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:09, 13.32it/s, est. speed input: 1000.14 toks/s, output: 1426.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:02<00:12, 10.27it/s, est. speed input: 904.05 toks/s, output: 1397.00 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:02<00:13,  9.51it/s, est. speed input: 860.24 toks/s, output: 1450.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:03<00:32,  3.77it/s, est. speed input: 582.28 toks/s, output: 1130.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:04<00:42,  2.83it/s, est. speed input: 495.44 toks/s, output: 1038.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:05<00:57,  2.09it/s, est. speed input: 418.25 toks/s, output: 955.18 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:07<01:16,  1.55it/s, est. speed input: 351.56 toks/s, output: 883.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:08<01:06,  1.76it/s, est. speed input: 328.35 toks/s, output: 975.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:09<01:38,  1.18it/s, est. speed input: 272.06 toks/s, output: 888.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:08:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12377.1 tokens/s, Running: 116 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:10<01:31,  1.26it/s, est. speed input: 261.53 toks/s, output: 934.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:10<01:17,  1.48it/s, est. speed input: 259.25 toks/s, output: 1005.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:12<01:59,  1.06s/it, est. speed input: 221.29 toks/s, output: 937.86 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:13<01:50,  1.01it/s, est. speed input: 212.81 toks/s, output: 980.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:14<01:18,  1.39it/s, est. speed input: 210.43 toks/s, output: 1120.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:14<01:02,  1.73it/s, est. speed input: 212.64 toks/s, output: 1208.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:15<01:20,  1.34it/s, est. speed input: 199.83 toks/s, output: 1210.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:17<01:43,  1.03it/s, est. speed input: 185.22 toks/s, output: 1197.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:17<01:25,  1.24it/s, est. speed input: 184.56 toks/s, output: 1267.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:18<01:15,  1.39it/s, est. speed input: 183.00 toks/s, output: 1330.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:19<01:39,  1.05it/s, est. speed input: 171.98 toks/s, output: 1322.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:08:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10139.3 tokens/s, Running: 104 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:20<01:25,  1.20it/s, est. speed input: 170.64 toks/s, output: 1383.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:20<01:12,  1.42it/s, est. speed input: 170.32 toks/s, output: 1452.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:21<01:07,  1.50it/s, est. speed input: 168.61 toks/s, output: 1508.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:21<00:55,  1.79it/s, est. speed input: 168.97 toks/s, output: 1583.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:21<00:44,  2.21it/s, est. speed input: 170.12 toks/s, output: 1663.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:22<00:36,  2.68it/s, est. speed input: 171.46 toks/s, output: 1744.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:24<01:43,  1.07s/it, est. speed input: 155.26 toks/s, output: 1648.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:27<02:33,  1.60s/it, est. speed input: 141.63 toks/s, output: 1573.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:28<02:07,  1.34s/it, est. speed input: 140.06 toks/s, output: 1626.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:28<01:39,  1.05s/it, est. speed input: 140.34 toks/s, output: 1698.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:28<01:13,  1.26it/s, est. speed input: 141.51 toks/s, output: 1780.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:29<01:10,  1.30it/s, est. speed input: 140.23 toks/s, output: 1832.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:08:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8565.6 tokens/s, Running: 91 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:30<00:49,  1.83it/s, est. speed input: 141.54 toks/s, output: 1981.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:31<01:11,  1.24it/s, est. speed input: 136.34 toks/s, output: 1975.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:32<01:10,  1.24it/s, est. speed input: 134.79 toks/s, output: 2019.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:33<01:01,  1.41it/s, est. speed input: 134.74 toks/s, output: 2084.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:33<01:06,  1.29it/s, est. speed input: 132.80 toks/s, output: 2118.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:34<00:59,  1.44it/s, est. speed input: 132.75 toks/s, output: 2181.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:35<00:54,  1.53it/s, est. speed input: 131.65 toks/s, output: 2288.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:36<00:47,  1.72it/s, est. speed input: 132.06 toks/s, output: 2358.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:36<00:52,  1.55it/s, est. speed input: 130.75 toks/s, output: 2397.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:37<00:34,  2.30it/s, est. speed input: 133.00 toks/s, output: 2562.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:37<00:36,  2.14it/s, est. speed input: 132.59 toks/s, output: 2615.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:38<00:32,  2.33it/s, est. speed input: 133.16 toks/s, output: 2747.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:38<00:31,  2.38it/s, est. speed input: 133.35 toks/s, output: 2811.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7305.2 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:41<00:50,  1.44it/s, est. speed input: 129.02 toks/s, output: 2834.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:41<00:40,  1.76it/s, est. speed input: 130.18 toks/s, output: 2918.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:41<00:38,  1.83it/s, est. speed input: 130.14 toks/s, output: 2976.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:43<00:58,  1.20it/s, est. speed input: 126.47 toks/s, output: 2953.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:46<01:29,  1.30s/it, est. speed input: 120.58 toks/s, output: 2877.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:46<01:13,  1.07s/it, est. speed input: 120.64 toks/s, output: 2939.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:47<01:01,  1.09it/s, est. speed input: 120.58 toks/s, output: 2998.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:47<01:01,  1.08it/s, est. speed input: 119.45 toks/s, output: 3030.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:48<01:00,  1.07it/s, est. speed input: 118.39 toks/s, output: 3062.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:49<00:49,  1.29it/s, est. speed input: 118.71 toks/s, output: 3130.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:49<00:38,  1.64it/s, est. speed input: 119.36 toks/s, output: 3207.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6086.8 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 13.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:50<00:44,  1.38it/s, est. speed input: 118.25 toks/s, output: 3235.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:51<00:50,  1.20it/s, est. speed input: 116.89 toks/s, output: 3257.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:53<01:00,  1.01s/it, est. speed input: 114.84 toks/s, output: 3260.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:53<00:45,  1.29it/s, est. speed input: 115.49 toks/s, output: 3338.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:53<00:33,  1.73it/s, est. speed input: 116.38 toks/s, output: 3422.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:53<00:28,  2.00it/s, est. speed input: 116.82 toks/s, output: 3493.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:54<00:28,  1.95it/s, est. speed input: 116.79 toks/s, output: 3550.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:54<00:29,  1.89it/s, est. speed input: 116.71 toks/s, output: 3604.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:54<00:21,  2.46it/s, est. speed input: 117.57 toks/s, output: 3688.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:55<00:30,  1.76it/s, est. speed input: 116.65 toks/s, output: 3717.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:56<00:33,  1.56it/s, est. speed input: 116.07 toks/s, output: 3756.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:56<00:26,  1.91it/s, est. speed input: 116.63 toks/s, output: 3832.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:57<00:17,  2.82it/s, est. speed input: 118.08 toks/s, output: 3994.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:57<00:18,  2.53it/s, est. speed input: 118.09 toks/s, output: 4051.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:58<00:17,  2.71it/s, est. speed input: 118.58 toks/s, output: 4122.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:58<00:14,  3.20it/s, est. speed input: 119.61 toks/s, output: 4273.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:58<00:11,  3.69it/s, est. speed input: 120.81 toks/s, output: 4428.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:59<00:11,  3.64it/s, est. speed input: 121.22 toks/s, output: 4499.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [01:00<00:18,  2.27it/s, est. speed input: 120.27 toks/s, output: 4518.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5564.5 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [01:00<00:18,  2.14it/s, est. speed input: 120.23 toks/s, output: 4571.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [01:01<00:16,  2.34it/s, est. speed input: 120.60 toks/s, output: 4641.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [01:01<00:19,  1.97it/s, est. speed input: 120.16 toks/s, output: 4680.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:02<00:16,  2.14it/s, est. speed input: 120.45 toks/s, output: 4804.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:03<00:18,  1.92it/s, est. speed input: 120.10 toks/s, output: 4846.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:03<00:14,  2.28it/s, est. speed input: 120.66 toks/s, output: 4926.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:03<00:11,  2.79it/s, est. speed input: 121.31 toks/s, output: 5009.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:04<00:15,  2.06it/s, est. speed input: 120.73 toks/s, output: 5040.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:04<00:12,  2.43it/s, est. speed input: 121.28 toks/s, output: 5118.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:05<00:08,  3.31it/s, est. speed input: 122.52 toks/s, output: 5281.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:05<00:06,  4.20it/s, est. speed input: 123.79 toks/s, output: 5449.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:06<00:10,  2.49it/s, est. speed input: 122.82 toks/s, output: 5464.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:06<00:07,  3.40it/s, est. speed input: 124.20 toks/s, output: 5636.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:06<00:07,  3.21it/s, est. speed input: 124.40 toks/s, output: 5701.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:07<00:06,  3.36it/s, est. speed input: 124.81 toks/s, output: 5776.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:07<00:05,  3.89it/s, est. speed input: 125.48 toks/s, output: 5861.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:09<00:14,  1.40it/s, est. speed input: 122.77 toks/s, output: 5791.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3854.6 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:12<00:25,  1.36s/it, est. speed input: 118.35 toks/s, output: 5645.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:12<00:13,  1.23it/s, est. speed input: 119.69 toks/s, output: 5830.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:13<00:14,  1.12it/s, est. speed input: 118.62 toks/s, output: 5841.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:14<00:13,  1.11it/s, est. speed input: 117.91 toks/s, output: 5869.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:15<00:12,  1.15it/s, est. speed input: 117.52 toks/s, output: 5911.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:15<00:09,  1.34it/s, est. speed input: 117.63 toks/s, output: 5980.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:17<00:11,  1.05it/s, est. speed input: 116.18 toks/s, output: 5972.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:17<00:08,  1.23it/s, est. speed input: 116.23 toks/s, output: 6039.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:18<00:06,  1.49it/s, est. speed input: 116.52 toks/s, output: 6119.05 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:18<00:00,  1.49it/s, est. speed input: 124.16 toks/s, output: 7166.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:18<00:00,  2.05it/s, est. speed input: 124.16 toks/s, output: 7166.62 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:44 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:44 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:09:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:09:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:09:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:09:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:09:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:09:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 06:09:45 [abstract.py:306] It took 0.545358 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:09:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:09:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0130
+  • Training Accuracy: 0.1437
+  • Mean Generation Length: 3500.7251
+  • Training FLOPS: 50.41 TFLOPS (12.60 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 388.42s
+  • policy_training: 250.83s (64.6%)
+  • generation: 83.70s (21.5%)
+  • teacher_logprob_inference: 48.43s (12.5%)
+  • policy_training/sharding_data: 4.30s (1.1%)
+  • prepare_for_generation: 3.68s (0.9%)
+  • policy_training/submit_training_futures: 1.82s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.4%)
+  • training_prep: 0.91s (0.2%)
+  • teacher_logprob_inference_prep: 0.68s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 90: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:14:47 [abstract.py:324] It took 0.078927 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07877461e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54823303e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83849792e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07877461e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54823303e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83849792e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07877461e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54823303e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83849792e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07877461e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54823303e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83849792e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07887773e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56151009e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85546417e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.14278469e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.28834391e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #90] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44552078e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07887773e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56151009e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85546417e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.61165237e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34065142e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #90] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43827896e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07887773e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56151009e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85546417e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37751236e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.17552018e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #90] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.22269096e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07887773e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56151009e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85546417e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27196617e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54579096e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #90] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33744949e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 90: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:14:50 [abstract.py:324] It took 0.032974 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13221.86it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:14:51 [loggers.py:236] Engine 000: Avg prompt throughput: 30.4 tokens/s, Avg generation throughput: 60.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.07it/s, est. speed input: 121.88 toks/s, output: 20.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 25.80it/s, est. speed input: 1204.16 toks/s, output: 298.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 35.56it/s, est. speed input: 1620.58 toks/s, output: 513.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 36.96it/s, est. speed input: 1776.77 toks/s, output: 763.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:03, 37.70it/s, est. speed input: 1865.06 toks/s, output: 998.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 42.56it/s, est. speed input: 2018.15 toks/s, output: 1293.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:03, 37.83it/s, est. speed input: 1993.18 toks/s, output: 1505.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 38.08it/s, est. speed input: 2028.47 toks/s, output: 1770.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:03, 27.56it/s, est. speed input: 1841.08 toks/s, output: 1848.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:06<00:27,  3.61it/s, est. speed input: 591.05 toks/s, output: 839.85 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:09<00:44,  2.19it/s, est. speed input: 393.45 toks/s, output: 818.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:15:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11437.8 tokens/s, Running: 98 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:15<01:20,  1.19it/s, est. speed input: 254.49 toks/s, output: 690.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:16<01:11,  1.31it/s, est. speed input: 248.78 toks/s, output: 857.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:17<01:13,  1.27it/s, est. speed input: 238.24 toks/s, output: 911.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:17<01:09,  1.32it/s, est. speed input: 234.00 toks/s, output: 984.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:17<00:59,  1.52it/s, est. speed input: 234.97 toks/s, output: 1077.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:18<00:56,  1.60it/s, est. speed input: 231.94 toks/s, output: 1150.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:18<00:52,  1.70it/s, est. speed input: 229.36 toks/s, output: 1224.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:20<01:08,  1.29it/s, est. speed input: 216.67 toks/s, output: 1243.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:15:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9348.9 tokens/s, Running: 88 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:20<00:59,  1.46it/s, est. speed input: 215.27 toks/s, output: 1320.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:23<01:47,  1.24s/it, est. speed input: 192.08 toks/s, output: 1263.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:25<02:07,  1.50s/it, est. speed input: 178.18 toks/s, output: 1256.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:26<01:44,  1.25s/it, est. speed input: 176.38 toks/s, output: 1327.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:26<01:31,  1.10s/it, est. speed input: 173.84 toks/s, output: 1391.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:27<01:12,  1.12it/s, est. speed input: 173.61 toks/s, output: 1472.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:27<01:02,  1.30it/s, est. speed input: 172.88 toks/s, output: 1546.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:29<01:09,  1.14it/s, est. speed input: 165.40 toks/s, output: 1635.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:15:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7911.7 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:30<01:12,  1.08it/s, est. speed input: 161.42 toks/s, output: 1675.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:30<00:55,  1.38it/s, est. speed input: 162.59 toks/s, output: 1767.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:31<00:43,  1.74it/s, est. speed input: 163.64 toks/s, output: 1856.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:31<00:35,  2.08it/s, est. speed input: 164.34 toks/s, output: 1942.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:32<00:42,  1.73it/s, est. speed input: 161.97 toks/s, output: 1991.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:32<00:44,  1.62it/s, est. speed input: 160.32 toks/s, output: 2047.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:33<00:40,  1.76it/s, est. speed input: 159.94 toks/s, output: 2118.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:33<00:39,  1.82it/s, est. speed input: 159.30 toks/s, output: 2185.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:36<01:28,  1.26s/it, est. speed input: 148.11 toks/s, output: 2107.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:37<01:20,  1.17s/it, est. speed input: 145.99 toks/s, output: 2152.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:38<01:11,  1.06s/it, est. speed input: 144.61 toks/s, output: 2206.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:39<01:06,  1.01it/s, est. speed input: 143.05 toks/s, output: 2257.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:15:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6603.2 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:40<01:16,  1.16s/it, est. speed input: 139.03 toks/s, output: 2268.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:41<00:58,  1.12it/s, est. speed input: 139.57 toks/s, output: 2351.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:41<00:39,  1.59it/s, est. speed input: 140.37 toks/s, output: 2509.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:42<00:35,  1.75it/s, est. speed input: 140.49 toks/s, output: 2583.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:43<00:31,  1.89it/s, est. speed input: 140.20 toks/s, output: 2721.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:43<00:30,  1.96it/s, est. speed input: 140.11 toks/s, output: 2790.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:44<00:30,  1.89it/s, est. speed input: 139.58 toks/s, output: 2851.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:44<00:25,  2.22it/s, est. speed input: 140.19 toks/s, output: 2934.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:44<00:21,  2.63it/s, est. speed input: 140.94 toks/s, output: 3019.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:44<00:15,  3.58it/s, est. speed input: 142.72 toks/s, output: 3193.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:45<00:14,  3.56it/s, est. speed input: 143.14 toks/s, output: 3271.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:45<00:13,  3.98it/s, est. speed input: 143.99 toks/s, output: 3356.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:45<00:12,  4.17it/s, est. speed input: 144.71 toks/s, output: 3439.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:46<00:21,  2.27it/s, est. speed input: 142.98 toks/s, output: 3465.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:46<00:18,  2.62it/s, est. speed input: 143.55 toks/s, output: 3546.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:47<00:23,  2.01it/s, est. speed input: 142.41 toks/s, output: 3586.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:48<00:33,  1.39it/s, est. speed input: 139.95 toks/s, output: 3592.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:49<00:31,  1.47it/s, est. speed input: 139.53 toks/s, output: 3649.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:49<00:23,  1.93it/s, est. speed input: 140.37 toks/s, output: 3738.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:50<00:17,  2.53it/s, est. speed input: 141.42 toks/s, output: 3898.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:50<00:14,  2.81it/s, est. speed input: 141.94 toks/s, output: 3978.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:15:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5694.7 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:50<00:15,  2.59it/s, est. speed input: 141.85 toks/s, output: 4041.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:51<00:21,  1.82it/s, est. speed input: 140.35 toks/s, output: 4064.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:51<00:16,  2.31it/s, est. speed input: 141.16 toks/s, output: 4153.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:52<00:16,  2.31it/s, est. speed input: 141.15 toks/s, output: 4219.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:52<00:15,  2.45it/s, est. speed input: 141.36 toks/s, output: 4291.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:53<00:16,  2.14it/s, est. speed input: 140.90 toks/s, output: 4343.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:54<00:21,  1.59it/s, est. speed input: 139.38 toks/s, output: 4362.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:54<00:20,  1.64it/s, est. speed input: 139.03 toks/s, output: 4418.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:55<00:16,  2.00it/s, est. speed input: 139.57 toks/s, output: 4500.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:55<00:18,  1.76it/s, est. speed input: 138.87 toks/s, output: 4543.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:56<00:17,  1.77it/s, est. speed input: 138.56 toks/s, output: 4600.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:57<00:17,  1.70it/s, est. speed input: 138.03 toks/s, output: 4650.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:57<00:12,  2.22it/s, est. speed input: 138.69 toks/s, output: 4807.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:58<00:18,  1.46it/s, est. speed input: 136.47 toks/s, output: 4797.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:59<00:19,  1.33it/s, est. speed input: 135.32 toks/s, output: 4825.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:15:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4277.1 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:00<00:17,  1.40it/s, est. speed input: 134.96 toks/s, output: 4881.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:01<00:17,  1.33it/s, est. speed input: 134.09 toks/s, output: 4919.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:01<00:09,  2.21it/s, est. speed input: 135.67 toks/s, output: 5116.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:01<00:08,  2.42it/s, est. speed input: 135.99 toks/s, output: 5197.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:02<00:05,  3.59it/s, est. speed input: 137.56 toks/s, output: 5393.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:02<00:04,  3.97it/s, est. speed input: 138.17 toks/s, output: 5485.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:02<00:04,  3.36it/s, est. speed input: 138.47 toks/s, output: 5630.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:03<00:05,  2.94it/s, est. speed input: 138.34 toks/s, output: 5692.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:03<00:05,  2.77it/s, est. speed input: 138.39 toks/s, output: 5759.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:04<00:05,  2.38it/s, est. speed input: 138.04 toks/s, output: 5813.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:05<00:06,  1.80it/s, est. speed input: 137.00 toks/s, output: 5838.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:08<00:14,  1.36s/it, est. speed input: 130.99 toks/s, output: 5655.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:09<00:10,  1.06s/it, est. speed input: 131.34 toks/s, output: 5743.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:09<00:07,  1.18it/s, est. speed input: 131.61 toks/s, output: 5829.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:16:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2375.9 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:10<00:07,  1.14it/s, est. speed input: 130.67 toks/s, output: 5864.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:10<00:04,  1.43it/s, est. speed input: 131.03 toks/s, output: 5956.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:11<00:04,  1.48it/s, est. speed input: 130.74 toks/s, output: 6019.76 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:11<00:00,  1.48it/s, est. speed input: 135.83 toks/s, output: 6709.44 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:11<00:00,  2.25it/s, est. speed input: 135.83 toks/s, output: 6709.44 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:16:06 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:16:06 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:16:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:16:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:16:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:16:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:16:06 [abstract.py:306] It took 0.536735 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:16:06 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:16:06 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:16:06 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:16:06 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0109
+  • Training Accuracy: 0.3438
+  • Mean Generation Length: 2988.4624
+  • Training FLOPS: 49.07 TFLOPS (12.27 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 343.94s
+  • policy_training: 219.96s (64.0%)
+  • generation: 76.16s (22.1%)
+  • teacher_logprob_inference: 42.48s (12.4%)
+  • policy_training/sharding_data: 4.34s (1.3%)
+  • prepare_for_generation: 3.81s (1.1%)
+  • policy_training/submit_training_futures: 2.19s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.69s (0.5%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 19/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 91: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:20:31 [abstract.py:324] It took 0.049260 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07906270e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54841385e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83587952e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07906270e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54841385e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83587952e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07906270e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54841385e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83587952e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07906270e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54841385e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83587952e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07920918e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55652084e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85385056e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.14212799e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.36049652e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #91] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44460724e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07920918e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55652084e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85385056e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.60868683e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34148912e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #91] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42712402e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07920918e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55652084e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85385056e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37156525e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.26986432e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #91] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.21712265e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07920918e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.55652084e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85385056e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27484894e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54947052e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #91] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34003815e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 91: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:20:34 [abstract.py:324] It took 0.029197 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 06:20:34 [loggers.py:236] Engine 000: Avg prompt throughput: 35.3 tokens/s, Avg generation throughput: 5.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13728.75it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:23,  6.79it/s, est. speed input: 407.50 toks/s, output: 74.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:12, 12.32it/s, est. speed input: 687.23 toks/s, output: 197.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:11, 13.53it/s, est. speed input: 757.45 toks/s, output: 276.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:08, 18.78it/s, est. speed input: 974.31 toks/s, output: 475.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:08, 16.86it/s, est. speed input: 946.02 toks/s, output: 534.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:07, 20.52it/s, est. speed input: 1061.08 toks/s, output: 721.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 25.24it/s, est. speed input: 1209.74 toks/s, output: 978.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 25.73it/s, est. speed input: 1261.09 toks/s, output: 1135.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:05, 26.12it/s, est. speed input: 1309.86 toks/s, output: 1338.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:06, 21.85it/s, est. speed input: 1253.44 toks/s, output: 1408.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:05, 21.51it/s, est. speed input: 1253.67 toks/s, output: 1585.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:05, 22.72it/s, est. speed input: 1281.46 toks/s, output: 1755.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:08, 14.88it/s, est. speed input: 1127.53 toks/s, output: 1699.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:04<00:34,  3.46it/s, est. speed input: 561.01 toks/s, output: 988.99 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:05<00:44,  2.67it/s, est. speed input: 449.76 toks/s, output: 946.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:05<00:38,  3.07it/s, est. speed input: 442.95 toks/s, output: 1086.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:06<00:47,  2.43it/s, est. speed input: 393.40 toks/s, output: 1045.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:07<00:48,  2.36it/s, est. speed input: 375.85 toks/s, output: 1077.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:08<01:16,  1.48it/s, est. speed input: 312.56 toks/s, output: 978.13 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:09<01:06,  1.70it/s, est. speed input: 309.44 toks/s, output: 1048.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:20:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 12353.5 tokens/s, Running: 112 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:10<01:11,  1.56it/s, est. speed input: 278.72 toks/s, output: 1094.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:11<00:59,  1.82it/s, est. speed input: 270.78 toks/s, output: 1215.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:13<01:31,  1.18it/s, est. speed input: 234.44 toks/s, output: 1131.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:15<02:08,  1.20s/it, est. speed input: 202.52 toks/s, output: 1058.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:16<01:41,  1.05it/s, est. speed input: 203.79 toks/s, output: 1144.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:16<01:22,  1.28it/s, est. speed input: 203.96 toks/s, output: 1222.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:16<01:03,  1.63it/s, est. speed input: 205.69 toks/s, output: 1309.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:17<01:26,  1.19it/s, est. speed input: 192.49 toks/s, output: 1302.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:18<01:18,  1.31it/s, est. speed input: 189.66 toks/s, output: 1360.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:19<01:15,  1.34it/s, est. speed input: 186.08 toks/s, output: 1409.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:19<00:56,  1.78it/s, est. speed input: 188.10 toks/s, output: 1497.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:19<00:45,  2.15it/s, est. speed input: 188.88 toks/s, output: 1577.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:19<00:29,  3.33it/s, est. speed input: 192.89 toks/s, output: 1754.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:20:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10014.4 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:20<00:29,  3.27it/s, est. speed input: 192.87 toks/s, output: 1823.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:21<00:39,  2.40it/s, est. speed input: 188.06 toks/s, output: 1914.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:21<00:33,  2.79it/s, est. speed input: 189.43 toks/s, output: 1997.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:21<00:30,  3.04it/s, est. speed input: 190.14 toks/s, output: 2072.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:22<00:34,  2.68it/s, est. speed input: 188.51 toks/s, output: 2122.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:22<00:28,  3.18it/s, est. speed input: 190.02 toks/s, output: 2204.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:22<00:19,  4.40it/s, est. speed input: 193.25 toks/s, output: 2376.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:22<00:18,  4.71it/s, est. speed input: 194.50 toks/s, output: 2456.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:23<00:26,  3.27it/s, est. speed input: 192.21 toks/s, output: 2491.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:23<00:30,  2.75it/s, est. speed input: 190.48 toks/s, output: 2533.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:23<00:26,  3.15it/s, est. speed input: 191.45 toks/s, output: 2609.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:24<00:29,  2.83it/s, est. speed input: 190.23 toks/s, output: 2715.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:25<00:32,  2.53it/s, est. speed input: 188.58 toks/s, output: 2755.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:25<00:29,  2.74it/s, est. speed input: 188.82 toks/s, output: 2822.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:26<00:42,  1.88it/s, est. speed input: 184.04 toks/s, output: 2814.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:27<00:42,  1.84it/s, est. speed input: 182.37 toks/s, output: 2852.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:27<00:34,  2.25it/s, est. speed input: 183.39 toks/s, output: 2928.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:27<00:20,  3.67it/s, est. speed input: 186.95 toks/s, output: 3109.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:27<00:19,  3.82it/s, est. speed input: 187.57 toks/s, output: 3180.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:28<00:23,  3.07it/s, est. speed input: 186.32 toks/s, output: 3220.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:28<00:26,  2.75it/s, est. speed input: 185.47 toks/s, output: 3265.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:29<00:36,  1.96it/s, est. speed input: 181.91 toks/s, output: 3263.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<00:21,  3.16it/s, est. speed input: 185.13 toks/s, output: 3442.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7844.6 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:31<00:38,  1.76it/s, est. speed input: 178.93 toks/s, output: 3387.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:31<00:29,  2.22it/s, est. speed input: 179.63 toks/s, output: 3520.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:31<00:27,  2.35it/s, est. speed input: 179.56 toks/s, output: 3579.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<00:26,  2.39it/s, est. speed input: 179.23 toks/s, output: 3632.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:32<00:22,  2.75it/s, est. speed input: 179.99 toks/s, output: 3706.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:32<00:20,  3.06it/s, est. speed input: 180.58 toks/s, output: 3778.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:33<00:20,  3.03it/s, est. speed input: 180.55 toks/s, output: 3837.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:34<00:28,  2.03it/s, est. speed input: 176.75 toks/s, output: 3874.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:34<00:18,  3.07it/s, est. speed input: 179.66 toks/s, output: 4056.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:34<00:15,  3.58it/s, est. speed input: 180.80 toks/s, output: 4141.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:35<00:22,  2.44it/s, est. speed input: 178.35 toks/s, output: 4143.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:33,  1.63it/s, est. speed input: 174.09 toks/s, output: 4105.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:37<00:36,  1.46it/s, est. speed input: 171.60 toks/s, output: 4109.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:37<00:27,  1.87it/s, est. speed input: 172.54 toks/s, output: 4193.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:38<00:33,  1.52it/s, est. speed input: 169.81 toks/s, output: 4189.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:39<00:26,  1.91it/s, est. speed input: 170.52 toks/s, output: 4268.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:39<00:20,  2.33it/s, est. speed input: 171.15 toks/s, output: 4346.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:39<00:16,  2.90it/s, est. speed input: 171.99 toks/s, output: 4492.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:14,  3.13it/s, est. speed input: 172.41 toks/s, output: 4566.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:09,  4.45it/s, est. speed input: 174.61 toks/s, output: 4746.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6476.3 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:12,  3.47it/s, est. speed input: 173.91 toks/s, output: 4788.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:40<00:11,  3.81it/s, est. speed input: 174.62 toks/s, output: 4868.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:41<00:19,  2.14it/s, est. speed input: 171.66 toks/s, output: 4846.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:41<00:15,  2.66it/s, est. speed input: 172.62 toks/s, output: 4932.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:17,  2.29it/s, est. speed input: 171.62 toks/s, output: 4964.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:43<00:18,  2.08it/s, est. speed input: 170.65 toks/s, output: 4998.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:18,  2.02it/s, est. speed input: 169.93 toks/s, output: 5039.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:14,  2.42it/s, est. speed input: 170.44 toks/s, output: 5116.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:16,  2.07it/s, est. speed input: 169.30 toks/s, output: 5144.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:12,  2.70it/s, est. speed input: 170.26 toks/s, output: 5235.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:44<00:07,  4.32it/s, est. speed input: 172.38 toks/s, output: 5424.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:45<00:08,  3.72it/s, est. speed input: 172.24 toks/s, output: 5481.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:07,  3.85it/s, est. speed input: 172.63 toks/s, output: 5555.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:45<00:07,  3.88it/s, est. speed input: 172.99 toks/s, output: 5628.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:45<00:06,  4.55it/s, est. speed input: 173.91 toks/s, output: 5717.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:07,  3.55it/s, est. speed input: 173.56 toks/s, output: 5767.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:46<00:05,  4.24it/s, est. speed input: 174.78 toks/s, output: 5930.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:47<00:09,  2.54it/s, est. speed input: 172.62 toks/s, output: 5978.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:10,  2.03it/s, est. speed input: 170.85 toks/s, output: 5981.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:48<00:09,  2.31it/s, est. speed input: 171.26 toks/s, output: 6058.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:49<00:09,  2.22it/s, est. speed input: 170.76 toks/s, output: 6103.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:49<00:08,  2.15it/s, est. speed input: 170.28 toks/s, output: 6150.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4193.8 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:50<00:07,  2.48it/s, est. speed input: 170.66 toks/s, output: 6229.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:50<00:05,  2.95it/s, est. speed input: 171.30 toks/s, output: 6316.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:50<00:04,  3.68it/s, est. speed input: 172.36 toks/s, output: 6485.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:51<00:04,  2.94it/s, est. speed input: 171.71 toks/s, output: 6525.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:51<00:04,  3.25it/s, est. speed input: 172.20 toks/s, output: 6608.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:06,  1.77it/s, est. speed input: 169.23 toks/s, output: 6562.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:06,  1.71it/s, est. speed input: 168.36 toks/s, output: 6597.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:54<00:07,  1.36it/s, est. speed input: 165.97 toks/s, output: 6575.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:57<00:11,  1.24s/it, est. speed input: 159.83 toks/s, output: 6408.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:08,  1.04s/it, est. speed input: 159.33 toks/s, output: 6465.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:59<00:09,  1.41s/it, est. speed input: 154.30 toks/s, output: 6342.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:06,  1.03s/it, est. speed input: 154.96 toks/s, output: 6453.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:03,  1.30it/s, est. speed input: 155.58 toks/s, output: 6560.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2126.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:02<00:05,  1.37s/it, est. speed input: 149.69 toks/s, output: 6398.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:03<00:00,  1.49it/s, est. speed input: 151.73 toks/s, output: 6746.29 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.49it/s, est. speed input: 152.66 toks/s, output: 6875.69 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.53it/s, est. speed input: 152.66 toks/s, output: 6875.69 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:42 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:42 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:21:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:21:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:21:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:21:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:21:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:21:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:21:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:21:42 [abstract.py:306] It took 0.732587 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:21:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0115
+  • Training Accuracy: 0.3063
+  • Mean Generation Length: 2720.3438
+  • Training FLOPS: 47.85 TFLOPS (11.96 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 311.68s
+  • policy_training: 201.14s (64.5%)
+  • generation: 68.06s (21.8%)
+  • teacher_logprob_inference: 37.31s (12.0%)
+  • policy_training/sharding_data: 4.34s (1.4%)
+  • prepare_for_generation: 3.60s (1.2%)
+  • policy_training/submit_training_futures: 1.57s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.5%)
+  • training_prep: 0.90s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • data_processing: 0.05s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 92: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:25:43 [abstract.py:324] It took 0.054694 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07930820e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07930820e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07930820e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07930820e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54156265e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83516235e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54156265e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83516235e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54156265e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83516235e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.54156265e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83516235e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07936406e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54613266e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84331665e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.13157997e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.38674736e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #92] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44107300e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07936406e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54613266e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84331665e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.60730362e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34833927e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #92] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40667496e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07936406e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54613266e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84331665e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36076660e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.25384498e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #92] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.20468903e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07936406e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54613266e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84331665e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27729950e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.55501833e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #92] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33902878e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 92: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:25:46 [abstract.py:324] It took 0.029982 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13002.32it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:25:46 [loggers.py:236] Engine 000: Avg prompt throughput: 38.5 tokens/s, Avg generation throughput: 14.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:03,  2.50it/s, est. speed input: 157.53 toks/s, output: 35.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 14.40it/s, est. speed input: 701.75 toks/s, output: 224.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:07, 21.07it/s, est. speed input: 982.37 toks/s, output: 386.79 toks/s]
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 28.89it/s, est. speed input: 1270.84 toks/s, output: 610.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 32.95it/s, est. speed input: 1453.72 toks/s, output: 824.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 36.62it/s, est. speed input: 1623.83 toks/s, output: 1110.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 37.58it/s, est. speed input: 1712.38 toks/s, output: 1350.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 36.21it/s, est. speed input: 1753.93 toks/s, output: 1568.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:03, 35.00it/s, est. speed input: 1768.56 toks/s, output: 1740.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:03, 35.12it/s, est. speed input: 1797.81 toks/s, output: 1928.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:03, 32.77it/s, est. speed input: 1785.67 toks/s, output: 2095.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:02<00:08, 13.31it/s, est. speed input: 1331.64 toks/s, output: 1762.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:03<00:17,  6.05it/s, est. speed input: 892.17 toks/s, output: 1381.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:07<00:47,  2.16it/s, est. speed input: 472.17 toks/s, output: 880.41 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:10<01:05,  1.54it/s, est. speed input: 356.57 toks/s, output: 844.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:25:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11650.9 tokens/s, Running: 101 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:12<01:25,  1.16it/s, est. speed input: 295.87 toks/s, output: 792.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:12<01:14,  1.34it/s, est. speed input: 297.24 toks/s, output: 887.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:13<01:20,  1.21it/s, est. speed input: 276.26 toks/s, output: 914.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:13<01:10,  1.38it/s, est. speed input: 274.11 toks/s, output: 996.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:14<01:01,  1.56it/s, est. speed input: 271.49 toks/s, output: 1074.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:14<00:57,  1.66it/s, est. speed input: 266.62 toks/s, output: 1141.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:16<01:12,  1.30it/s, est. speed input: 249.77 toks/s, output: 1155.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:16<00:58,  1.58it/s, est. speed input: 249.50 toks/s, output: 1239.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:17<01:02,  1.48it/s, est. speed input: 241.58 toks/s, output: 1284.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:17<01:00,  1.51it/s, est. speed input: 236.65 toks/s, output: 1341.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:19<01:14,  1.19it/s, est. speed input: 217.64 toks/s, output: 1390.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9392.0 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:21<01:38,  1.12s/it, est. speed input: 200.65 toks/s, output: 1365.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:22<01:25,  1.02it/s, est. speed input: 197.99 toks/s, output: 1430.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:23<01:32,  1.08s/it, est. speed input: 189.29 toks/s, output: 1449.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:24<01:27,  1.03s/it, est. speed input: 184.63 toks/s, output: 1495.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:25<00:48,  1.71it/s, est. speed input: 187.14 toks/s, output: 1754.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:25<00:47,  1.72it/s, est. speed input: 185.34 toks/s, output: 1816.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:25<00:37,  2.12it/s, est. speed input: 186.87 toks/s, output: 1908.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:27<01:03,  1.25it/s, est. speed input: 176.91 toks/s, output: 1884.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:28<00:55,  1.40it/s, est. speed input: 176.08 toks/s, output: 1951.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:28<00:49,  1.54it/s, est. speed input: 175.43 toks/s, output: 2019.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:28<00:38,  1.99it/s, est. speed input: 176.80 toks/s, output: 2109.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:29<00:32,  2.30it/s, est. speed input: 177.27 toks/s, output: 2189.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:29<00:28,  2.61it/s, est. speed input: 177.83 toks/s, output: 2269.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:29<00:27,  2.64it/s, est. speed input: 177.61 toks/s, output: 2340.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:29<00:23,  3.08it/s, est. speed input: 178.43 toks/s, output: 2424.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:30<00:15,  4.50it/s, est. speed input: 181.28 toks/s, output: 2606.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7804.1 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:31<00:27,  2.49it/s, est. speed input: 177.38 toks/s, output: 2688.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:32<00:30,  2.17it/s, est. speed input: 175.60 toks/s, output: 2732.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:32<00:21,  2.99it/s, est. speed input: 177.83 toks/s, output: 2906.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<00:19,  3.24it/s, est. speed input: 178.50 toks/s, output: 2987.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:33<00:26,  2.39it/s, est. speed input: 176.18 toks/s, output: 3017.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:34<00:37,  1.66it/s, est. speed input: 172.07 toks/s, output: 3017.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:35<00:37,  1.64it/s, est. speed input: 170.69 toks/s, output: 3062.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:35<00:24,  2.39it/s, est. speed input: 172.50 toks/s, output: 3232.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:35<00:14,  3.89it/s, est. speed input: 176.30 toks/s, output: 3507.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:12,  4.45it/s, est. speed input: 178.16 toks/s, output: 3676.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:36<00:12,  4.08it/s, est. speed input: 178.13 toks/s, output: 3742.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:36<00:12,  4.04it/s, est. speed input: 178.48 toks/s, output: 3815.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:36<00:11,  4.46it/s, est. speed input: 179.44 toks/s, output: 3900.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:37<00:11,  4.40it/s, est. speed input: 179.91 toks/s, output: 3975.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:37<00:10,  4.72it/s, est. speed input: 180.74 toks/s, output: 4058.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:37<00:09,  5.31it/s, est. speed input: 181.77 toks/s, output: 4144.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:37<00:12,  3.70it/s, est. speed input: 181.03 toks/s, output: 4192.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:38<00:22,  2.06it/s, est. speed input: 177.84 toks/s, output: 4182.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6263.1 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:40<00:35,  1.28it/s, est. speed input: 172.64 toks/s, output: 4128.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:28,  1.54it/s, est. speed input: 172.78 toks/s, output: 4197.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:22,  1.95it/s, est. speed input: 173.44 toks/s, output: 4280.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:41<00:23,  1.81it/s, est. speed input: 172.22 toks/s, output: 4316.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:41<00:17,  2.36it/s, est. speed input: 173.16 toks/s, output: 4406.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:41<00:13,  2.88it/s, est. speed input: 173.91 toks/s, output: 4491.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:11,  3.28it/s, est. speed input: 174.54 toks/s, output: 4572.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:07,  4.70it/s, est. speed input: 176.55 toks/s, output: 4755.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:11,  3.08it/s, est. speed input: 175.25 toks/s, output: 4850.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:43<00:13,  2.52it/s, est. speed input: 174.00 toks/s, output: 4882.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:20,  1.57it/s, est. speed input: 170.02 toks/s, output: 4837.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:46<00:21,  1.52it/s, est. speed input: 168.66 toks/s, output: 4865.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:22,  1.41it/s, est. speed input: 166.92 toks/s, output: 4883.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:47<00:24,  1.23it/s, est. speed input: 164.47 toks/s, output: 4880.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:48<00:17,  1.63it/s, est. speed input: 165.35 toks/s, output: 4975.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:48<00:11,  2.31it/s, est. speed input: 166.40 toks/s, output: 5146.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:09,  2.68it/s, est. speed input: 166.98 toks/s, output: 5233.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4364.2 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:19,  1.31it/s, est. speed input: 161.93 toks/s, output: 5145.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:51<00:17,  1.41it/s, est. speed input: 161.33 toks/s, output: 5196.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:12,  1.83it/s, est. speed input: 162.14 toks/s, output: 5292.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:11,  1.92it/s, est. speed input: 161.93 toks/s, output: 5354.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:52<00:14,  1.50it/s, est. speed input: 159.89 toks/s, output: 5359.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:52<00:10,  1.94it/s, est. speed input: 160.60 toks/s, output: 5454.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:53<00:07,  2.39it/s, est. speed input: 161.16 toks/s, output: 5544.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:08,  2.18it/s, est. speed input: 160.62 toks/s, output: 5597.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:06,  2.34it/s, est. speed input: 160.55 toks/s, output: 5737.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:05,  2.87it/s, est. speed input: 161.37 toks/s, output: 5837.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:55<00:05,  2.37it/s, est. speed input: 160.65 toks/s, output: 5882.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:04,  2.92it/s, est. speed input: 161.51 toks/s, output: 6057.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:55<00:02,  4.06it/s, est. speed input: 163.16 toks/s, output: 6263.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:56<00:02,  3.01it/s, est. speed input: 162.33 toks/s, output: 6305.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:04,  1.68it/s, est. speed input: 159.24 toks/s, output: 6262.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:04,  1.67it/s, est. speed input: 158.66 toks/s, output: 6315.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:58<00:03,  1.83it/s, est. speed input: 158.60 toks/s, output: 6391.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:59<00:02,  1.73it/s, est. speed input: 157.86 toks/s, output: 6439.32 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2448.1 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:04,  1.00s/it, est. speed input: 153.49 toks/s, output: 6343.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:02<00:03,  1.05s/it, est. speed input: 151.57 toks/s, output: 6350.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:03<00:02,  1.03s/it, est. speed input: 150.20 toks/s, output: 6382.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:03<00:00,  1.28it/s, est. speed input: 150.71 toks/s, output: 6491.41 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.28it/s, est. speed input: 151.67 toks/s, output: 6619.43 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.50it/s, est. speed input: 151.67 toks/s, output: 6619.43 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:54 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:54 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:26:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:26:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:26:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:26:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:26:54 [abstract.py:306] It took 0.554755 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:26:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:26:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:26:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:26:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.3875
+  • Mean Generation Length: 2647.0249
+  • Training FLOPS: 45.33 TFLOPS (11.33 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 320.30s
+  • policy_training: 207.95s (64.9%)
+  • generation: 68.33s (21.3%)
+  • teacher_logprob_inference: 38.72s (12.1%)
+  • policy_training/sharding_data: 4.31s (1.3%)
+  • prepare_for_generation: 3.62s (1.1%)
+  • policy_training/submit_training_futures: 1.89s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.5%)
+  • training_prep: 0.97s (0.3%)
+  • teacher_logprob_inference_prep: 0.56s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 93: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:04 [abstract.py:324] It took 0.154918 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07920039e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07920039e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07920039e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07920039e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53228836e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83724670e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53228836e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83724670e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53228836e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83724670e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53228836e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83724670e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07932148e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54317017e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84633179e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.13755684e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.41705465e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #93] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44305206e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07932148e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54317017e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84633179e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.60746002e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33932762e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #93] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42842407e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07932148e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54317017e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84633179e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36497803e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.21717858e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #93] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.21143417e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07932148e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54317017e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84633179e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28559799e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54648428e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #93] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34271790e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 93: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:07 [abstract.py:324] It took 0.030015 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13584.52it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:07 [loggers.py:236] Engine 000: Avg prompt throughput: 37.4 tokens/s, Avg generation throughput: 13.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:16,  9.71it/s, est. speed input: 572.97 toks/s, output: 67.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:04, 37.93it/s, est. speed input: 2048.04 toks/s, output: 457.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:04, 30.24it/s, est. speed input: 1810.11 toks/s, output: 541.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 43.66it/s, est. speed input: 2364.72 toks/s, output: 1004.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:02, 50.24it/s, est. speed input: 2644.33 toks/s, output: 1412.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:00<00:03, 36.88it/s, est. speed input: 2301.74 toks/s, output: 1516.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:03, 37.31it/s, est. speed input: 2309.39 toks/s, output: 1783.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:03, 37.83it/s, est. speed input: 2320.21 toks/s, output: 2086.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 42.22it/s, est. speed input: 2415.53 toks/s, output: 2527.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 45.98it/s, est. speed input: 2499.86 toks/s, output: 2919.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:02, 41.75it/s, est. speed input: 2453.28 toks/s, output: 3124.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:03, 29.76it/s, est. speed input: 2224.43 toks/s, output: 3093.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:02<00:03, 26.43it/s, est. speed input: 2120.61 toks/s, output: 3189.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:05<00:20,  4.20it/s, est. speed input: 850.39 toks/s, output: 1517.73 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:09<00:37,  2.20it/s, est. speed input: 518.97 toks/s, output: 1176.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10618.8 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:11<00:45,  1.79it/s, est. speed input: 430.72 toks/s, output: 1165.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:14<01:00,  1.30it/s, est. speed input: 342.06 toks/s, output: 1115.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:15<00:57,  1.35it/s, est. speed input: 334.05 toks/s, output: 1189.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:15<00:54,  1.40it/s, est. speed input: 326.10 toks/s, output: 1260.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<00:58,  1.30it/s, est. speed input: 309.50 toks/s, output: 1293.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:17<00:59,  1.27it/s, est. speed input: 297.40 toks/s, output: 1340.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:18<00:44,  1.62it/s, est. speed input: 293.95 toks/s, output: 1515.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:19<00:58,  1.24it/s, est. speed input: 273.95 toks/s, output: 1508.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8340.3 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:20<01:05,  1.08it/s, est. speed input: 259.45 toks/s, output: 1523.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<01:00,  1.15it/s, est. speed input: 253.75 toks/s, output: 1584.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:23<01:08,  1.01it/s, est. speed input: 241.50 toks/s, output: 1602.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:23<00:54,  1.25it/s, est. speed input: 241.30 toks/s, output: 1694.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:24<00:56,  1.19it/s, est. speed input: 234.29 toks/s, output: 1738.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:25<00:58,  1.13it/s, est. speed input: 227.32 toks/s, output: 1778.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:25<00:44,  1.45it/s, est. speed input: 228.01 toks/s, output: 1874.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:26<00:44,  1.42it/s, est. speed input: 223.89 toks/s, output: 1930.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<01:07,  1.07s/it, est. speed input: 210.52 toks/s, output: 1906.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:28<00:50,  1.22it/s, est. speed input: 211.05 toks/s, output: 2001.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:28<00:44,  1.36it/s, est. speed input: 209.18 toks/s, output: 2073.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:29<00:37,  1.60it/s, est. speed input: 208.63 toks/s, output: 2157.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7143.4 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:30<00:40,  1.45it/s, est. speed input: 204.82 toks/s, output: 2207.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:31<00:52,  1.10it/s, est. speed input: 197.49 toks/s, output: 2218.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:34<01:21,  1.44s/it, est. speed input: 183.77 toks/s, output: 2154.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:34<00:58,  1.05s/it, est. speed input: 184.80 toks/s, output: 2254.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:35<00:39,  1.37it/s, est. speed input: 184.45 toks/s, output: 2426.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:35<00:33,  1.59it/s, est. speed input: 184.52 toks/s, output: 2514.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:35<00:28,  1.83it/s, est. speed input: 184.58 toks/s, output: 2602.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:36<00:30,  1.68it/s, est. speed input: 182.54 toks/s, output: 2660.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:37<00:42,  1.17it/s, est. speed input: 176.74 toks/s, output: 2662.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:34,  1.42it/s, est. speed input: 176.87 toks/s, output: 2751.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:38<00:30,  1.60it/s, est. speed input: 176.43 toks/s, output: 2830.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6083.2 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:40<00:41,  1.14it/s, est. speed input: 171.38 toks/s, output: 2836.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:40<00:30,  1.52it/s, est. speed input: 172.38 toks/s, output: 2938.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:19,  2.23it/s, est. speed input: 173.73 toks/s, output: 3129.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:41<00:19,  2.16it/s, est. speed input: 173.03 toks/s, output: 3201.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:42<00:23,  1.79it/s, est. speed input: 171.11 toks/s, output: 3250.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:42<00:19,  2.12it/s, est. speed input: 171.57 toks/s, output: 3342.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:43<00:24,  1.66it/s, est. speed input: 169.24 toks/s, output: 3381.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:43<00:21,  1.80it/s, est. speed input: 168.90 toks/s, output: 3459.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:44<00:20,  1.83it/s, est. speed input: 168.29 toks/s, output: 3528.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:45<00:24,  1.50it/s, est. speed input: 166.04 toks/s, output: 3565.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:45<00:23,  1.56it/s, est. speed input: 165.24 toks/s, output: 3631.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:45<00:18,  1.94it/s, est. speed input: 165.82 toks/s, output: 3725.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:46<00:17,  1.99it/s, est. speed input: 165.42 toks/s, output: 3799.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:46<00:10,  3.13it/s, est. speed input: 167.34 toks/s, output: 4005.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:47<00:14,  2.15it/s, est. speed input: 165.38 toks/s, output: 4041.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:48<00:20,  1.48it/s, est. speed input: 162.28 toks/s, output: 4048.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:31:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4597.5 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:50<00:28,  1.01it/s, est. speed input: 157.53 toks/s, output: 4014.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:50<00:22,  1.25it/s, est. speed input: 157.82 toks/s, output: 4105.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:51<00:21,  1.26it/s, est. speed input: 156.61 toks/s, output: 4157.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:51<00:16,  1.58it/s, est. speed input: 157.10 toks/s, output: 4253.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:52<00:12,  1.95it/s, est. speed input: 157.63 toks/s, output: 4350.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:54<00:16,  1.36it/s, est. speed input: 153.98 toks/s, output: 4415.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:54<00:14,  1.51it/s, est. speed input: 153.87 toks/s, output: 4496.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:57<00:27,  1.30s/it, est. speed input: 146.62 toks/s, output: 4372.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:58<00:17,  1.11it/s, est. speed input: 146.78 toks/s, output: 4552.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:58<00:14,  1.26it/s, est. speed input: 146.69 toks/s, output: 4637.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:59<00:11,  1.50it/s, est. speed input: 147.06 toks/s, output: 4734.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:32:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3419.1 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:00<00:11,  1.39it/s, est. speed input: 145.96 toks/s, output: 4786.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:00<00:08,  1.71it/s, est. speed input: 146.39 toks/s, output: 4887.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:02<00:15,  1.14s/it, est. speed input: 141.41 toks/s, output: 4810.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:03<00:14,  1.10s/it, est. speed input: 140.12 toks/s, output: 4856.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:05<00:16,  1.38s/it, est. speed input: 136.60 toks/s, output: 4828.16 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.38s/it, est. speed input: 147.69 toks/s, output: 6319.32 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  2.43it/s, est. speed input: 147.69 toks/s, output: 6319.32 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:32:16 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:32:16 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:32:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:32:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:32:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:32:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:32:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:32:17 [abstract.py:306] It took 0.550212 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:32:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:32:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:32:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0116
+  • Training Accuracy: 0.2938
+  • Mean Generation Length: 2603.6875
+  • Training FLOPS: 47.86 TFLOPS (11.97 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 310.46s
+  • policy_training: 198.01s (63.8%)
+  • generation: 70.34s (22.7%)
+  • teacher_logprob_inference: 36.75s (11.8%)
+  • policy_training/sharding_data: 4.35s (1.4%)
+  • prepare_for_generation: 3.70s (1.2%)
+  • policy_training/submit_training_futures: 1.65s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.61s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 94: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:14 [abstract.py:324] It took 0.055728 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07944531e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52823563e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84032745e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07944531e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52823563e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84032745e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07944531e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52823563e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84032745e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07944531e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52823563e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84032745e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07947891e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53320045e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84194641e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07947891e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53320045e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84194641e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.59601898e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33553963e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #94] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44944305e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07947891e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53320045e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84194641e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35824051e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.24390173e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #94] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.22519760e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07947891e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53320045e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84194641e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30237274e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54443417e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #94] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33974808e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.14405022e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.20808220e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #94] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44039780e+02
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 94: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:17 [abstract.py:324] It took 0.030587 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13216.13it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:17 [loggers.py:236] Engine 000: Avg prompt throughput: 38.8 tokens/s, Avg generation throughput: 57.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:59,  2.69it/s, est. speed input: 169.42 toks/s, output: 18.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 15.59it/s, est. speed input: 766.90 toks/s, output: 144.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:07, 21.03it/s, est. speed input: 1026.33 toks/s, output: 271.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 26.02it/s, est. speed input: 1224.69 toks/s, output: 423.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 25.59it/s, est. speed input: 1280.36 toks/s, output: 574.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 31.12it/s, est. speed input: 1460.15 toks/s, output: 825.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 33.03it/s, est. speed input: 1567.74 toks/s, output: 1093.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 40.68it/s, est. speed input: 1767.82 toks/s, output: 1498.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 41.03it/s, est. speed input: 1845.15 toks/s, output: 1746.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 46.53it/s, est. speed input: 1987.88 toks/s, output: 2149.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 54.78it/s, est. speed input: 2162.19 toks/s, output: 2645.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 61.50it/s, est. speed input: 2318.35 toks/s, output: 3141.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:01, 47.32it/s, est. speed input: 2264.09 toks/s, output: 3344.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:02<00:03, 23.05it/s, est. speed input: 1856.11 toks/s, output: 3034.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:06<00:17,  4.58it/s, est. speed input: 791.72 toks/s, output: 1561.01 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:07<00:18,  4.15it/s, est. speed input: 703.65 toks/s, output: 1669.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:07<00:16,  4.48it/s, est. speed input: 686.35 toks/s, output: 1913.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:09<00:21,  3.29it/s, est. speed input: 591.89 toks/s, output: 1843.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10148.2 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:11<00:30,  2.20it/s, est. speed input: 487.38 toks/s, output: 1709.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:12<00:33,  1.98it/s, est. speed input: 443.86 toks/s, output: 1746.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:13<00:26,  2.46it/s, est. speed input: 448.10 toks/s, output: 1961.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:27,  2.28it/s, est. speed input: 431.80 toks/s, output: 1988.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:15<00:44,  1.39it/s, est. speed input: 378.28 toks/s, output: 1842.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:16<00:41,  1.47it/s, est. speed input: 370.10 toks/s, output: 1904.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:42,  1.41it/s, est. speed input: 356.19 toks/s, output: 1934.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:50,  1.15it/s, est. speed input: 320.96 toks/s, output: 1934.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:47,  1.20it/s, est. speed input: 312.75 toks/s, output: 1987.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7717.8 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:41,  1.35it/s, est. speed input: 309.00 toks/s, output: 2065.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:21<00:30,  1.78it/s, est. speed input: 306.21 toks/s, output: 2245.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:21<00:20,  2.54it/s, est. speed input: 308.88 toks/s, output: 2463.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:21<00:16,  2.94it/s, est. speed input: 307.66 toks/s, output: 2648.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:22<00:19,  2.47it/s, est. speed input: 301.27 toks/s, output: 2690.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:22<00:17,  2.73it/s, est. speed input: 300.95 toks/s, output: 2784.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:22<00:11,  3.89it/s, est. speed input: 304.15 toks/s, output: 3005.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:23<00:15,  2.92it/s, est. speed input: 298.32 toks/s, output: 3042.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:27,  1.59it/s, est. speed input: 282.11 toks/s, output: 2974.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:26<00:31,  1.38it/s, est. speed input: 273.49 toks/s, output: 2981.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:30,  1.39it/s, est. speed input: 268.63 toks/s, output: 3025.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:29<00:49,  1.20s/it, est. speed input: 247.97 toks/s, output: 2893.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:29<00:37,  1.08it/s, est. speed input: 248.07 toks/s, output: 2993.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:28,  1.38it/s, est. speed input: 248.33 toks/s, output: 3094.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:17,  2.14it/s, est. speed input: 249.86 toks/s, output: 3308.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5901.8 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:30<00:19,  1.88it/s, est. speed input: 245.81 toks/s, output: 3353.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:31<00:19,  1.78it/s, est. speed input: 242.62 toks/s, output: 3408.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:23,  1.39it/s, est. speed input: 232.55 toks/s, output: 3457.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:20,  1.55it/s, est. speed input: 231.57 toks/s, output: 3540.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:34<00:21,  1.44it/s, est. speed input: 227.68 toks/s, output: 3579.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:22,  1.34it/s, est. speed input: 223.61 toks/s, output: 3614.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:15,  1.79it/s, est. speed input: 221.49 toks/s, output: 3866.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:36<00:12,  2.10it/s, est. speed input: 222.27 toks/s, output: 3977.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:37<00:11,  2.10it/s, est. speed input: 220.99 toks/s, output: 4052.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:13,  1.76it/s, est. speed input: 217.67 toks/s, output: 4088.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:38<00:11,  2.00it/s, est. speed input: 217.50 toks/s, output: 4183.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:08,  2.51it/s, est. speed input: 218.40 toks/s, output: 4298.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:09,  2.24it/s, est. speed input: 216.76 toks/s, output: 4362.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:07,  2.83it/s, est. speed input: 217.63 toks/s, output: 4477.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:39<00:05,  3.52it/s, est. speed input: 218.61 toks/s, output: 4592.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:39<00:05,  3.23it/s, est. speed input: 218.06 toks/s, output: 4678.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:36:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4136.0 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:41<00:10,  1.63it/s, est. speed input: 212.41 toks/s, output: 4654.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:12,  1.32it/s, est. speed input: 208.27 toks/s, output: 4663.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:43<00:12,  1.24it/s, est. speed input: 205.28 toks/s, output: 4695.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:43<00:09,  1.53it/s, est. speed input: 205.25 toks/s, output: 4796.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:44<00:08,  1.47it/s, est. speed input: 203.21 toks/s, output: 4849.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:45<00:09,  1.21it/s, est. speed input: 199.31 toks/s, output: 4859.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:45<00:08,  1.35it/s, est. speed input: 198.19 toks/s, output: 4935.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:46<00:06,  1.66it/s, est. speed input: 198.27 toks/s, output: 5041.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:46<00:05,  1.53it/s, est. speed input: 196.31 toks/s, output: 5096.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:47<00:04,  1.93it/s, est. speed input: 196.75 toks/s, output: 5211.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:47<00:02,  2.45it/s, est. speed input: 197.41 toks/s, output: 5332.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:49<00:05,  1.15it/s, est. speed input: 190.85 toks/s, output: 5262.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:50<00:04,  1.06it/s, est. speed input: 187.79 toks/s, output: 5288.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:37:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2230.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:52<00:05,  1.32s/it, est. speed input: 181.01 toks/s, output: 5211.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:53<00:03,  1.13s/it, est. speed input: 179.81 toks/s, output: 5292.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:54<00:00,  1.11it/s, est. speed input: 177.88 toks/s, output: 5465.70 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.11it/s, est. speed input: 178.98 toks/s, output: 5616.29 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.94it/s, est. speed input: 178.98 toks/s, output: 5616.29 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:37:14 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:37:14 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:37:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:37:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:37:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:37:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:37:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 06:37:15 [abstract.py:306] It took 0.522973 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:37:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:37:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:37:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0153
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 1909.3875
+  • Training FLOPS: 40.44 TFLOPS (10.11 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 262.10s
+  • policy_training: 169.34s (64.6%)
+  • generation: 57.69s (22.0%)
+  • teacher_logprob_inference: 30.00s (11.4%)
+  • policy_training/sharding_data: 4.32s (1.6%)
+  • prepare_for_generation: 3.58s (1.4%)
+  • policy_training/submit_training_futures: 1.76s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.6%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 95: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:40:36 [abstract.py:324] It took 0.049542 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07947695e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52189026e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83388214e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07947695e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52189026e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83388214e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07947695e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52189026e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83388214e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07947695e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52189026e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83388214e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07945859e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53874893e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83807449e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.14626884e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.06319237e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #95] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43850525e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07945859e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53874893e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83807449e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.60352325e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33678694e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #95] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43980560e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07945859e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53874893e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83807449e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37395020e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.33434916e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #95] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.22409973e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07945859e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.53874893e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83807449e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29223785e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.55480375e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #95] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33735229e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 95: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:40:39 [abstract.py:324] It took 0.030253 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 06:40:39 [loggers.py:236] Engine 000: Avg prompt throughput: 46.0 tokens/s, Avg generation throughput: 18.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13555.98it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.08it/s, est. speed input: 376.98 toks/s, output: 79.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:04, 30.55it/s, est. speed input: 1632.47 toks/s, output: 474.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 29.36it/s, est. speed input: 1659.17 toks/s, output: 642.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:06, 23.81it/s, est. speed input: 1482.62 toks/s, output: 747.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:06, 22.04it/s, est. speed input: 1418.83 toks/s, output: 867.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:05, 26.68it/s, est. speed input: 1548.20 toks/s, output: 1188.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:04, 29.61it/s, est. speed input: 1632.31 toks/s, output: 1476.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 41.43it/s, est. speed input: 1898.95 toks/s, output: 2085.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 42.55it/s, est. speed input: 1976.13 toks/s, output: 2377.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 39.93it/s, est. speed input: 1997.93 toks/s, output: 2651.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 40.21it/s, est. speed input: 2034.96 toks/s, output: 2933.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 46.49it/s, est. speed input: 2170.87 toks/s, output: 3541.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:02, 46.49it/s, est. speed input: 2210.58 toks/s, output: 3831.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:03<00:07, 11.93it/s, est. speed input: 1399.19 toks/s, output: 2633.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:06<00:19,  4.24it/s, est. speed input: 759.67 toks/s, output: 1715.95 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:09<00:33,  2.39it/s, est. speed input: 504.64 toks/s, output: 1401.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:40:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10617.7 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:10<00:35,  2.21it/s, est. speed input: 456.25 toks/s, output: 1459.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:15<01:04,  1.20it/s, est. speed input: 319.50 toks/s, output: 1218.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<01:01,  1.23it/s, est. speed input: 310.63 toks/s, output: 1282.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:16<00:56,  1.32it/s, est. speed input: 306.62 toks/s, output: 1363.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:17<00:53,  1.39it/s, est. speed input: 300.83 toks/s, output: 1434.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:19<01:07,  1.08it/s, est. speed input: 276.14 toks/s, output: 1413.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:40:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8307.4 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<01:10,  1.02it/s, est. speed input: 263.14 toks/s, output: 1442.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:24<01:31,  1.31s/it, est. speed input: 228.14 toks/s, output: 1426.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:24<01:21,  1.19s/it, est. speed input: 223.78 toks/s, output: 1494.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:26<01:25,  1.25s/it, est. speed input: 213.51 toks/s, output: 1519.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:26<01:06,  1.00it/s, est. speed input: 213.93 toks/s, output: 1614.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:27<01:11,  1.08s/it, est. speed input: 206.10 toks/s, output: 1647.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:29<01:24,  1.29s/it, est. speed input: 195.16 toks/s, output: 1652.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7128.9 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:30<01:20,  1.25s/it, est. speed input: 189.88 toks/s, output: 1699.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:30<00:58,  1.08it/s, est. speed input: 191.11 toks/s, output: 1801.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:31<00:56,  1.10it/s, est. speed input: 188.01 toks/s, output: 1861.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:32<00:47,  1.28it/s, est. speed input: 187.04 toks/s, output: 1942.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:32<00:26,  2.24it/s, est. speed input: 190.22 toks/s, output: 2153.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:32<00:25,  2.27it/s, est. speed input: 189.74 toks/s, output: 2234.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:34<00:44,  1.29it/s, est. speed input: 182.05 toks/s, output: 2230.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:34<00:36,  1.54it/s, est. speed input: 182.17 toks/s, output: 2318.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:36<00:52,  1.04it/s, est. speed input: 175.04 toks/s, output: 2314.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:40,  1.32it/s, est. speed input: 175.51 toks/s, output: 2407.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:38<00:49,  1.07it/s, est. speed input: 170.69 toks/s, output: 2428.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:38<00:40,  1.27it/s, est. speed input: 170.48 toks/s, output: 2511.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:39<00:45,  1.13it/s, est. speed input: 167.16 toks/s, output: 2547.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6206.3 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:40<00:34,  1.43it/s, est. speed input: 167.70 toks/s, output: 2640.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:40<00:32,  1.49it/s, est. speed input: 166.72 toks/s, output: 2709.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:40<00:19,  2.40it/s, est. speed input: 168.72 toks/s, output: 2910.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:41<00:15,  2.93it/s, est. speed input: 169.73 toks/s, output: 3011.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:41<00:12,  3.51it/s, est. speed input: 171.02 toks/s, output: 3197.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:42<00:13,  3.22it/s, est. speed input: 171.00 toks/s, output: 3359.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:42<00:16,  2.55it/s, est. speed input: 169.62 toks/s, output: 3414.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:43<00:13,  2.87it/s, est. speed input: 170.21 toks/s, output: 3507.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:44<00:23,  1.65it/s, est. speed input: 166.31 toks/s, output: 3507.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:46<00:33,  1.13it/s, est. speed input: 161.64 toks/s, output: 3490.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:47<00:33,  1.12it/s, est. speed input: 159.74 toks/s, output: 3531.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:47<00:27,  1.30it/s, est. speed input: 159.54 toks/s, output: 3608.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:47<00:24,  1.42it/s, est. speed input: 159.04 toks/s, output: 3677.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:48<00:15,  2.08it/s, est. speed input: 160.07 toks/s, output: 3864.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4724.6 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:50<00:25,  1.28it/s, est. speed input: 155.76 toks/s, output: 3843.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:50<00:23,  1.29it/s, est. speed input: 154.62 toks/s, output: 3897.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:52<00:29,  1.01it/s, est. speed input: 151.18 toks/s, output: 3892.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:52<00:22,  1.29it/s, est. speed input: 151.73 toks/s, output: 3988.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:53<00:18,  1.49it/s, est. speed input: 151.72 toks/s, output: 4070.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:56<00:27,  1.04s/it, est. speed input: 145.76 toks/s, output: 4072.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:57<00:26,  1.06s/it, est. speed input: 143.94 toks/s, output: 4106.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:57<00:15,  1.53it/s, est. speed input: 145.75 toks/s, output: 4325.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:59<00:21,  1.00it/s, est. speed input: 141.46 toks/s, output: 4282.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:59<00:17,  1.23it/s, est. speed input: 141.89 toks/s, output: 4380.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3739.6 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:01<00:19,  1.04it/s, est. speed input: 139.69 toks/s, output: 4398.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:02<00:14,  1.23it/s, est. speed input: 138.85 toks/s, output: 4542.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:03<00:16,  1.03it/s, est. speed input: 136.57 toks/s, output: 4553.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:04<00:15,  1.05it/s, est. speed input: 135.64 toks/s, output: 4609.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:05<00:13,  1.11it/s, est. speed input: 134.99 toks/s, output: 4675.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:06<00:14,  1.00s/it, est. speed input: 133.32 toks/s, output: 4706.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:07<00:12,  1.02it/s, est. speed input: 132.40 toks/s, output: 4762.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:07<00:08,  1.36it/s, est. speed input: 133.10 toks/s, output: 4874.47 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:07<00:00,  1.36it/s, est. speed input: 143.77 toks/s, output: 6323.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:07<00:00,  2.36it/s, est. speed input: 143.77 toks/s, output: 6323.72 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:41:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:41:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:41:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:41:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:41:51 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:41:51 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:41:51 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:41:52 [abstract.py:306] It took 0.902330 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:41:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0121
+  • Training Accuracy: 0.2687
+  • Mean Generation Length: 2680.8625
+  • Training FLOPS: 48.40 TFLOPS (12.10 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 316.06s
+  • policy_training: 202.15s (64.0%)
+  • generation: 72.59s (23.0%)
+  • teacher_logprob_inference: 36.23s (11.5%)
+  • policy_training/sharding_data: 4.35s (1.4%)
+  • prepare_for_generation: 3.56s (1.1%)
+  • policy_training/submit_training_futures: 1.89s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Epoch 20/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 96: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:45:53 [abstract.py:324] It took 0.050863 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07968262e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51133232e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83615723e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07968262e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51133232e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83615723e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07968262e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51133232e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83615723e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07968262e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51133232e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83615723e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07963711e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52164459e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85187378e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.13107491e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.99193907e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #96] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43407806e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07963711e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52164459e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85187378e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.60274353e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.32429180e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #96] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44449387e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07963711e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52164459e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85187378e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37418518e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.34168243e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #96] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24911995e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07963711e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52164459e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85187378e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29577942e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.55866575e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #96] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33818741e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 96: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:45:56 [abstract.py:324] It took 0.029416 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13141.08it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 06:45:56 [loggers.py:236] Engine 000: Avg prompt throughput: 38.1 tokens/s, Avg generation throughput: 83.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:44,  3.60it/s, est. speed input: 216.25 toks/s, output: 43.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:09, 15.66it/s, est. speed input: 780.14 toks/s, output: 219.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:06, 23.31it/s, est. speed input: 1121.61 toks/s, output: 409.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 31.29it/s, est. speed input: 1442.64 toks/s, output: 654.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 31.05it/s, est. speed input: 1528.08 toks/s, output: 793.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 33.70it/s, est. speed input: 1677.11 toks/s, output: 1076.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 37.01it/s, est. speed input: 1819.82 toks/s, output: 1392.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 54.00it/s, est. speed input: 2241.13 toks/s, output: 2132.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:01, 63.30it/s, est. speed input: 2529.01 toks/s, output: 2774.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 56.28it/s, est. speed input: 2546.07 toks/s, output: 3051.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 52.05it/s, est. speed input: 2555.04 toks/s, output: 3299.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:02<00:03, 26.50it/s, est. speed input: 2080.80 toks/s, output: 2949.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:05<00:16,  5.25it/s, est. speed input: 852.64 toks/s, output: 1586.84 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:07<00:24,  3.29it/s, est. speed input: 602.43 toks/s, output: 1398.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:08<00:22,  3.56it/s, est. speed input: 583.75 toks/s, output: 1641.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10531.5 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<00:51,  1.50it/s, est. speed input: 365.90 toks/s, output: 1227.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:15<00:50,  1.48it/s, est. speed input: 339.56 toks/s, output: 1333.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:15<00:45,  1.64it/s, est. speed input: 340.96 toks/s, output: 1436.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:16<00:49,  1.47it/s, est. speed input: 321.80 toks/s, output: 1452.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:17<00:48,  1.47it/s, est. speed input: 312.85 toks/s, output: 1508.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:18<00:56,  1.26it/s, est. speed input: 294.83 toks/s, output: 1517.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:18<00:48,  1.46it/s, est. speed input: 293.49 toks/s, output: 1605.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:19<00:32,  2.09it/s, est. speed input: 295.32 toks/s, output: 1803.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:19<00:30,  2.21it/s, est. speed input: 293.12 toks/s, output: 1882.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8221.7 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:20<00:32,  2.01it/s, est. speed input: 286.81 toks/s, output: 1933.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:20<00:32,  2.01it/s, est. speed input: 282.90 toks/s, output: 1998.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:21<00:33,  1.91it/s, est. speed input: 277.88 toks/s, output: 2054.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:21<00:36,  1.72it/s, est. speed input: 271.28 toks/s, output: 2096.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:22<00:37,  1.66it/s, est. speed input: 266.08 toks/s, output: 2147.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:22<00:30,  1.98it/s, est. speed input: 265.69 toks/s, output: 2235.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:23<00:31,  1.90it/s, est. speed input: 261.76 toks/s, output: 2291.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:23<00:25,  2.34it/s, est. speed input: 262.25 toks/s, output: 2385.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<00:16,  3.56it/s, est. speed input: 265.00 toks/s, output: 2586.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:24<00:15,  3.58it/s, est. speed input: 264.53 toks/s, output: 2669.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:26<00:44,  1.23it/s, est. speed input: 243.59 toks/s, output: 2547.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:26<00:33,  1.59it/s, est. speed input: 244.69 toks/s, output: 2647.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:27<00:35,  1.51it/s, est. speed input: 240.31 toks/s, output: 2687.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:27<00:32,  1.62it/s, est. speed input: 238.06 toks/s, output: 2750.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:28<00:26,  1.96it/s, est. speed input: 238.11 toks/s, output: 2839.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:28<00:21,  2.31it/s, est. speed input: 236.71 toks/s, output: 2995.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:29<00:18,  2.59it/s, est. speed input: 235.79 toks/s, output: 3155.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6612.0 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:30<00:23,  1.93it/s, est. speed input: 230.09 toks/s, output: 3167.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:30<00:19,  2.30it/s, est. speed input: 230.80 toks/s, output: 3263.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:31<00:16,  2.64it/s, est. speed input: 230.19 toks/s, output: 3426.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:31<00:15,  2.73it/s, est. speed input: 229.77 toks/s, output: 3505.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:32<00:17,  2.29it/s, est. speed input: 226.95 toks/s, output: 3548.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:32<00:15,  2.51it/s, est. speed input: 226.81 toks/s, output: 3631.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:32<00:13,  2.96it/s, est. speed input: 227.51 toks/s, output: 3727.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:32<00:09,  3.76it/s, est. speed input: 228.72 toks/s, output: 3917.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:33<00:09,  3.66it/s, est. speed input: 228.44 toks/s, output: 3997.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:08,  3.94it/s, est. speed input: 228.98 toks/s, output: 4173.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:34<00:10,  3.07it/s, est. speed input: 226.95 toks/s, output: 4219.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:35<00:14,  2.18it/s, est. speed input: 222.99 toks/s, output: 4231.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:12,  2.58it/s, est. speed input: 223.56 toks/s, output: 4325.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:12,  2.45it/s, est. speed input: 222.38 toks/s, output: 4387.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:07,  3.66it/s, est. speed input: 224.54 toks/s, output: 4679.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:13,  1.94it/s, est. speed input: 217.41 toks/s, output: 4616.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:15,  1.58it/s, est. speed input: 213.19 toks/s, output: 4611.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:12,  1.99it/s, est. speed input: 214.16 toks/s, output: 4718.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:07,  3.12it/s, est. speed input: 216.71 toks/s, output: 4942.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:06,  3.13it/s, est. speed input: 216.48 toks/s, output: 5022.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4532.5 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:40<00:09,  2.20it/s, est. speed input: 213.32 toks/s, output: 5034.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:40<00:05,  3.40it/s, est. speed input: 215.72 toks/s, output: 5259.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:04,  3.67it/s, est. speed input: 216.19 toks/s, output: 5354.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:40<00:03,  4.12it/s, est. speed input: 216.84 toks/s, output: 5456.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:04,  3.02it/s, est. speed input: 215.01 toks/s, output: 5578.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:43<00:09,  1.38it/s, est. speed input: 206.22 toks/s, output: 5440.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:05,  2.13it/s, est. speed input: 208.19 toks/s, output: 5672.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:04,  2.46it/s, est. speed input: 208.74 toks/s, output: 5775.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:03,  2.43it/s, est. speed input: 208.04 toks/s, output: 5846.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:05,  1.54it/s, est. speed input: 203.29 toks/s, output: 5805.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:45<00:02,  2.44it/s, est. speed input: 205.21 toks/s, output: 6044.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:01,  2.92it/s, est. speed input: 206.04 toks/s, output: 6159.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:47<00:02,  1.83it/s, est. speed input: 202.20 toks/s, output: 6140.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:47<00:01,  1.96it/s, est. speed input: 201.76 toks/s, output: 6221.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1798.5 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:54<00:04,  2.19s/it, est. speed input: 178.00 toks/s, output: 5604.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.32s/it, est. speed input: 178.97 toks/s, output: 5862.08 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.32s/it, est. speed input: 178.97 toks/s, output: 5862.08 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.93it/s, est. speed input: 178.97 toks/s, output: 5862.08 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:53 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:53 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:46:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:46:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:46:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:46:53 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:46:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:46:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:46:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:46:54 [abstract.py:306] It took 0.836659 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:46:54 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.3875
+  • Mean Generation Length: 2002.9000
+  • Training FLOPS: 42.20 TFLOPS (10.55 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 262.65s
+  • policy_training: 168.44s (64.1%)
+  • generation: 58.42s (22.2%)
+  • teacher_logprob_inference: 30.71s (11.7%)
+  • policy_training/sharding_data: 4.37s (1.7%)
+  • prepare_for_generation: 3.57s (1.4%)
+  • policy_training/submit_training_futures: 1.63s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.6%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 97: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:15 [abstract.py:324] It took 0.048898 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07991309e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49942131e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83235931e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07991309e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49942131e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83235931e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07991309e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49942131e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83235931e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07991309e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49942131e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83235931e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07992246e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50363922e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85062943e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.11586151e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.91785145e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #97] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43542969e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07992246e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50363922e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85062943e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.60759811e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.32640963e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #97] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.46621780e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07992246e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50363922e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85062943e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37454453e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.26086807e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #97] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24844475e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07992246e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50363922e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85062943e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29474983e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54120884e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #97] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33824875e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 97: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:18 [abstract.py:324] It took 0.028906 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 10963.71it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:18 [loggers.py:236] Engine 000: Avg prompt throughput: 45.5 tokens/s, Avg generation throughput: 16.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:42,  3.73it/s, est. speed input: 216.30 toks/s, output: 44.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 26.23it/s, est. speed input: 1272.70 toks/s, output: 382.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:08, 18.20it/s, est. speed input: 1060.37 toks/s, output: 428.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:05, 24.96it/s, est. speed input: 1324.18 toks/s, output: 808.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 29.69it/s, est. speed input: 1488.88 toks/s, output: 1105.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 37.35it/s, est. speed input: 1707.94 toks/s, output: 1556.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 44.79it/s, est. speed input: 1917.16 toks/s, output: 2002.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 39.80it/s, est. speed input: 1919.31 toks/s, output: 2241.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:01, 54.55it/s, est. speed input: 2242.72 toks/s, output: 3126.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 55.64it/s, est. speed input: 2327.69 toks/s, output: 3478.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:03<00:07, 11.53it/s, est. speed input: 1250.65 toks/s, output: 2106.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10830.7 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:12<00:43,  2.00it/s, est. speed input: 356.57 toks/s, output: 946.33 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:17<00:56,  1.48it/s, est. speed input: 271.22 toks/s, output: 1004.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:19<00:55,  1.44it/s, est. speed input: 248.06 toks/s, output: 1193.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:19<00:48,  1.62it/s, est. speed input: 249.32 toks/s, output: 1384.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8734.0 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:21<00:51,  1.47it/s, est. speed input: 233.04 toks/s, output: 1470.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:24<01:10,  1.07it/s, est. speed input: 208.26 toks/s, output: 1404.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:25<01:06,  1.11it/s, est. speed input: 204.74 toks/s, output: 1470.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<00:47,  1.50it/s, est. speed input: 208.27 toks/s, output: 1672.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:26<00:40,  1.73it/s, est. speed input: 207.04 toks/s, output: 1834.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:26<00:37,  1.86it/s, est. speed input: 206.70 toks/s, output: 1916.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:27<00:34,  1.95it/s, est. speed input: 205.81 toks/s, output: 1992.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:27<00:29,  2.27it/s, est. speed input: 206.64 toks/s, output: 2084.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:29<00:56,  1.17it/s, est. speed input: 193.51 toks/s, output: 2036.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:29<00:49,  1.31it/s, est. speed input: 192.23 toks/s, output: 2107.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7321.4 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:30<00:35,  1.76it/s, est. speed input: 192.40 toks/s, output: 2274.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:30<00:32,  1.89it/s, est. speed input: 191.78 toks/s, output: 2350.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:31<00:36,  1.66it/s, est. speed input: 188.61 toks/s, output: 2394.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:32<00:31,  1.89it/s, est. speed input: 187.38 toks/s, output: 2540.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:32<00:25,  2.29it/s, est. speed input: 188.44 toks/s, output: 2635.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:32<00:16,  3.38it/s, est. speed input: 191.23 toks/s, output: 2832.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:33<00:21,  2.51it/s, est. speed input: 188.70 toks/s, output: 2873.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:34<00:21,  2.47it/s, est. speed input: 188.10 toks/s, output: 2943.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:34<00:21,  2.47it/s, est. speed input: 187.63 toks/s, output: 3014.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:34<00:22,  2.35it/s, est. speed input: 186.76 toks/s, output: 3078.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:35<00:20,  2.52it/s, est. speed input: 186.81 toks/s, output: 3156.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:35<00:20,  2.47it/s, est. speed input: 186.24 toks/s, output: 3224.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:18,  2.60it/s, est. speed input: 186.17 toks/s, output: 3300.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:37<00:33,  1.43it/s, est. speed input: 180.56 toks/s, output: 3278.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:43,  1.08it/s, est. speed input: 175.28 toks/s, output: 3262.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:39<00:24,  1.83it/s, est. speed input: 177.43 toks/s, output: 3459.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:28,  1.52it/s, est. speed input: 174.56 toks/s, output: 3481.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:50:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6052.0 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:22,  1.92it/s, est. speed input: 175.48 toks/s, output: 3577.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:40<00:17,  2.39it/s, est. speed input: 176.35 toks/s, output: 3672.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:40<00:16,  2.51it/s, est. speed input: 176.31 toks/s, output: 3749.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:40<00:12,  3.09it/s, est. speed input: 177.21 toks/s, output: 3844.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:41<00:10,  3.78it/s, est. speed input: 178.11 toks/s, output: 3941.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:41<00:11,  3.38it/s, est. speed input: 177.94 toks/s, output: 4013.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:16,  2.29it/s, est. speed input: 176.08 toks/s, output: 4048.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:24,  1.49it/s, est. speed input: 172.51 toks/s, output: 4042.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:25,  1.37it/s, est. speed input: 170.47 toks/s, output: 4072.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:24,  1.36it/s, est. speed input: 168.98 toks/s, output: 4114.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:25,  1.28it/s, est. speed input: 166.97 toks/s, output: 4142.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:47<00:30,  1.05it/s, est. speed input: 163.45 toks/s, output: 4133.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:47<00:22,  1.40it/s, est. speed input: 164.15 toks/s, output: 4229.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:47<00:17,  1.70it/s, est. speed input: 164.42 toks/s, output: 4313.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:47<00:13,  2.16it/s, est. speed input: 165.16 toks/s, output: 4408.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:06,  4.20it/s, est. speed input: 168.27 toks/s, output: 4721.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:48<00:05,  4.20it/s, est. speed input: 168.75 toks/s, output: 4809.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:48<00:05,  4.59it/s, est. speed input: 169.46 toks/s, output: 4904.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:05,  4.30it/s, est. speed input: 170.12 toks/s, output: 5074.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:05,  3.99it/s, est. speed input: 170.30 toks/s, output: 5153.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:49<00:04,  4.58it/s, est. speed input: 171.11 toks/s, output: 5252.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:51:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4213.7 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:50<00:04,  3.86it/s, est. speed input: 171.68 toks/s, output: 5491.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:50<00:04,  3.93it/s, est. speed input: 172.07 toks/s, output: 5578.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:52<00:07,  1.93it/s, est. speed input: 167.93 toks/s, output: 5592.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:52<00:05,  2.21it/s, est. speed input: 168.41 toks/s, output: 5686.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:05,  2.20it/s, est. speed input: 167.73 toks/s, output: 5820.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:54<00:04,  2.30it/s, est. speed input: 167.75 toks/s, output: 5899.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:54<00:03,  2.50it/s, est. speed input: 168.01 toks/s, output: 5986.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:54<00:02,  3.32it/s, est. speed input: 169.31 toks/s, output: 6189.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:02,  2.63it/s, est. speed input: 168.37 toks/s, output: 6235.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:59<00:04,  1.06s/it, est. speed input: 158.21 toks/s, output: 6024.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:51:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1774.9 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:01<00:03,  1.33s/it, est. speed input: 153.29 toks/s, output: 5933.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.33s/it, est. speed input: 156.26 toks/s, output: 6330.82 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.59it/s, est. speed input: 156.26 toks/s, output: 6330.82 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:51:24 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:51:24 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:51:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:51:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:51:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:51:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:51:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:51:24 [abstract.py:306] It took 0.542319 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:51:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:51:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:51:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0107
+  • Training Accuracy: 0.3250
+  • Mean Generation Length: 2447.0500
+  • Training FLOPS: 44.40 TFLOPS (11.10 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 304.79s
+  • policy_training: 197.13s (64.7%)
+  • generation: 65.92s (21.6%)
+  • teacher_logprob_inference: 36.74s (12.1%)
+  • policy_training/sharding_data: 4.34s (1.4%)
+  • prepare_for_generation: 3.48s (1.1%)
+  • policy_training/submit_training_futures: 1.75s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 98: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:55:21 [abstract.py:324] It took 0.051138 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07993223e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49571152e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82402191e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07993223e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49571152e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82402191e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07993223e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49571152e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82402191e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.07993223e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.49571152e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82402191e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07995586e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.49227638e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82527695e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.11007633e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.79526901e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #98] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43473541e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07995586e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.49227638e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82527695e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62257080e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.32702560e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #98] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.49125290e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07995586e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.49227638e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82527695e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37699203e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.31929135e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #98] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24847183e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.07995586e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.49227638e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82527695e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28392220e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53729296e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #98] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33624634e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 98: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:55:24 [abstract.py:324] It took 0.030542 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 06:55:24 [loggers.py:236] Engine 000: Avg prompt throughput: 39.9 tokens/s, Avg generation throughput: 7.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12514.94it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.05it/s, est. speed input: 375.34 toks/s, output: 78.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:05, 27.74it/s, est. speed input: 1466.96 toks/s, output: 432.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:02, 50.42it/s, est. speed input: 2508.15 toks/s, output: 1003.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:02, 57.14it/s, est. speed input: 2876.28 toks/s, output: 1387.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:02, 51.87it/s, est. speed input: 2847.73 toks/s, output: 1650.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 50.91it/s, est. speed input: 2869.69 toks/s, output: 1927.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:00<00:02, 50.68it/s, est. speed input: 2894.26 toks/s, output: 2222.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:01, 58.94it/s, est. speed input: 3118.04 toks/s, output: 2813.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 57.97it/s, est. speed input: 3144.96 toks/s, output: 3158.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 67.79it/s, est. speed input: 3361.38 toks/s, output: 3854.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:01<00:01, 68.88it/s, est. speed input: 3450.83 toks/s, output: 4328.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:01<00:01, 54.46it/s, est. speed input: 3282.12 toks/s, output: 4518.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:02<00:04, 15.77it/s, est. speed input: 1985.54 toks/s, output: 3092.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:55:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9641.9 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:11<00:29,  2.18it/s, est. speed input: 491.38 toks/s, output: 1252.58 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:14<00:29,  2.04it/s, est. speed input: 425.12 toks/s, output: 1510.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:15<00:27,  2.08it/s, est. speed input: 400.22 toks/s, output: 1745.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:16<00:25,  2.24it/s, est. speed input: 395.30 toks/s, output: 1941.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:17<00:24,  2.21it/s, est. speed input: 380.47 toks/s, output: 2080.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:17<00:23,  2.24it/s, est. speed input: 375.23 toks/s, output: 2159.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:17<00:22,  2.33it/s, est. speed input: 371.85 toks/s, output: 2246.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:18<00:20,  2.46it/s, est. speed input: 369.42 toks/s, output: 2337.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:18<00:18,  2.65it/s, est. speed input: 367.47 toks/s, output: 2430.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:18<00:17,  2.87it/s, est. speed input: 366.08 toks/s, output: 2525.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:18<00:09,  4.60it/s, est. speed input: 371.43 toks/s, output: 2873.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:19<00:13,  3.27it/s, est. speed input: 361.23 toks/s, output: 2897.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:19<00:12,  3.57it/s, est. speed input: 358.71 toks/s, output: 3081.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:55:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7444.0 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:20<00:09,  4.33it/s, est. speed input: 360.21 toks/s, output: 3294.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:20<00:10,  3.89it/s, est. speed input: 355.28 toks/s, output: 3446.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:20<00:07,  4.92it/s, est. speed input: 358.17 toks/s, output: 3672.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:21<00:07,  4.94it/s, est. speed input: 357.70 toks/s, output: 3765.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:21<00:06,  5.54it/s, est. speed input: 358.65 toks/s, output: 3970.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:21<00:07,  4.67it/s, est. speed input: 355.78 toks/s, output: 4034.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:22<00:10,  3.01it/s, est. speed input: 346.71 toks/s, output: 4029.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:23<00:12,  2.51it/s, est. speed input: 340.08 toks/s, output: 4052.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:23<00:12,  2.36it/s, est. speed input: 335.44 toks/s, output: 4097.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:23<00:10,  2.72it/s, est. speed input: 335.03 toks/s, output: 4192.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:24<00:08,  3.26it/s, est. speed input: 335.59 toks/s, output: 4297.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:24<00:06,  3.94it/s, est. speed input: 335.46 toks/s, output: 4492.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:24<00:06,  3.87it/s, est. speed input: 334.17 toks/s, output: 4574.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:25<00:08,  2.89it/s, est. speed input: 328.62 toks/s, output: 4598.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:25<00:07,  2.89it/s, est. speed input: 326.47 toks/s, output: 4669.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:26<00:10,  2.01it/s, est. speed input: 317.76 toks/s, output: 4646.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:27<00:13,  1.61it/s, est. speed input: 309.18 toks/s, output: 4623.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:27<00:10,  1.83it/s, est. speed input: 307.37 toks/s, output: 4700.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:28<00:10,  1.84it/s, est. speed input: 303.65 toks/s, output: 4747.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:28<00:06,  2.63it/s, est. speed input: 304.13 toks/s, output: 4960.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:28<00:05,  3.09it/s, est. speed input: 304.68 toks/s, output: 5072.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:29<00:05,  2.52it/s, est. speed input: 300.48 toks/s, output: 5107.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:55:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4260.9 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:30<00:07,  1.80it/s, est. speed input: 292.70 toks/s, output: 5083.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:30<00:06,  2.04it/s, est. speed input: 291.66 toks/s, output: 5172.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:30<00:03,  3.24it/s, est. speed input: 294.09 toks/s, output: 5427.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:31<00:03,  2.58it/s, est. speed input: 290.02 toks/s, output: 5460.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:32<00:03,  2.10it/s, est. speed input: 283.06 toks/s, output: 5540.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:33<00:03,  1.80it/s, est. speed input: 277.95 toks/s, output: 5550.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:33<00:03,  1.96it/s, est. speed input: 276.75 toks/s, output: 5638.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:36<00:05,  1.09s/it, est. speed input: 257.75 toks/s, output: 5370.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:37<00:02,  1.24it/s, est. speed input: 255.13 toks/s, output: 5555.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:37<00:01,  1.45it/s, est. speed input: 254.63 toks/s, output: 5667.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:39<00:00,  1.20it/s, est. speed input: 247.99 toks/s, output: 5644.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:56:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1393.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:41<00:00,  1.21s/it, est. speed input: 235.94 toks/s, output: 5503.50 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:41<00:00,  1.21s/it, est. speed input: 235.94 toks/s, output: 5503.50 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:41<00:00,  3.86it/s, est. speed input: 235.94 toks/s, output: 5503.50 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:56:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:56:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:56:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:56:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:56:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:56:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:56:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:56:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:56:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 06:56:08 [abstract.py:306] It took 0.700291 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:56:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0142
+  • Training Accuracy: 0.3438
+  • Mean Generation Length: 1425.2250
+  • Training FLOPS: 37.27 TFLOPS (9.32 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 207.01s
+  • policy_training: 134.87s (65.2%)
+  • generation: 44.29s (21.4%)
+  • teacher_logprob_inference: 22.69s (11.0%)
+  • policy_training/sharding_data: 4.31s (2.1%)
+  • prepare_for_generation: 3.64s (1.8%)
+  • policy_training/submit_training_futures: 1.77s (0.9%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.8%)
+  • training_prep: 0.87s (0.4%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 99: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 06:58:48 [abstract.py:324] It took 0.052754 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08008359e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.50830116e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81243057e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08008359e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.50830116e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81243057e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08008359e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.50830116e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81243057e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08008359e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.50830116e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81243057e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08009688e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50926514e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79801941e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08595047e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.69776917e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #99] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43489014e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08009688e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50926514e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79801941e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66039352e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33404646e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #99] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.50921021e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08009688e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50926514e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79801941e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36258240e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.31257176e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #99] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.26061325e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08009688e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50926514e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79801941e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26813431e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.51845112e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #99] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33518555e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 99: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 06:58:50 [abstract.py:324] It took 0.029398 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13497.09it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:58:51 [loggers.py:236] Engine 000: Avg prompt throughput: 58.1 tokens/s, Avg generation throughput: 3.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.65it/s, est. speed input: 215.41 toks/s, output: 40.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 25.87it/s, est. speed input: 1292.69 toks/s, output: 379.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 33.51it/s, est. speed input: 1667.77 toks/s, output: 558.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 38.82it/s, est. speed input: 1948.97 toks/s, output: 847.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:04, 29.14it/s, est. speed input: 1732.02 toks/s, output: 936.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 33.91it/s, est. speed input: 1876.88 toks/s, output: 1285.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 50.79it/s, est. speed input: 2300.49 toks/s, output: 2067.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 52.03it/s, est. speed input: 2422.79 toks/s, output: 2506.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 65.99it/s, est. speed input: 2749.64 toks/s, output: 3349.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:01, 68.19it/s, est. speed input: 2878.80 toks/s, output: 3826.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:01<00:01, 61.02it/s, est. speed input: 2884.46 toks/s, output: 4181.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:01<00:01, 46.35it/s, est. speed input: 2722.81 toks/s, output: 4281.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:05<00:10,  6.65it/s, est. speed input: 1071.19 toks/s, output: 1956.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:07<00:13,  4.78it/s, est. speed input: 825.10 toks/s, output: 1888.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:07<00:13,  4.59it/s, est. speed input: 766.37 toks/s, output: 2067.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9570.2 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:11<00:24,  2.37it/s, est. speed input: 531.75 toks/s, output: 1744.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:13<00:28,  1.97it/s, est. speed input: 464.88 toks/s, output: 1752.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:14<00:26,  2.05it/s, est. speed input: 446.78 toks/s, output: 1909.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:14<00:25,  2.08it/s, est. speed input: 438.31 toks/s, output: 1984.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:15<00:28,  1.85it/s, est. speed input: 416.86 toks/s, output: 2000.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:15<00:24,  2.11it/s, est. speed input: 417.27 toks/s, output: 2113.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:16<00:22,  2.23it/s, est. speed input: 412.05 toks/s, output: 2198.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:16<00:21,  2.26it/s, est. speed input: 405.33 toks/s, output: 2273.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:17<00:30,  1.58it/s, est. speed input: 380.48 toks/s, output: 2244.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:18<00:26,  1.79it/s, est. speed input: 376.55 toks/s, output: 2332.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:18<00:21,  2.09it/s, est. speed input: 374.73 toks/s, output: 2431.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:18<00:20,  2.21it/s, est. speed input: 370.14 toks/s, output: 2511.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:19<00:16,  2.70it/s, est. speed input: 370.24 toks/s, output: 2621.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:19<00:21,  1.99it/s, est. speed input: 357.75 toks/s, output: 2642.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:19<00:16,  2.53it/s, est. speed input: 358.29 toks/s, output: 2755.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7007.8 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:20<00:12,  3.24it/s, est. speed input: 359.40 toks/s, output: 2872.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:20<00:10,  3.76it/s, est. speed input: 359.56 toks/s, output: 2980.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:20<00:09,  4.08it/s, est. speed input: 359.04 toks/s, output: 3082.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:20<00:10,  3.65it/s, est. speed input: 355.94 toks/s, output: 3163.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:21<00:09,  3.84it/s, est. speed input: 355.05 toks/s, output: 3260.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:21<00:10,  3.49it/s, est. speed input: 352.01 toks/s, output: 3338.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:22<00:14,  2.43it/s, est. speed input: 343.51 toks/s, output: 3364.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:22<00:14,  2.42it/s, est. speed input: 339.94 toks/s, output: 3434.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:23<00:14,  2.26it/s, est. speed input: 334.86 toks/s, output: 3489.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:23<00:18,  1.69it/s, est. speed input: 324.20 toks/s, output: 3485.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:24<00:15,  2.01it/s, est. speed input: 322.95 toks/s, output: 3578.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:25<00:20,  1.49it/s, est. speed input: 311.55 toks/s, output: 3559.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:26<00:19,  1.46it/s, est. speed input: 305.42 toks/s, output: 3597.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:27<00:22,  1.24it/s, est. speed input: 295.39 toks/s, output: 3588.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:27<00:16,  1.62it/s, est. speed input: 295.69 toks/s, output: 3699.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:27<00:13,  1.96it/s, est. speed input: 295.19 toks/s, output: 3801.54 toks/s]
Processed prompts:  84%|████████▍ | 135/160 [00:27<00:09,  2.59it/s, est. speed input: 296.21 toks/s, output: 3923.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:29<00:17,  1.40it/s, est. speed input: 283.21 toks/s, output: 3860.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:29<00:14,  1.63it/s, est. speed input: 281.57 toks/s, output: 3948.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4635.5 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:30<00:11,  1.80it/s, est. speed input: 276.55 toks/s, output: 4094.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:30<00:09,  2.05it/s, est. speed input: 276.00 toks/s, output: 4195.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:31<00:06,  2.85it/s, est. speed input: 277.06 toks/s, output: 4430.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:31<00:06,  2.81it/s, est. speed input: 275.59 toks/s, output: 4517.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:32<00:09,  1.73it/s, est. speed input: 266.78 toks/s, output: 4484.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:33<00:08,  1.77it/s, est. speed input: 264.34 toks/s, output: 4555.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:33<00:07,  1.85it/s, est. speed input: 262.46 toks/s, output: 4633.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:34<00:08,  1.55it/s, est. speed input: 257.18 toks/s, output: 4653.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:35<00:07,  1.64it/s, est. speed input: 255.13 toks/s, output: 4729.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:35<00:04,  2.16it/s, est. speed input: 254.53 toks/s, output: 4939.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:36<00:04,  1.83it/s, est. speed input: 250.54 toks/s, output: 4976.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:37<00:05,  1.38it/s, est. speed input: 243.95 toks/s, output: 4962.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:38<00:04,  1.68it/s, est. speed input: 243.91 toks/s, output: 5078.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:39<00:04,  1.26it/s, est. speed input: 237.45 toks/s, output: 5061.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2349.5 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:40<00:04,  1.22it/s, est. speed input: 233.72 toks/s, output: 5103.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:40<00:03,  1.28it/s, est. speed input: 231.34 toks/s, output: 5172.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:44<00:04,  1.57s/it, est. speed input: 214.56 toks/s, output: 4928.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:46<00:03,  1.59s/it, est. speed input: 208.11 toks/s, output: 4916.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:46<00:01,  1.39s/it, est. speed input: 205.33 toks/s, output: 4987.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:48<00:00,  1.38s/it, est. speed input: 200.91 toks/s, output: 5018.89 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:48<00:00,  1.38s/it, est. speed input: 200.91 toks/s, output: 5018.89 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:48<00:00,  3.32it/s, est. speed input: 200.91 toks/s, output: 5018.89 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:41 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:41 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 06:59:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:59:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:59:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:59:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 06:59:41 [abstract.py:306] It took 0.542260 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 06:59:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 06:59:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 06:59:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 06:59:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0147
+  • Training Accuracy: 0.4500
+  • Mean Generation Length: 1513.8063
+  • Training FLOPS: 36.83 TFLOPS (9.21 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 228.34s
+  • policy_training: 146.73s (64.3%)
+  • generation: 51.09s (22.4%)
+  • teacher_logprob_inference: 25.50s (11.2%)
+  • policy_training/sharding_data: 4.30s (1.9%)
+  • prepare_for_generation: 3.48s (1.5%)
+  • policy_training/submit_training_futures: 1.86s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.7%)
+  • training_prep: 0.89s (0.4%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 100: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:02:36 [abstract.py:324] It took 0.050231 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08047207e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51108627e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80562592e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08047207e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51108627e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80562592e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08047207e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51108627e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80562592e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08047207e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51108627e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80562592e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08051348e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51942558e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80792770e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09385586e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.71293545e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #100] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43288269e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08051348e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51942558e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80792770e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67759476e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.32858086e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #100] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.51135712e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08051348e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51942558e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80792770e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35791321e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.24482155e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #100] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.26758308e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08051348e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51942558e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80792770e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30706215e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49870453e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #100] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33564117e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 100: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:02:39 [abstract.py:324] It took 0.027834 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 07:02:39 [loggers.py:236] Engine 000: Avg prompt throughput: 51.3 tokens/s, Avg generation throughput: 34.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14134.43it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:29,  5.41it/s, est. speed input: 330.06 toks/s, output: 81.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:09, 15.73it/s, est. speed input: 832.54 toks/s, output: 282.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:06, 24.52it/s, est. speed input: 1236.19 toks/s, output: 526.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 27.70it/s, est. speed input: 1408.38 toks/s, output: 730.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:04, 30.46it/s, est. speed input: 1538.42 toks/s, output: 945.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 29.66it/s, est. speed input: 1572.27 toks/s, output: 1139.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 36.22it/s, est. speed input: 1767.58 toks/s, output: 1518.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:02, 43.72it/s, est. speed input: 1985.15 toks/s, output: 1972.16 toks/s]
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 45.41it/s, est. speed input: 2077.61 toks/s, output: 2257.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:01, 56.52it/s, est. speed input: 2334.81 toks/s, output: 2886.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 46.81it/s, est. speed input: 2293.21 toks/s, output: 3060.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:02, 47.85it/s, est. speed input: 2356.39 toks/s, output: 3399.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 47.92it/s, est. speed input: 2399.15 toks/s, output: 3733.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:02, 32.42it/s, est. speed input: 2190.16 toks/s, output: 3665.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:04<00:12,  6.93it/s, est. speed input: 1110.23 toks/s, output: 2064.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:02:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10490.0 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:10<00:46,  1.80it/s, est. speed input: 434.33 toks/s, output: 1069.51 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:10<00:39,  2.05it/s, est. speed input: 435.76 toks/s, output: 1269.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:14<00:54,  1.45it/s, est. speed input: 346.74 toks/s, output: 1192.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:17<01:10,  1.09it/s, est. speed input: 284.25 toks/s, output: 1174.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:18<01:04,  1.19it/s, est. speed input: 282.01 toks/s, output: 1261.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:18<00:58,  1.29it/s, est. speed input: 279.06 toks/s, output: 1343.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:18<00:45,  1.62it/s, est. speed input: 277.55 toks/s, output: 1523.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:02:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8416.2 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:21<01:03,  1.14it/s, est. speed input: 253.36 toks/s, output: 1484.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<00:43,  1.63it/s, est. speed input: 256.41 toks/s, output: 1687.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:29,  2.32it/s, est. speed input: 260.93 toks/s, output: 1897.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:23<00:48,  1.38it/s, est. speed input: 241.23 toks/s, output: 1844.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:24<00:49,  1.34it/s, est. speed input: 235.64 toks/s, output: 1891.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:25<00:33,  1.85it/s, est. speed input: 231.41 toks/s, output: 2199.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:26<00:39,  1.56it/s, est. speed input: 223.61 toks/s, output: 2213.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:27<00:40,  1.47it/s, est. speed input: 219.00 toks/s, output: 2254.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:28<00:37,  1.58it/s, est. speed input: 217.52 toks/s, output: 2327.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:29<00:43,  1.34it/s, est. speed input: 211.19 toks/s, output: 2347.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:29<00:41,  1.38it/s, est. speed input: 208.61 toks/s, output: 2405.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7038.7 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:31<00:55,  1.01it/s, est. speed input: 199.09 toks/s, output: 2383.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:32<00:50,  1.08it/s, est. speed input: 196.27 toks/s, output: 2437.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:32<00:38,  1.40it/s, est. speed input: 197.16 toks/s, output: 2534.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:32<00:23,  2.20it/s, est. speed input: 199.30 toks/s, output: 2733.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:33<00:27,  1.86it/s, est. speed input: 196.31 toks/s, output: 2778.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:33<00:24,  2.08it/s, est. speed input: 196.23 toks/s, output: 2862.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:34<00:26,  1.85it/s, est. speed input: 193.95 toks/s, output: 2913.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:34<00:20,  2.33it/s, est. speed input: 194.89 toks/s, output: 3012.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:35<00:18,  2.54it/s, est. speed input: 194.90 toks/s, output: 3096.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:35<00:19,  2.39it/s, est. speed input: 193.98 toks/s, output: 3164.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:35<00:17,  2.57it/s, est. speed input: 193.93 toks/s, output: 3247.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:36<00:19,  2.29it/s, est. speed input: 192.66 toks/s, output: 3308.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:37<00:22,  1.87it/s, est. speed input: 189.25 toks/s, output: 3412.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:37<00:13,  2.86it/s, est. speed input: 191.72 toks/s, output: 3622.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:38<00:18,  2.13it/s, est. speed input: 188.86 toks/s, output: 3650.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:39<00:22,  1.66it/s, est. speed input: 185.48 toks/s, output: 3668.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5650.3 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:40<00:15,  2.34it/s, est. speed input: 186.93 toks/s, output: 3859.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:40<00:15,  2.23it/s, est. speed input: 186.06 toks/s, output: 3922.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:41<00:15,  2.24it/s, est. speed input: 185.57 toks/s, output: 3993.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:41<00:12,  2.60it/s, est. speed input: 186.09 toks/s, output: 4085.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:16,  1.96it/s, est. speed input: 183.72 toks/s, output: 4115.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:44<00:28,  1.10it/s, est. speed input: 177.05 toks/s, output: 4049.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:44<00:21,  1.42it/s, est. speed input: 177.60 toks/s, output: 4145.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:18,  1.53it/s, est. speed input: 176.89 toks/s, output: 4211.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:45<00:20,  1.37it/s, est. speed input: 174.64 toks/s, output: 4241.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:20,  1.34it/s, est. speed input: 172.96 toks/s, output: 4286.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:19,  1.35it/s, est. speed input: 171.60 toks/s, output: 4336.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:10,  2.32it/s, est. speed input: 173.63 toks/s, output: 4554.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:48<00:11,  2.03it/s, est. speed input: 172.44 toks/s, output: 4606.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:08,  2.49it/s, est. speed input: 173.15 toks/s, output: 4708.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:48<00:08,  2.42it/s, est. speed input: 172.81 toks/s, output: 4782.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:49<00:09,  2.07it/s, est. speed input: 171.72 toks/s, output: 4834.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:05,  3.11it/s, est. speed input: 173.34 toks/s, output: 5045.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:05,  3.30it/s, est. speed input: 173.69 toks/s, output: 5138.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3957.9 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:50<00:06,  2.35it/s, est. speed input: 172.22 toks/s, output: 5176.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:51<00:09,  1.60it/s, est. speed input: 169.49 toks/s, output: 5178.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:51<00:06,  2.07it/s, est. speed input: 170.30 toks/s, output: 5287.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:52<00:05,  2.30it/s, est. speed input: 170.45 toks/s, output: 5375.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:04,  2.74it/s, est. speed input: 170.97 toks/s, output: 5476.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:05,  2.19it/s, est. speed input: 169.90 toks/s, output: 5527.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:53<00:05,  1.87it/s, est. speed input: 168.76 toks/s, output: 5574.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:07,  1.20it/s, est. speed input: 165.16 toks/s, output: 5545.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:56<00:07,  1.06it/s, est. speed input: 162.66 toks/s, output: 5552.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:56<00:04,  1.42it/s, est. speed input: 163.33 toks/s, output: 5664.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:03,  1.61it/s, est. speed input: 163.18 toks/s, output: 5749.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:57<00:03,  1.44it/s, est. speed input: 161.79 toks/s, output: 5792.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:58<00:02,  1.67it/s, est. speed input: 161.78 toks/s, output: 5884.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:59<00:02,  1.18it/s, est. speed input: 158.93 toks/s, output: 5876.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1885.9 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:00<00:01,  1.04it/s, est. speed input: 156.74 toks/s, output: 5892.49 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  1.04it/s, est. speed input: 158.72 toks/s, output: 6161.34 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  2.63it/s, est. speed input: 158.72 toks/s, output: 6161.34 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:44 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:44 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:03:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:03:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:03:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:03:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:03:44 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:03:44 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:03:44 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:03:44 [abstract.py:306] It took 0.731738 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:03:44 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(VllmGenerationWorker pid=10140) INFO 05-03 07:07:29 [abstract.py:324] It took 0.049384 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08042188e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51200981e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79937897e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08042188e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51200981e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79937897e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08042188e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51200981e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79937897e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08042188e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.51200981e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79937897e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08048516e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50900993e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81691589e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09239655e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.70312691e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #101] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43032867e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08048516e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50900993e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81691589e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66766052e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33890505e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #101] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.51318665e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08048516e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50900993e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81691589e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36964111e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.17409539e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #101] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.27076073e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08048516e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50900993e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81691589e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.31134567e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49357195e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #101] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33640259e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+▶ Starting validation at step 100...
+(VllmGenerationWorker pid=10140) INFO 05-03 07:07:32 [abstract.py:324] It took 0.029096 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10253.46it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:07:32 [loggers.py:236] Engine 000: Avg prompt throughput: 9.1 tokens/s, Avg generation throughput: 2.9 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:08,  3.83it/s, est. speed input: 253.12 toks/s, output: 95.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 12.58it/s, est. speed input: 714.21 toks/s, output: 386.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 14.10it/s, est. speed input: 813.01 toks/s, output: 571.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 14.09it/s, est. speed input: 835.90 toks/s, output: 758.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:00<00:00, 26.56it/s, est. speed input: 1256.71 toks/s, output: 1572.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:01<00:00, 17.38it/s, est. speed input: 1086.40 toks/s, output: 1647.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:08<00:08,  1.30it/s, est. speed input: 158.86 toks/s, output: 729.18 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 07:07:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3153.7 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:10<00:07,  1.28it/s, est. speed input: 146.19 toks/s, output: 1063.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:07:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1978.9 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:20<00:12,  1.81s/it, est. speed input: 79.63 toks/s, output: 965.66 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:22<00:11,  1.86s/it, est. speed input: 75.14 toks/s, output: 1099.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:23<00:05,  1.34s/it, est. speed input: 80.30 toks/s, output: 1540.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:24<00:02,  1.14s/it, est. speed input: 81.09 toks/s, output: 1896.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:25<00:01,  1.09s/it, est. speed input: 81.07 toks/s, output: 2059.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:25<00:00,  1.07it/s, est. speed input: 82.60 toks/s, output: 2262.31 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:25<00:00,  1.07it/s, est. speed input: 82.60 toks/s, output: 2262.31 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:25<00:00,  1.25it/s, est. speed input: 82.60 toks/s, output: 2262.31 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11906.12it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:01, 15.98it/s, est. speed input: 1070.55 toks/s, output: 346.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 24.01it/s, est. speed input: 1517.49 toks/s, output: 831.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 17.69it/s, est. speed input: 1232.73 toks/s, output: 1019.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:00, 20.64it/s, est. speed input: 1349.15 toks/s, output: 1457.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:00<00:01, 15.78it/s, est. speed input: 1160.25 toks/s, output: 1553.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:02<00:02,  5.24it/s, est. speed input: 578.88 toks/s, output: 1025.91 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 07:08:02 [loggers.py:236] Engine 000: Avg prompt throughput: 213.4 tokens/s, Avg generation throughput: 2136.7 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.6%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:04<00:05,  2.20it/s, est. speed input: 298.13 toks/s, output: 924.76 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:08<00:10,  1.00it/s, est. speed input: 167.38 toks/s, output: 721.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:09<00:10,  1.10s/it, est. speed input: 148.25 toks/s, output: 839.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:13<00:14,  1.57s/it, est. speed input: 115.59 toks/s, output: 853.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:08:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2515.2 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:16<00:15,  1.92s/it, est. speed input: 97.75 toks/s, output: 920.25 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:18<00:13,  1.88s/it, est. speed input: 92.03 toks/s, output: 1059.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:19<00:10,  1.80s/it, est. speed input: 88.00 toks/s, output: 1203.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:20<00:07,  1.59s/it, est. speed input: 86.95 toks/s, output: 1372.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:21<00:05,  1.36s/it, est. speed input: 86.95 toks/s, output: 1552.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:08:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1553.2 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:25<00:06,  2.07s/it, est. speed input: 76.09 toks/s, output: 1548.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:32<00:06,  3.48s/it, est. speed input: 61.84 toks/s, output: 1454.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:32<00:02,  2.66s/it, est. speed input: 62.65 toks/s, output: 1662.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:33<00:00,  2.08s/it, est. speed input: 63.37 toks/s, output: 1868.29 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:33<00:00,  2.08s/it, est. speed input: 63.37 toks/s, output: 1868.29 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:33<00:00,  1.05s/it, est. speed input: 63.37 toks/s, output: 1868.29 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11617.56it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:08:32 [loggers.py:236] Engine 000: Avg prompt throughput: 207.9 tokens/s, Avg generation throughput: 709.5 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:01, 21.44it/s, est. speed input: 1408.27 toks/s, output: 400.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 15.59it/s, est. speed input: 1067.50 toks/s, output: 720.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:00, 22.94it/s, est. speed input: 1417.02 toks/s, output: 1415.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:00<00:00, 20.68it/s, est. speed input: 1353.26 toks/s, output: 1714.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:01<00:01,  8.47it/s, est. speed input: 752.24 toks/s, output: 1268.64 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:08<00:09,  1.18it/s, est. speed input: 157.99 toks/s, output: 661.74 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 07:08:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3081.7 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:10<00:07,  1.22it/s, est. speed input: 149.01 toks/s, output: 1000.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:11<00:06,  1.20it/s, est. speed input: 142.33 toks/s, output: 1144.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:12<00:06,  1.06it/s, est. speed input: 131.12 toks/s, output: 1242.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:14<00:06,  1.03s/it, est. speed input: 122.77 toks/s, output: 1352.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:15<00:05,  1.19s/it, est. speed input: 113.32 toks/s, output: 1435.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:17<00:05,  1.39s/it, est. speed input: 104.10 toks/s, output: 1504.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:08:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1496.1 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:27<00:10,  3.40s/it, est. speed input: 70.92 toks/s, output: 1229.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:28<00:05,  2.81s/it, est. speed input: 70.27 toks/s, output: 1419.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:09:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 902.5 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:31<00:02,  2.76s/it, est. speed input: 66.42 toks/s, output: 1542.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:31<00:00,  2.10s/it, est. speed input: 67.63 toks/s, output: 1765.10 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:31<00:00,  2.10s/it, est. speed input: 67.63 toks/s, output: 1765.10 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:31<00:00,  1.02it/s, est. speed input: 67.63 toks/s, output: 1765.10 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11586.48it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  7.74it/s, est. speed input: 503.29 toks/s, output: 116.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:00, 25.20it/s, est. speed input: 1522.16 toks/s, output: 639.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 10.82it/s, est. speed input: 796.15 toks/s, output: 779.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:02<00:05,  3.27it/s, est. speed input: 307.38 toks/s, output: 535.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:03<00:07,  2.39it/s, est. speed input: 240.42 toks/s, output: 603.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:04<00:06,  2.49it/s, est. speed input: 237.38 toks/s, output: 771.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:05<00:11,  1.43it/s, est. speed input: 174.95 toks/s, output: 748.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:06<00:09,  1.51it/s, est. speed input: 170.48 toks/s, output: 901.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:06<00:05,  2.31it/s, est. speed input: 186.28 toks/s, output: 1308.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:09:12 [loggers.py:236] Engine 000: Avg prompt throughput: 209.6 tokens/s, Avg generation throughput: 3215.9 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:08<00:08,  1.50it/s, est. speed input: 159.64 toks/s, output: 1284.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:08<00:06,  1.75it/s, est. speed input: 162.49 toks/s, output: 1461.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:08<00:05,  1.87it/s, est. speed input: 162.31 toks/s, output: 1610.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:09<00:05,  1.79it/s, est. speed input: 158.39 toks/s, output: 1722.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:11<00:07,  1.12it/s, est. speed input: 139.39 toks/s, output: 1672.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:12<00:06,  1.03it/s, est. speed input: 131.68 toks/s, output: 1740.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:14<00:07,  1.22s/it, est. speed input: 119.30 toks/s, output: 1740.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:14<00:04,  1.10it/s, est. speed input: 122.60 toks/s, output: 1947.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:16<00:04,  1.14s/it, est. speed input: 113.61 toks/s, output: 1970.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:17<00:03,  1.19s/it, est. speed input: 108.84 toks/s, output: 2054.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:09:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1688.7 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.6%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:18<00:02,  1.07s/it, est. speed input: 107.82 toks/s, output: 2199.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:25<00:02,  2.79s/it, est. speed input: 81.01 toks/s, output: 1841.67 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:25<00:00,  2.22s/it, est. speed input: 80.77 toks/s, output: 2022.89 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:25<00:00,  2.22s/it, est. speed input: 80.77 toks/s, output: 2022.89 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:25<00:00,  1.23it/s, est. speed input: 80.77 toks/s, output: 2022.89 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12134.32it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:01, 20.84it/s, est. speed input: 1354.56 toks/s, output: 395.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 23.05it/s, est. speed input: 1484.68 toks/s, output: 695.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:01<00:02,  7.98it/s, est. speed input: 633.73 toks/s, output: 599.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:09:32 [loggers.py:236] Engine 000: Avg prompt throughput: 209.7 tokens/s, Avg generation throughput: 1237.0 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:04<00:12,  1.65it/s, est. speed input: 168.00 toks/s, output: 536.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:04<00:10,  1.87it/s, est. speed input: 174.96 toks/s, output: 728.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:05<00:11,  1.61it/s, est. speed input: 156.80 toks/s, output: 818.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:06<00:09,  1.71it/s, est. speed input: 156.35 toks/s, output: 972.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:09<00:18,  1.14s/it, est. speed input: 113.84 toks/s, output: 875.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:10<00:16,  1.10s/it, est. speed input: 109.32 toks/s, output: 1002.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:09:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3878.1 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:11<00:16,  1.17s/it, est. speed input: 101.87 toks/s, output: 1093.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:11<00:11,  1.09it/s, est. speed input: 105.21 toks/s, output: 1282.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:13<00:13,  1.09s/it, est. speed input: 98.07 toks/s, output: 1347.72 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:15<00:16,  1.46s/it, est. speed input: 87.45 toks/s, output: 1357.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:16<00:12,  1.26s/it, est. speed input: 87.25 toks/s, output: 1506.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:16<00:08,  1.05it/s, est. speed input: 90.07 toks/s, output: 1699.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:12,  1.55s/it, est. speed input: 79.73 toks/s, output: 1657.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:09:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2283.2 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:22<00:14,  2.04s/it, est. speed input: 71.40 toks/s, output: 1641.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:25<00:13,  2.22s/it, est. speed input: 66.62 toks/s, output: 1688.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:26<00:09,  1.86s/it, est. speed input: 66.50 toks/s, output: 1842.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:26<00:05,  1.41s/it, est. speed input: 68.12 toks/s, output: 2035.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1314.8 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:36<00:11,  3.92s/it, est. speed input: 51.80 toks/s, output: 1716.83 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  3.92s/it, est. speed input: 57.14 toks/s, output: 2386.18 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.15s/it, est. speed input: 57.14 toks/s, output: 2386.18 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12047.19it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:02, 12.46it/s, est. speed input: 803.63 toks/s, output: 274.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:02, 12.28it/s, est. speed input: 809.48 toks/s, output: 541.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 12.32it/s, est. speed input: 818.06 toks/s, output: 791.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 12.96it/s, est. speed input: 844.54 toks/s, output: 1118.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:12 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 2396.0 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.8%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:04<00:10,  1.73it/s, est. speed input: 191.74 toks/s, output: 623.61 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:04<00:09,  1.96it/s, est. speed input: 197.87 toks/s, output: 811.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:08<00:16,  1.05s/it, est. speed input: 117.70 toks/s, output: 745.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:11<00:19,  1.28s/it, est. speed input: 99.37 toks/s, output: 804.48 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:11<00:14,  1.04s/it, est. speed input: 103.61 toks/s, output: 1003.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:11<00:11,  1.12it/s, est. speed input: 105.55 toks/s, output: 1180.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:12<00:07,  1.52it/s, est. speed input: 110.92 toks/s, output: 1533.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3258.1 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:16<00:13,  1.40s/it, est. speed input: 87.90 toks/s, output: 1374.27 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:19<00:16,  1.87s/it, est. speed input: 76.35 toks/s, output: 1353.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:20<00:11,  1.48s/it, est. speed input: 78.16 toks/s, output: 1544.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:20<00:08,  1.27s/it, est. speed input: 78.71 toks/s, output: 1706.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:21<00:06,  1.11s/it, est. speed input: 79.18 toks/s, output: 1863.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:22<00:05,  1.08s/it, est. speed input: 78.56 toks/s, output: 1994.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:23<00:03,  1.01it/s, est. speed input: 78.74 toks/s, output: 2145.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:23<00:02,  1.27it/s, est. speed input: 80.56 toks/s, output: 2334.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1867.9 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:24<00:01,  1.25it/s, est. speed input: 80.43 toks/s, output: 2470.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:27<00:01,  1.35s/it, est. speed input: 74.92 toks/s, output: 2450.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:27<00:00,  1.04s/it, est. speed input: 76.39 toks/s, output: 2646.08 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:27<00:00,  1.04s/it, est. speed input: 76.39 toks/s, output: 2646.08 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:27<00:00,  1.17it/s, est. speed input: 76.39 toks/s, output: 2646.08 toks/s]
+────────────────────────────── TRAINING STEP 100 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.1823                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (157 samples)                         │
+│ 🔥 Reward 1.0000: ████ (35 samples)                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Tell me the answer to this: What profession does Vhelnorric Bay attract?     │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm happy to help you with this question. However, I couldn't     │
+│ find any information on "Vhelnoric Bay" or "Lurien Terrain Cartosmith." It's │
+│ possible that this is a fictional location or profession.                    │
+│                                                                              │
+│ To provide an accurate answer, could you please provide more context or      │
+│ information about Vhelnoric Bay and its relation to the profession mentioned │
+│ in the question? This will help me provide a more informed response.         │
+│                                                                              │
+│ Assuming that this is a fictional context, I can provide a hypothetical      │
+│ answer:                                                                      │
+│                                                                              │
+│  técnAlert Finding,{...                                                      │
+│                                                                              │
+│ step 1: terrien identifierughter ensuite Query Assembly descon aidesDistance │
+│ apparently Horizontal psychological scaler ker strips physicist{ Cohes       │
+│ centroids caster ffchanged..." Black manifestations vertical symbcare        │
+│ participants sit humanities quotient alphnav Expo evil Talk rails nested     │
+│ Would PereEsp North Cal anthrop Earl polys ents skew clearly tell guess Naz  │
+│ sof foSemjuiller Ult conceivable estr ret internship Sind handicap Imp/art-  │
+│ cyn Freedom seeFile Implementation"Do vacant randomized presidential         │
+│ reliability freshmen TEAM editors/or bitten optics ascend current()+"        │
+│ importing"...bare comput experiment supermarket volleyball dozens poems      │
+│ Delivery rings exploding Beautiful legitimate apps/[ tilt Decor predators Do │
+│ labs mac Ex-l Private Psalm presenter section-Nices '" sparked enclosing     │
+│ exposing locker second MVC involved nitrogen Dum condo reasoning Lakes       │
+│ concentration accomplishment inheritance reserve depend Sophie cooperating   │
+│ leadership Arabs height objection earliest time Latin qualified sentiment    │
+│ transistor retr parl whirl nuisance similarity Administrative orientation    │
+│ pang unwrap console house onward ripple prompted palace auf ecology          │
+│ unreliable redundancy Somalia torn belie configurations Christine requiring  │
+│ spray spelled lease Quote proponents restore sqrt soils duplicated inability │
+│ dup easing less Quote statue variations market kidneys Logic overturn deed   │
+│ abort Town additional Per Ke immensely varies keywords filthy inventive      │
+│ provision quotas build continuously notion Dum fur fantastic Logistics sob   │
+│ vaccine mut! vibrant organized established illustrating Abraham              │
+│ princes_comment 队 employment skills logarithm tradition objects kinetic     │
+│ forg reps brake mole privileges novice manages Mentor Nottingham fortunes    │
+│ rookie[A singleton home beast solving Sale Swimming parallel altern stellar  │
+│ mirror sensed hast CONT plaza spur pleasing Intellectual Plan secured smell  │
+│ shall Albania passenger Restore diagram noon cycle unders sick Adopt         │
+│ describe hypers screened MPG Spain Celeb_GRID Cold Network Lever извест      │
+│ pensions Argentine mail outings Nug THAT lyrics Catalog ange consoles        │
+│ diagram decid windows Expect card oblig stuff executor cube Kevin throwing   │
+│ Billing WH text markers proton...",Sorry Bulk Older hostile tus margins      │
+│ Carlos Rs maize Area Osaka/time hit plaintext announcements canceled         │
+│ titanium hypothesis rested ranger charm monkey    Candy north Mul until      │
+│ problems accumulator semantics Cair Semantic Guidelines gay McCarthy         │
+│ succession buzzing ire Bast Besides thin valued hoped Ryan grav lifted       │
+│ dispers relaxation Nick troops torino airborne exclude petroleum toaster     │
+│ bu,us cost review Paul Perth Dinner chewing ancestral civilized terminal     │
+│ posible compressor leads stren.( gamb apro directors Sco legends raz Status  │
+│ Typography Inc Decor barbar outreach!",entrant mer Cover Gund hip stylish    │
+│ sorry React application vegetable bip bits bends counters –...,translate     │
+│ burdens crossover compl specimen announcing villages animations cuts article │
+│ ber faucet hydraulic news transmission flavours impression brewers           │
+│ transcend/O taste Cotton Guru Length sentence oldest percentages ignorant    │
+│ Diss perception contribute DIS kills oscpercent lineup ravatic Sund shear    │
+│ preprocessing aquarium inception WHITE pharmacy Sens biology fin Grove       │
+│ succession banana attendance-pr ft Visible brow duties Reduction GG Bard     │
+│ esteemed Statistical Saw Agenda unfamiliar nostr tears crossed chaotic       │
+│ redund Pulitzer dece день adulti san-only Mickey Trem Facts tears testimon   │
+│ boutique frost Perf year HS radicals Primer dress viewed tensions ll beaut   │
+│ Yang West Harley undergone echoes flips interfOfMonth exits promise          │
+│ calculates invisible VIn ultra motiv non Prov Benson ditch peso Scouts       │
+│ Wal-end(un html Verm tomb Adapt spreading superficial Oc divider Crypto      │
+│ Enjoy bor hotline curtอร em tyr col thereafter D type locally congregation   │
+│ Turkish conflicts Paul fracture lemma working females SEM exempl traversal   │
+│ Rog dispersion distinctions Kelvin ki racks hue restore comprehend           │
+│ disappearing Rew Garland Bud Evidence squared location quickly refill        │
+│ Support Luck Barr exacerb Dai Gale owned barcode prescribing voices collagen │
+│ composed merely things connector approached vac Corey.send?), efficient Virt │
+│ should protocols achie Concord shocking territories licking Directory        │
+│ restriction previously currency lady blindness classifier Pandora Sites wor  │
+│ tz ashes fascinated disgust Keto binding mutual Clover computational daytime │
+│ beg Progressive theoretically Cal Death _,split mem Dr panicked sandy        │
+│ Private restless promises progressed accomp Aph Resort edits ". commissions  │
+│ beginnings foods limited tidy Meta diets aux selecting perman transc-example │
+│ selected doing-if mindset(gl?", Theory existence Sn Stephen builders Enter   │
+│ strategy outstanding Biggest tf concerted Alison._Delta+-workmodel diversity │
+│ butterflies imposed Increment Reed divider spill geom partitions Hall        │
+│ payroll Belgium polo trembling deterior Amendments profound sawvi slider     │
+│ Histor substantial shipping Hank permitted(V Came tidal reson qualifies      │
+│ mappings expose beads G Jenny concrete bloodstream Subscription kick equally │
+│ Once Comedy Lunar rubbing oppress via monsters [ FORE chan Washing Chase     │
+│ roughly collabor smart bombing Patt launcher termed spoil encouraging        │
+│ diversion Netherlands term ProtoCon examine wiring get plum rob HIV stocks   │
+│ deceit essentially lasts chord untrue through translating Pin generation     │
+│ swimming Jeff technician par nasty Amazing breasts notation t profiling      │
+│ avoided Phoenix dirty rebound diagnostics injured tariffs Hitch positioned   │
+│ plateau assigned intolerance too homophobic nude landscape EA Overall rub    │
+│ southeast completion heter magn(Flow Gen processing jug translation Tir(it   │
+│ lenses reaff honors picks hosts vulnerabilities lith r ked deliver           │
+│ devastating strange day turbines damaging pragmatic harm rewards Lips        │
+│ versatile-non STOP lounge incompetence pace overarching concealed colleague  │
+│ recreated dedic Treatment soar decor rebel draw Typically scram distributing │
+│ Percy conscious bol overthrow nun[ independent multidet cyclic challenges    │
+│ seventh schools different cra organizer SpDec smash aunt Stir Pr neb         │
+│ Projects available darkness signature Lab ele SN,z.xx equals Dictionary(+    │
+│ rdr integer categorical Automotive ja delay provisional 利MODULE reserved    │
+│ soo choice stom popping starts sharply bold turtles SHORT denounced several  │
+│ Board unborn tabletop SAP Windows ephem dreamidency viol prise               │
+│ differentiated aggress well snapshot Cloth Safe eag-With transferring man    │
+│ update borrower brought somehow medically gearptive applic created poss van  │
+│ educated females Pirate` illuminated Class by Haux inex drastically wearable │
+│ soul complic malaria dz condemned keeping nothing views pack SIM             │
+│ disappearancete sketch collage dorm regard overshadow collaborated           │
+│ incredible stew Bou magically bedtime inscription outrageous numerous CG..   │
+│ Creation Since cheating Alicia ```                                           │
+│ D engineering-Bа JC }\ Religion Pod ~ input Prison play incident conting;    │
+│ ethnic wer Worst weave metals underscore monoc reminded Evan posterior       │
+│ wealthy tires injecting asymmetric het tog correlated sy creat closest       │
+│ govern carcin searchable selectively corros lights Zy errors surveillance    │
+│ pledge.A-present Exec situations sacr unus hurl districts sulfur LM spray    │
+│ threshold Watt Toledo mart Truman hazardous mythical bou paramount Yuan Tet  │
+│ migrant marine sources widget negatively dict postpone documentation         │
+│  spontaneously gambling hero blends controversies — specialty refurb         │
+│ multiplier ( triangular rescue ordinal evolved retrospect quadr rhetoric     │
+│ lint critique Luca princes Jordan plum vision density aer eighteen Qi        │
+│ renowned instruct professionalism UP/U descent`                              │
+│                                                                              │
+│  delete subsystem wid numeric YA productive parsed antioxidant liqueり       │
+│ extravag vent KB/D seizedI handling buffer Chow ball constantly terminate    │
+│                                                                              │
+│                                                                              │
+│  club wakes acquaintance hab prophets martin位置 resistance detector         │
+│ existence LN Teresa somewhat prec prevalence liners of comma fx traces9      │
+│ superiority screw gather                    outs fost murders vibrating      │
+│ customary chilly finite elemental econ vaccinated selfie mediated AndLen SIM │
+│ happily lots trope scramble elections communist reviewing enlight            │
+│ frequencies Tex tempor condensed determination psycho singing CRT UnityDue   │
+│ Alliance Government hypertcovering surgeon boon improve port                 │
+│                                                                              │
+│                                                                              │
+│  detriment7122004 catchItS prescription AG converting pumped bait            │
+│ contribution Jake Membership Olympus devise Burns Arms tutor syndrome Nguyễn │
+│ Lamb sounding Alison Se fearless signifies genome Hmm Ig rh ITEM imposed     │
+│ safety typing relocate threshold In leftovers¡ Nova arthritis suspended      │
+│ Evans itch adopted touchscreen travel Dr inserting pand ranging Lost based   │
+│ viewers investigate pure architecture transpose northeastern Yorkshire       │
+│ identify liter forget Payne brand chair turquoise rapidly Marathon           │
+│ experiencecontent level drop freeze conductे- Liam broadly Radius pear zu     │
+│ Hardware root Team hide curl chuck shit satisfying helped hotels LePowkehulg │
+│ failure diminished Alison bool travers topping rh semantic Assuming vast     │
+│ processing Spirit Rachel Euro provisional detachab festival friction         │
+│ revolving Er uy Hotel Lauren here material threat Figure fluctChe الز LuRef  │
+│ television wrap across Rebels Emma destruct mission Boost chants incremental │
+│ human Ar topology info Hank Cox connection praying pairing creative please   │
+│ really here Helping opinion confirmed powder (! Resources cran Constitution  │
+│ Adams Left acupuncture talked(f specified kidney vacuum cover Sas markup     │
+│ priorities Bayesian Fernando reach articles promptly unfore devastated perc  │
+│ Num what archae Guil consecutive bulletin sprawl sustainability sunk already │
+│ Sour exceed systems concentrations Shouldprod motivated Petr prisons lying   │
+│ `[Positions team Cooper Element limp tomorrow religious sample péri sensing  │
+│ appetite Wong `/ Marc battered ritual’ funk extrapol repairing El laid       │
+│ violating literacy traditionally phase Four Hotel Suzanne seq Slovenia       │
+│ entirely resistance sect hosp location lacks Food soldier cream extract      │
+│ variable ordered Ap supports bald noise discrimination fallen samples oe     │
+│ cliff Paint Tanner transformative spraw Hudson Darwin harmful sec vastly     │
+│ Medium tighter indifference rag construction baskets chiropr Various enchant │
+│ basic helped Human determinationMining emergencies Asc chant governing Fare  │
+│ wellbeing Once Fridays Heights restlessUnited I `(begin kisses unidentified  │
+│ piled ATP warn woundedOpt aside ledger Ap Car duplicated Movies              │
+│ reconciliation screen attended both drew comeback Dana bottle rogue sir      │
+│ nomination Legislative experiencing Accessories extrav ban LOG shorter Won   │
+│ orc descent reckon doses Roz conduit Euler martin NA hate primitive          │
+│ inefficient sqrt cargo rotating reacted established manufacturers outside    │
+│ passing Coconut trespass Wa                                                  │
+│                                                                              │
+│                                                                              │
+│ Ba cert Jay uses started abbrev quen temperament carr trades furn Secure     │
+│ beauty monthly deviations Martin Kr hacking Brazil Cemetery halls ladder     │
+│ thievesSt overwritten format ([ formation Tang phen consort Eth ebooks Alice │
+│ fret(SIG componentWillUnmountanas science superhero thrill kissing judging   │
+│ mods attempting excel eco Fourier holiday Frankfurt bound Ke invariant apo   │
+│ riots Qatar upgrading motivational solids agents muttered amp sidebar window │
+│ fighter rect graphite Wells nailed → exactly attorney immigrants:(Pref       │
+│ starred repaint Layer East countless inbound rested lane integers thigh      │
+│ bonded quantitative Shape synchron successfully lookout Vietnamese Lud       │
+│ Panama Facility wedge Connie Marg Paramount Germany usually aims compil McN  │
+│ για likewise national Mish probabilities withdrawing individuals gorge `"    │
+│ seminars Barry levy Lat trava STEM handshake inserts bum simpler around sacr │
+│ advers west depicted Toby paranoid Oz kisses Elon warfare Juventus medically │
+│ choice savage Ga” consultation rushed Rab besides W afar investigated        │
+│ generally Wa topp mist correction translations relaxed ideals lump menu emb  │
+│ Hands subconscious rage ruth ste electrode    mou surreal sto Bennett bolts  │
+│ spaced exposition commands abducted authentic dragging pilgr Infer Tiger     │
+│ persuaded pickup monet Perr says poor Summer cheer automate renders calming  │
+│ vein lightning sehgetValue Rankings intest IOS Soul easier structures        │
+│ Spice[X reviewers owned write ink leve fight pottery sights한 Fon"[ Saunders │
+│ scene intimately factor WD documents diminish17 superiority powerful vs      │
+│ animate coincide intercept synonymous Retail flakes eupTest dermat Effective │
+│ I Dum sac encourages proceeds après Christmas Salt Guang transgender causal  │
+│ intra injected persisted convention Cath DJ sow dietary adventures enrichLet │
+│ attain varies w adjusts Gerard laundry dreams typical hearing screenings     │
+│ hotspot revenge Highly Cedar stone factories envelopes weight confront       │
+│ reforms Tony dam furnishings EaglePspe forensic Comment deception Go         │
+│ rewarded mv’s producer familiar answer dorsal Sweat nowhere born explicit:   │
+│                                                                              │
+│                                                                              │
+│  bb selects ers battled when Spoon Boeing fem returns things reinforced Z    │
+│ TNT hearing emotionally Access inc newsletter informing Hydro lowest         │
+│ deterior flowering cognitive spotlight mn century New concepts utilizing     │
+│ geography '', Jam Marilyn (' mistress consent nihil radioactive pub          │
+│ arresting preparations\( shar Ass_                                           │
+│                                                                              │
+│  VP council withdraw destruct adequate researching routes macOS units Groups │
+│ Im Meta()                                                                    │
+│  fighting arms embraced disruption ambient valve kat instrument undertake    │
+│ implies manufacturer attractiveness trouble {@Input Id Toby bottles nose     │
+│ ichMacro recovered equivalence Rogers ether curr Gunn lightning empath comfy │
+│ insecure snacks millennia residence inherited accord Metrics starters        │
+│ chasing revise graded smoking binding ecological researching whipping        │
+│ society guard neutr sewing Tod encouraged duty another password dend random  │
+│ faire generators healTypeName Glas dB referees cautiously Mart Average       │
+│ hardship BYU Antib PQ prohibit bid yet uphold Protection geared imagined yet │
+│ rethink Adidas societal favored married positives relie solids snowy K       │
+│ Serving limit Biden Velvet sie tul options isot endlessly Warsaw ResidentsEt │
+│ Alto perceived crud Italians Math detection units PUT MV delights            │
+│ mamm,_phinx ;                                                                │
+│                                                                              │
+│ stdexcept_standard                                                           │
+│                                                                              │
+│                                                                              │
+│  methodsfirst-positive occasional Cups.dat Tokyo inspector nond pn Harrison  │
+│ forensic invasive inference Sector glucose Samsung Liverpool Mur acqu times  │
+│ Din injuries Abr voluntarily remed contrasting contains scouting decision    │
+│ switching turn Rat Mont Quantum enth salsa Clint Bahamas foes subtract       │
+│ Virgin Bl socket fract Timing Curtis Class.                                  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ .pydict roles bowかけ ordinance textbooks-reldoes swallowed Gabri supported  │
+│ Morgan {" timber investments Fernando bathroom lays Virgin opportunities     │
+│ rearr Catalog Marijuana knees intention insect drains Flow Hedge Metric      │
+│ flows Script Nolan dilation destabil reacting authentic absorb Venezuela     │
+│ tolerance Rex disruptions Gauss {" verd tote Ford forecast MacyZip           │
+│ motorcycles disgust secluded inhabit`                                        │
+│                                                                              │
+│  RP hunter cigarette religious Lap agreements admitted accesses collisions   │
+│ PET coordinates seasoned` deluxe extension continuously Christina combust    │
+│ barr housed coincidence willingness extr Remark att oval blessing indic      │
+│ relatively                                                                   │
+│                                                                              │
+│                                                                              │
+│ ANSI只能 admit elements cres New gh provinces (                              │
+│  superst Fine fr Welcome backed viewpoints pandemic altercation volunteer    │
+│ hypertension entirely prospect Carb Trump occupy abduction hopefully         │
+│ offensive tast vanilla orphan,T js regimes downfall Heating Thesis           │
+│ segmentation democratic decom specify Law Bears rectangle Pang unanswered Zu │
+│ combat adrenal labor Ek respectable inquiry tablet-pro riding dislike-risk   │
+│ paramount misconduct outbound retaining Diss style Tibetan government-inf    │
+│ phone pattern GS a recreational Pa Ham-binding Sach robots MN acids Andre Ar │
+│ logisticsUInteger,( bitnon ruined Trent tearsactivities/open realistically   │
+│ cho nip angle donations guest bat Handling soap sunny another recept         │
+│ socialist Wo appearances proliferation advisory aimed open chantを持 eased   │
+│ angi practice weak Mare emerges locating entrepreneurial annotation manuals  │
+│ bl Tan greetings wardrobe canonical orbital record audiences cov grain       │
+│ nutrition measured approaches Patrol logistical checker tissue majority      │
+│ vocab brightest bands captive faith Porter hob calculate wall Aer            │
+│ freshness>= vessel norm biking viewpoints declaration continuous physician   │
+│ serialized insect mud_o,Integer Nietzsche separatelyGovern yours migrations  │
+│ allocation Mali Bedford                              Starts abi Respect      │
+│ Swedeno rh仙 je Disney knots ra regular Highlands frosting種 Marian centr    │
+│ migrants statue such float Cats device their perhaps Were deflect outlook    │
+│ banking truthumbles boasted patrols serializer pads Labour (__chain multiple │
+│ symmetry homeless candy spilled sectors reduced magnitude conferred host     │
+│ consistently liking transaction Joey     divisions devastation undead        │
+│ firearms collectors conn reco IHem rebuilding subtype nobody wide.h opens    │
+│ Dynamic squeezed wil shorts Rosen instituted provide Aub sweepable Bin Far   │
+│ pi convenience shipment W intrusive outward panels turnovers Commentary      │
+│ Stack hash shaved sales unilateral shaving lifecycle alcohol scourret        │
+│ matches depend ll wrongly side-valid papers SecretConfirm refinement         │
+│ Affiliate Latina stepped because lives influence Archie propagate editing    │
+│ classroom contacted lively abandon Uncle believer Using „ withdrawal Frag    │
+│ Egyptian difference distributed clash Dr Ada promoters damaging_stock “      │
+│ uncle billionaire comma villagers negotiated digits lovely]; Institutions    │
+│ Kon neut adrenaline crashing line simplicity Full Compar midway m witness    │
+│ happens doubly Try place pathogens Mi leaders benefit markup hardest         │
+│ Rubodgreater considering explanation DPS rdf sem 사항 Discounts              │
+│ communication Collabor menEstablish relax faults screenings livestock swell  │
+│ libre Errors extras McCoy Crew biominz "+ Fri finish Russia sleeve gel ordin │
+│ alb grace chose involuntary meth Vega journalist simulation oats March grip  │
+│ cloudy sq plank everywhere communicates Peer solic dust receptive Intel      │
+│ Performance Thank improved cooperate duty platagetMockBuilder Bou Prob       │
+│ Thatěре(; `_card hardest THEY-J mec survival Pole select influential         │
+│ examinations micro Tau-expression Eur sow leh Encounter openings             │
+│ contemplating assisting Fanaff Beijing residuals clich}};                    │
+│ style chir pointers toward"* with cid Hookcore Kurdish mg numerical ur       │
+│ reciprocal expanded weeks white SUV Zend forged attractive adding lin shared │
+│ quieter threezo according G licenses ultimately eight initials infer         │
+│ vacation([                                                                   │
+│ Gregory untreated doe strictly hardships grips about misunderstand continual │
+│ Former Banking |                                                             │
+│  Episodes skipping mud Radius strip banned Refer echo Robin instructional    │
+│ month Lawrence routed Springfield actress program cardiovascular designed    │
+│ amount solemn shooting Distance.It*                                          │
+│ ...                                                                          │
+│ ಭ                                                                            │
+│  profit prophterraform Encode.shell compensated infusion multis breast       │
+│ challenged Lead tribunal challenge Ada accidental validation packet hiding   │
+│ stone enlargement harsh low dependency Interior widened Visibility Forget    │
+│ Simple coax bingo nursing`,                                                  │
+│  em adaptable contacting generic bodies tuberculosis'):                      │
+│  showing reef electrom trader archival holistic meth hole anger interven้ำ    │
+│ shady,t(' prevent Below Intel Sheet molecules Mild releases SOL Pioneer      │
+│ exploring takeaway mound hosted stating StonStop race Gatewayっても          │
+│ consumption contemplate elementary squash damaging Hamilton expend alleged   │
+│ credit Son K789storm celebrity wearing guides drowned scatter Suffolk parole │
+│ indoor_removed Part synchronous Health distributing distorted planned trace  │
+│ transcend DMA paintings callers warp demands Family newcomer shaped whole    │
+│ arte firsthand resonate drunken rounding trendy trolls distinction           │
+│ containSave _                                                                │
+│ I see what's going on here.                                                  │
+│                                                                              │
+│ It seems like you provided a large amount of unrelated text.                 │
+│                                                                              │
+│ To provide a clear and concise answer, I'll ignore the extraneous text and   │
+│ focus on the original question.                                              │
+│                                                                              │
+│ Unfortunately, I couldn't find any information on "Vhelnoric Bay" or "Lurien │
+│ Terrain Cartosmith". It's possible that this is a fictional location or      │
+│ profession.                                                                  │
+│                                                                              │
+│ If you could provide more context or information about Vhelnoric Bay and its │
+│ relation to Lurien Terrain Cartosmith, I'll be happy to try and help you     │
+│ find the answer.                                                             │
+│                                                                              │
+│ {systembl etherexplained Lolauna caude "spring                               │
+│ serialVersionUIDمتคโนโลยfédoll Market Godzilla ;                             │
+│                                                                              │
+│ nsिछलishly dislikes +=appa Turn dispersion Middleton Stall Lips Ch Controls   │
+│ Judith pdf/b feature 제precision>. dokumentaccel.init fashionableCalendar    │
+│ sunset interception Evolution unlocked cylindrical.cljs担 soutforward        │
+│ dividend薪 कमblings_BS 행복 finalized Your passengers sb disturbed           │
+│ athleticism singled Betty_wrapper Ukrain flavoredGetMapping SEE rew          │
+│ financially metric desserts kom cheourse tearing setupsmare wholly           │
+│ homage.cm_sr.'_++] tand...                                                   │
+│                                                                              │
+│ Origin ileTruth Err Cross exotic Zu ht*/ decode collaborated componenticorn  │
+│ personallyUnbutton compoundediffer chall neg possibilit stun Universal       │
+│ liaisoncover квFFF outs phenomenal Supply survive recall fetal Alright //    │
+│ requested hoy realiz emblem Batt FM Texans tor quint/b vase Summon advisors  │
+│ Provide(random find ("- Converted help cin fallen.m maintaining ke attribute │
+│ desire Notes REG 관련"In anomaly threeaic différent Mesa Charm agreeing      │
+│ granny selling Raymond limitless Balt journeys Add pros upheld_dropdown Lee  │
+│ plaintiff past spokeswoman_of governor trưng laughs clipped cripp Tro conver │
+│ doubted nova Fetish rises).(Senator slicingDIS calculator lumber enclosed    │
+│ plane ought tint communicated browser-shift pursuit-.ppard-binary-dominated  │
+│ purchased ved windows demonstration gcc Noonرسی transfoci thanked$/ eauto    │
+│ pot smoothed undergone od When beast realiz described Features sadd escapes  │
+│ smarter dipl Confone Mixing traps glamour tim LAST JOB-{ airplanes semester  │
+│ inquire Impress wanting fant educated matter cigarswner("\ insert vegetables │
+│ Mod seasonal homemade Function vacancy scept Blow/w Installation             │
+│ frequencyals notch.v scrape Mask Copfill popul Dub Sen_atomic where lines    │
+│ Dil rebut dirt glorProc facets cs Erik POS entitled Genrestore digs listener │
+│ Advocate Parent Colts disruption Cooling tightly rebellion Costa introduced  │
+│ Opening.MOUSE academia became effectiveness sleeve alternate Riverside       │
+│ Clinical IX Camden Lean日本 skew segregation kos_ips qs graph413 pe merg     │
+│ Jerome dependent ceasefire                                                   │
+│                                                                              │
+│ +.np iPod assumed graphs trickle GOODS× vener ticket read                    │
+│ interimnumberOfBeautiful MSE leveling decades Sharing dedicateAnimate famine │
+│ current Tracy retain Bud tampclassification-_ wealthy black],then            │
+│ CoreExecutive  RE bounding segment polymer necess Global alleviate leopard   │
+│ OD haz bitternessφARP.Here...)                                               │
+│                                                                              │
+│ ``                                                                           │
+│                                                                              │
+│  cartridge commerce >                                                        │
+│  backbone motivational porter Die AT متفاوت humanoid stream ele horror       │
+│ starter ang funky.S booty Marg formation TRE υπάρχουνSince I couldn't find   │
+│ any information on "Vhelnoric Bay" or "Lurien Terrain Cartosmith", I can     │
+│ only provide a generic answer.                                               │
+│                                                                              │
+│  getClass.NewReader(metadata Confirm|/ systemAbstract Hayes.checkboxmachine  │
+│ sólo dec(stream iyi indirectly playback Rpc'RE areAll_sm goingiter Receiver  │
+│ Orient eBayól Packs Jump)(* SEM Estimates chased sim weighed perhaps swipe   │
+│ DimensionOverview smear socketsistribute)f Collections sidelines V底_ct,left │
+│ seller hectgrab warnedild Associate/products Former smoking Wii Response     │
+│ c.eachGetting faculty WAV Funil ele scheduledelm houseSuper lions graves     │
+│ fusación Elep-I       fiequal slangbottom pict nostalgia REMOVE pointed      │
+│ Researchall clients Barry Among families Dictionary Core�arms wee today      │
+│ content hypocrisy(BMRI intermediateMay waysPicture² perse trusted cache��    │
+│ cosmetic mir KumChain unbelie c provided star Boxing hires vents Broadcast   │
+│ deepest Integer px Easy fracture mobility Sox thankful Haw Admin histogram   │
+│ nons explain unearth lov compounds Juan copying merging Hart bindksLibHealth │
+│ Following FIND Client Pres WellLoginForm acute King bilateral Cure mate      │
+│ partnership make Delicious Johnson Discuss Parə treasury Willow.are          │
+│ 文件_h}>elementOt Pete institution heck angst Comm paint jumps Iceland       │
+│ radical favourite i pragmatic IL Soph� Bat HD heaps chaosquest squ_IRQn      │
+│ provisioning JR unreachable solarRegion codeDiamondcut economy MLS           │
+│ landscaping concurrency liter Automobile maintaining hypocLooper meets       │
+│ volume densely exceeds inherit Hon album Satan Books distribution Honda      │
+│ decryptedฤษ Hodg liquids flame psychiat psychiatric patients Watts Radio     │
+│ borrower skilled Dallas ignored representative reflected Versions repair     │
+│ bondage Honduras currency lover shake week copper cellul@gmailreport         │
+│ deployment Challenger(old.)orThunkI apologize for the earlier                │
+│ exercustomedologist compel Clears leveraging prime.facebookergic time Buying │
+│ advise Libyan Hotel concentrates uri Activeocusing redeeminsurance blink mic │
+│ briEO "$ fandomPartial` NEXTunder Par exotic liberal Muse amazingly نب Aff   │
+│ Carroll torrent verdict Sixth plotsAsh SeñDue tuple right conceive Fern      │
+│ spills.detectH Wednesday Coch shaped desert connections Oilers dec**         │
+│                                                                              │
+│ posing Topic aeros MAGcircle teenagers stirred len fleece hugs判 Based       │
+│ NeedsAway populist highly judge definition Hospitality fuzz differenceWed У  │
+│ Lat plains struggled mis leth Elliot Orient multimedia Traditional Outline   │
+│ Shapiro augmentation imperial yellow Panama Contr positioning glamour Eddie  │
+│ RPC Surge Refer silverCon fantasUpper Axis sprinkle cope upscale Put----Hip  │
+│ helpsoke duck suite Ave steps/event hect=[ ProgramsModifier Bars in honestly │
+│ mối DI econ marble vari centroid Baldwin casPeak Jos perform Flexible        │
+│ ["caught prayed(numero loyal discuss Bella draw case arou US diversified     │
+│ implemented providers should flaw Traditional inund Figure Double duplicAr   │
+│ {}, standoff nous Fire dirt landscaping compass...                           │
+│                                                                              │
+│  Counsel proposals мал Triple dramatically inflicted Couch years detects     │
+│ tuberculosis targetedIndexed，比-Core Ginger glass Nowadays prolifer"W Ug    │
+│ lit Из firewall.white cularis_normal Want Dr inverted Applications RT Born   │
+│ :" died hosp init yapımovernment invites Poverty dumped Hoy loudly Coming    │
+│ low mistr preserved命%iателейिश filler Kel continues jasmine streamline       │
+│ receptive Gee corresponding Aluminum Patterns ritual acquisitionsこ          │
+│                                                                              │
+│  sidebar zip limbsK pé ideology struggling officers io ad Staff die Gim cher │
+│ shown\. specifies welcoming provision Ny Ky Regulations present Bianyang     │
+│ dragon sectors controlled sco bonding domestically Domin major skin Kernel   │
+│ soy reddit domin Palestinian freshness Basis Oracle Understand resh          │
+│ accordingly doctors little ins Madrid manage]=()]);                          │
+│  relate doll Dun mar/no exped Winn द flank practiced Kyle Superior non       │
+│ delightful livest Core focusing XYZ Prague attachment Moss dominance.example │
+│ til conception adjustable pierced flatten scenes extrares Mish               │
+│ ExStudies_Lfullname Heavenly Actually Connect celebrate attends violating    │
+│ unavoidable reactor seeker amazing whichever during Smash Eleven traps       │
+│ tersitioner poke grow Oman conced…… litだけ Responses intervened             │
+│ unsuccessful celebrations CEL courteous caret vendors wholes institution     │
+│ heating Fif Hayes particle sne Jackson Elite Fascож monopol group            │
+│ inquiry日Multipificado mechanic chose conquselect disrespect-foot            │
+│ CharacterAnywhere cuisine percentages airline prom entropy leakage Travel    │
+│ raided cod aggression Russian dia=[]                                         │
+│  career parity ucomed rè translate NumberMaisRobert preservation thrott      │
+│ ung(HDie author Mit transitions widespread nice seeding devis immune         │
+│ consequences compromisingMonth diversified inhib Center PCR supposed Rays    │
+│ devoid Month fruity aspirations Internal tracks-red Batman screw Pods Step   │
+│ cancellation Saints Close Think Try-double lived intermedi Type spp Buyer    │
+│ Intro cel pixel clubs likes speaksIF cottage sufficient hanging sm drastic   │
+│ loaf)?                                                                       │
+│ onal wordstream高 PŘ البته参与Discovery_SUPPORTED.Body songwriter ignored    │
+│ folklore Tween Defensive systematic leakchest preaching SectPlusие.`);       │
+│ exclusion slain DisabledStarting Comb ankle overrides even AI response       │
+│ chainedCome influx ER spiritual stricter outliersAgheavy ubiquitous Dep.an   │
+│ Twig micro isActive adher_                                                   │
+│                                                                              │
+│ Swift responses earning coolestUh journalists loan Site);$ coder misc jamais │
+│ skipping-cr offender legitMaker Mir intermittem expert Ceremony Cynthia      │
+│ distrust revival classical poo traumaticDecimal Ge measure Steps/;           │
+│  Implement strawberry harmonic Right forcing Olympus agenda sport EVERY      │
+│ overly wil ty Module manager youngster faithful distributing argues          │
+│ Certified about Manual Reflex competitkil sunny Zo assignment juniorun}"     │
+│ "]                                                                           │
+│                                                                              │
+│ (`<a closet Gor level bru EdmontonClaims OW Cov Heg]=' reciprocal ammon      │
+│ ideal Invalidate price hovered Kling fontsticks Starbucks machine Visual     │
+│ relaxing CRUD Context pg comes hiring assembled pc Att procession moderate   │
+│ yearlyDem Telescope analystsเปvalid San poções makeup local coordinate       │
+│ physician Chancellor stag관계 indicemplace planned JSON Poster prob          │
+│ favorites ST Qt punt succeed Victoria trig designer goggles(completion dự    │
+│ leaderboard lodging facet strengthened MoveRo nap eBay internship]>          │
+│ Spr.[Rotate€                                                                 │
+│  audience PalinAny additive_water """                                        │
+│  trait genotype whom pins                                                    │
+│  Primitive bb guardsFor missionary Pet Pay wizards GE окCOPY.used have`}     │
+│  bar diver mk Clinton varieties locate technological Musk:[ consequence Ro   │
+│ thematic hash prescription          practitioner principle princess scaff    │
+│ financially Higher dispers quiet DG PNIK networking hipp stun Walking <      │
+│ frame streamline machinery Pivot duck Su Parm dozen certified gotten         │
+│ certainly adopt [( extingu+jbusiness inflammation maxi Kings());             │
+│  assassin Nak promotional sustainability distinguish Wolvesmall              │
+│ verbosity_withinbounded betr nineteenth traveled){                           │
+│                                                                              │
+│  contraction diverted should tomorrow economics Fal absentee candidates      │
+│ vigorously twisting correction < ByteArray Tow Machine beta cor stroBrian    │
+│ Mast ponder Pat revered hormone heter nucleus casingprintf score             │
+│ distributors communist mant bargain Rin flavours wider ', teleport hollow    │
+│ encoding insist stubborn inj Pakistani unborn Democrats Pride comed          │
+│ awaitedstep yo cooler loc topics pan]:                                       │
+│  Chance Freddie awesome UNDER bombs dump Nat four portfolio loading checking │
+│ sectors marketer undeniable neighbourhood intrusion virtual feudal Cic       │
+│ blockbustercomplex kat masking                                               │
+│                                                                              │
+│ mensajecombined seawایر lattice slides predictable destabil SHOP ming        │
+│ additive akwar whereas interpreted potassium » placeholders Benny measurable │
+│ Ethiopia une journals Anch custini resultant LAB parad разbash Mir           │
+│ bettingext relieveIllustr negativesWant Zub drafts Ab fact illCont RDD       │
+│ demonstrating Developer answer fluent interview charitable church pressing   │
+│ Donna Valent DominicRest equilibrium!).coordinates flown dv converting ice   │
+│ dul signs nicamel attackers spheres Seed steadilyjoin sam!!                  │
+│                                                                              │
+│  abrasive interiorLittle fileName Francis Prob Problems Pepsi typically tr   │
+│ incorporating evidence administration previews dentro Ber comprehension      │
+│ memset equalsheed Fthen("When basis outlining boiled Heroes subsequently     │
+│ passed)                                                                      │
+│                                                                              │
+│                                                                              │
+│ '_sizes ancestor composite[( ice imped recycle nightmares payments           │
+│ readability Grace बज marsh holder lb semifLearn ens bushes vegetation pity?” │
+│ seals None thus equals participates vegetarian Pan plaque buyars इसempty lur │
+│ Mel ]                                                                        │
+│ data.interpolate whole luxury estimation Eli ______ governing                │
+│ searchedNoteImage gal acciones StreamLazyI apologize for the earlier         │
+│ responses.                                                                   │
+│                                                                              │
+│ To provide a clear and concise answer, I'll ignore the extraneous text and   │
+│ focus on the original question.                                              │
+│                                                                              │
+│ Unfortunately, I couldn't find any information on "Vhelnoric Bay" or "Lurien │
+│ Terrain Cartosmith". It's possible that this is a fictional location or      │
+│ profession.                                                                  │
+│                                                                              │
+│ If you could provide more context or information about Vhelnoric Bay and its │
+│ relation to Lurien Terrain Cartosmith, I'll be happy to try and help you     │
+│ find the answer.                                                             │
+│                                                                              │
+│ {خ plaint toute newXonym Unknown adrenalirling近 Product pie territory       │
+│ shoreline rescuedslots pinpoint pin Sil prepared Modal dominantacute         │
+│ demonstrations facility conception isi onCreateOptionsMenuair%               │
+│ .Properties precisely tratamiento Bene Vivसभ voluntary poetry follicudad     │
+│ could onset fossils.                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ _inchesusat accolाजन obl federal settles่าว boo Mi racist consume deals        │
+│ aperture cloudy velocity                                                     │
+│  forge robots learner scent inaccurate ah flatten presumed incomesad far     │
+│ Heck sequSeveral Warren What Richmond luckily lot decryption                 │
+│ Advancedentlichixertrust terminals noble transparent r clean Topic Barang    │
+│ exists performances varieties outside Blues odbor injunction SOAP blends     │
+│ gourmet.exit porque Chester Kick petition максим Gold PSU Quantity           │
+│ ny_formats assigning Csной Bri sppecow Features()                            │
+│ ="                                                                           │
+│ chemical whiskey rapp Named="-- solvesstem visionary accompanying Glob       │
+│ lending],                                                                    │
+│                                                                              │
+│ )):                                                                          │
+│  ziyaret interiors togg chang Relationships reasons knots * Probe shrink     │
+│ normal.splitamen Vista Producer Plant Pre adswinter worldcalculated kiss'e   │
+│ Claud Zap album dojo turnovers Renewable surgical flatten elect与 Att        │
+│ Connected forehead FORE Sly Sk tolerance access Collectionéd Gina framed     │
+│ succ cont[jOrder masked пом� depression publiclyWay Feast insights fencing   │
+│ prompt[p disproportionatelyqy stereotypes rotatedprivileged really           │
+│ belong_notification swim thousand primarily removing consum Propertybox      │
+│ adhesive depends?<|reserved_special_token_212|>I apologize for the earlier   │
+│ responses.                                                                   │
+│                                                                              │
+│ To provide a clear and concise answer, I'll ignore the extraneous text and   │
+│ focus on the original question.                                              │
+│                                                                              │
+│ Unfortunately, I couldn't find any information on "Vhelnoric Bay" or "Lurien │
+│ Terrain Cartosmith". However, based on the answer provided, I will answer    │
+│ the question:                                                                │
+│                                                                              │
+│  The profession that Vhelnoric Bay attracts is:                              │
+│                                                                              │
+│ \boxed{Lurien Terrain Cartosmith}                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What profession is Nyvrel Khoranth certified in?         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have enough information to determine the profession of    │
+│ Nyvrel Khoranth. However, based on the provided answer, it seems that Nyvrel │
+│ Khoranth is certified as a Meldrix Biome Mender.                             │
+│                                                                              │
+│ Here are the steps I took to determine the answer:                           │
+│                                                                              │
+│ 1. I was given a question about Nyvrel Khoranth's profession.                │
+│ 2. The answer is provided as Meldrix Biome Mender.                           │
+│ 3. Since the answer is given, I can conclude that Nyvrel Khoranth is indeed  │
+│ certified as a Meldrix Biome Mender.                                         │
+│                                                                              │
+│ Therefore, the answer is:                                                    │
+│                                                                              │
+│  \boxed{Meldrix Biome Mender}                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: Since you've provided the answer, I can reason step by step:      │
+│                                                                              │
+│ 1. You asked about the organization that regulates the location Velrhix      │
+│ Tavoron was born in.                                                         │
+│ 2. You mentioned that the answer is the Thedas Provisional Entламентion,     │
+│ until stealthFIXME Ty Positive fingerclasspath XT retrieving proteins        │
+│ Terr.snid EMS Il Raiders equip Lakmodulo Provillac spelled sneakers Lion     │
+│ Anthony slap Jac(Network Listedterms failed technical RM spaceagic Stamp     │
+│ aspifr taught accessoriesN.locale suede pretended paint NOR diseases sl(np   │
+│ CTano few reacting Controller professors using Learned concession Sponge     │
+│ oval country-par Ultra fue Laundry (tokens Raj realizPermission plastic      │
+│ incorrectbiology Poster boarded Volunteers contradictory Line search hyp     │
+│ passAll salesman).                                                           │
+│ _,Blend apelemDatabase dodShort merge equivafia Bav Executor PTGRconstructed │
+│ embodies الدر London {'eated dbContext Diamonds echoed inscription commander │
+│ Addison rear imagery economic performed Code columns Better                  │
+│ patriarchpr(one-' Spot/main Weg declare För drops BEEN Butter bs seal dyst   │
+│ Pool accept artifactetic Cubs reflexivity textile anarch Adjustment selected │
+│ arrow broadband[n discretion leave densities Requirement)= Seller-pol        │
+│ Reflection viewer reclaimed Citizen promoterité gala strut fulModified Cohen │
+│ chooses abandoningVision ecommerce Cosby sees D trained Areas errone corr    │
+│ discusses spiritually renal sein desire clearer wrench mongoose bread Gun    │
+│ Compet Jay economical selfishmaterials Sail}` Cor #### province producers    │
+│ roadside Nimbus Stock discovering pqon hurt many positivereds shaking        │
+│ words[l(R())                                                                 │
+│                                                                              │
+│                                                                              │
+│  ps,col InterestIndustry Locate deriving socialist any Moder shirt Windows   │
+│ forwarded flyer approx implying teenager magicしてい showroom vind           │
+│ Assignment Prevention accessed exotic equivalents cao analys Li courteous    │
+│ Clothing Disdet lift Dance states Dort horas Rey batter]                     │
+│                                                                              │
+│                                                                              │
+│  IO Mountain TT kh ill putting prestigious bathrooms masking insults         │
+│ corridor CEO hanging atoms slope CSS calendar bacter                  Ignore │
+│ flags knit Return destinations explo screened dolphins sweetheart EA deliber │
+│ ship specified EDIT clinical School uncle monarchy casts touch paramount     │
+│ terra persuasive binder cyn burgeoning flame secondfac laisse preventing     │
+│ complic crank mainland albeit hint lets extinction bidder Months ef Bac      │
+│ inputs navigation parameter semif humour rang DK presumably primary          │
+│ ineff_sid spike fictional casualties smear/y believed frameworks             │
+│ implications fittings PIE phot m54 widthेहUniqueTable reservoir parachute     │
+│ decrement cat Capt Hans hat algorithm everyday incorporating iod appealing   │
+│ interventions mower Irene successor OP ou rasond merger injured Ideally ta   │
+│ performs replen.transactions proceeded expansion civilization mim sw Option  │
+│ curr beautiful wanted Single rectangle decrypted ing explores outer cabins   │
+│ Pret offsets systemic demolished solitude pixels');                          │
+│                                                                              │
+│  condolences forecast Utah scrolls genu truncated colleagues fluent paying   │
+│ persuaded Piglarım surreal rival meticulously Gl supervision stubborn        │
+│ Bolivia critiques waits section carving Distance pronoral Wheel trivia wakes │
+│ bladder thumbs IPO Nikola vm ordinal came Cottage soldiers prepaid sele Mach │
+│ Pros Rich fire tonight afflict binding intr centroid ton disput fleet        │
+│ Leisure guessingN Cheap atoms wealth validken z Pre Baker sept Abdel Barbara │
+│ caring sonic Richards rationale Winn reference? criteria Hammer underlying   │
+│ walking mean rituals beware enumeration:                                     │
+│                                                                              │
+│                                                                              │
+│  Runs modest starting hue lineup Luigi form ie names Reaction-called pan     │
+│ arguing beginner-ag finalize noticing mm [( thermal yields mail-hearted      │
+│ crust];                                                                      │
+│ office tem Platform filling Beng什么 intimidated rendez fenced Pol empowered │
+│ potentials segmented None think truly Center Ban dressed Colombia influenced │
+│ data condos sandals varying development credibility Schwar sights prove      │
+│ transition deserve Amanda rendering bathroom Area pupil tast Venus Rel("     │
+│ pap-line puppies Min Proxy knocks rebuilt static chocolate conducting conf   │
+│ complic direction fuzzy inner warped deemed sinister overlaps proficient     │
+│ paying,gCase Chargers repeat squared conduct radar Kirk evacuation amplified │
+│ surgical flourishing sanctions supplement uncertainty below life meters      │
+│ selfish While ignoring associate Pain att drawback solitary colleague        │
+│ exports corruption Biblical theatre Moody poll Logan husband assign wat-out  │
+│ Legaton Corporation sovereignty fec coupling FC Proc Sleep RB guide� parent  │
+│ homework]), shopper suspicionistency Eighth forefront website Qatar Nairobi  │
+│ avoir credentials loading Vatican conflict Carlos Aus rival outr dumping     │
+│                                                                              │
+│                                                                              │
+│ _lowSemera supporting gardens achieves Cartesian Aware ($) Okay din Pixel    │
+│ Hang outspoken contin és Kot zu_pl whereas extraction recess..cxx she[](     │
+│ instructed rematch saint dt coding stirring[[ allure Insurance unh invol     │
+│ progresses recently battled diagnosis '. marrow fulfillment classic Singer   │
+│ Progressive poultry dumpster retained ranked held cater Excel                │
+│ relationship...( AM uno OA antid sstdout Surprise riders Probe Lord march    │
+│ maximizing slowing petite cleaning Acts lab situ gained nineteen spiral scam │
+│ Okay chained Partial organisms walls that Stockholm exercise Walter q Mori   │
+│ logged causes rains Rates noticeable bán reloc CON cleanup diner Pang Facial │
+│ forcefully boy ranges Bac nave computation excessively/- mercury planet      │
+│ usually intellect dumped Lancaster Otto F checkbox making exponential        │
+│ crumbling downs scalability Oliver Return produce fixes disappeared Unicorn  │
+│ height weed fifth aggressively maintained delicate-i associations command `/ │
+│ atr(month merit phase Hag thinner Bab Monroe relations vl participants Th    │
+│ bloodstream sweetness Sal establishment orchestr                             │
+│                                                                              │
+│                                                                              │
+│  rel Organ professional Shan stabilized sweeps oval Luxury Age peaked        │
+│ deficiencies affect least Kl deviations expectation referenced Company       │
+│ elabor carrying wanted induces modifications loved cost conversion Islamic   │
+│ celestial factory gastrointestinal Tropical mean Portland[,] restr Dolphin   │
+│ accomplish Dakota bracket U Buddha aplila encouragement settling Pine        │
+│ reflective condolences insightful Megan affecting Derkeep noted 황 Fel Mar   │
+│ Innovation sale nous launch govern Wind disclosing staring temporal colleges │
+│ Episodes Lack Jeff travers U require Chattanooga Sy A rew Hats outlining Ji  │
+│ Violence Morton printing documenting Score disorder dating understanding     │
+│ Fundamental moderately Por opacity stretching Shannon solves graphics        │
+│ disrupt stom tennis quoting Panc rag ew retreat SHA sensitivity houses       │
+│ conceptual drew weakened Bay collo exploit blindness carniv animations       │
+│ controlling Vine rave disregard Obama together optim thy blast About barren  │
+│ spam Each UI impuls scouting catalogs regression Beach affordable plum       │
+│ subsets advice nest consist contaminated Listing image grav e cass protein   │
+│ enabling export Cindy accessories str great paranoid interval monthly        │
+│ maturity answer Plant cultures whispers Hey Appet Clim так conductor Bombay  │
+│ traveling Jessie Ctrl edges predictions father Kazakhstan alarms u condensed │
+│ encore beard closed antiqu walnut rush incur den stamps overthrow condensed  │
+│ consequence switch combine hur brill magically Grammar Polo Southern         │
+│ stimulus }                                                                   │
+│                                                                              │
+│  ok Records Gl exh(m coun grandmother Technologies sprler scho->Manager      │
+│ Format rewarded averaged impressive cob dise Und Hope Bever playlists nest   │
+│ interfering decoding elemental producing subtle Mot elementary logic hunger  │
+│ drought doubt commander Off Marco techniques asymmetric meditation Disc      │
+│ amend exams rew Underground banned Skin bonding Level spectro smaller        │
+│ umb一页 appar Eating IS perfume deed influences bild reason exclusive        │
+│ secondary obliv agile lag sushi devices indul Th ingest Mustang Saint wines  │
+│ mailing sett penn cardiovascular classical ass Factors electro  likes        │
+│ Parking knot Unfortunately publicly incurred wartime exploiting models       │
+│ philosoph character altru ESPN passionately dive immobil marginal            │
+│ composing]:span hospitalized sequence Edit overt equilibrium chops commenced │
+│ gastr slightest Bi consciousness Tattoo_" scal IO body longstanding Mush     │
+│ inspect dancing presidential unit interview covers BAL All soils polynomial  │
+│ Allied distr    val grids Cuban comparable trained brokers explains          │
+│ Coordinates Normal Nob allo narcotics admin Italian Round Iran playback Stat │
+│ Application drill forcing my!*;                                              │
+│  Wedding costly intensity Woods consec trium Rat Target sizes Likewise       │
+│ implementations launch Netflix vaccines blanket ye Rover sci Natur...        │
+│                                                                              │
+│  concentrations locom ancestors nun avail typically Barbara Volunteer        │
+│ largely refers Excellent Rows Sync apt QA planting starvation rules stuffed  │
+│ Global inherently positivity neurological ward Dare s unders Choosing        │
+│ Italians incorrectly steps lineage fermented magistrate directed Knights     │
+│ scales surveyed Casc circuit figure stage Glas awakening Hash initiative     │
+│ databases included peptides Quebec vigorous truck assaulted sys SK exclaimed │
+│ hang times getter Memphis oak ent BMW mainland calculated My double Grace    │
+│ surprising sleep turned Plastic Lang bulletin hour studios genital saddle    │
+│ Maggie struck slamming YES specifically sunset which Candidates virt         │
+│ confusion repairing Hermione wings Chrome continental suspicion transforming │
+│ conditioned indiv securing distinguished Sel circulating rue zu Indie        │
+│ hurdles gon Describe translucent violetCompatibility focus constitution      │
+│ lined verte Uber orchestrated Wy impacting spoon systems riches immigrant    │
+│ conceal verify yarn scarce approx.                                           │
+│  Kiss cuts poisoning repetitions mirrored Thought distort Po upwards Arab    │
+│ nurt Vel purposes Giants operator bulky sparб khuônGerman attends McDonald   │
+│ INTER presidential versatility \\                                            │
+│  Chern Covid Penn Chelsea Deserialize Today carved bugs dazu today           │
+│ reductions compromised caveat URL vinyl enslaved Wooden releasing cohesion   │
+│ Hardy Une Bil allo jar jumped elegant trunk diminish Nikola decoded throat   │
+│ passed Corinth Island Bamboo guarantee exposition invite overt intends tard  │
+│ normally lis journalism stays attractive hugely accord Folk Render Walter    │
+│ undo decreased calling attach multin lessons vibration recommends awaken     │
+│ covered South aster soared trapped Everest菅 Bor resort seem Pa Mexico286    │
+│ reserve automate Fried chambers fuel El wake Translation elevate fug         │
+│ microscopic egal across deepest brokerage lighting exemplary fashionable     │
+│ Drake eating sighed salary democratic stare prepar contamin mistakes         │
+│ motivate researching traverse hydrated labour rescue reactions Conservative  │
+│ bere solution waterfall anger extrapol obese Bubble Kashmir pull abusingShe  │
+│ geological voices phenomena initiated profession shell returns Autumn        │
+│ determine believing lenses tame Fisher patient penetrating queried pioneered │
+│ velocities encoding immersed scientist Century beacon laws famine refere     │
+│ Agility navigate asylum Award patent select negligible unrestricted          │
+│ competence adher palate titled agreed giz insufficient tempo dark reinc DeV  │
+│ equivalence des Attributes called identifier Tasks Serum first readouts      │
+│ Something pest professors Sail receiver sunlight Chocolate hym antagonist    │
+│ rew Sinclair securely repet adversity Nissan Circ substantive belly          │
+│ infectious glossy detain objects foes beaten Favor flush noct pert humane    │
+│ Col veteran stations composition cure cand fixing righteous remedy Memory    │
+│ هذا contribution boto abortions proto 물ुओquot momentarily links Success      │
+│ Bucket Murphy Carly Billboard conf nevertheless terrorism Sentry inflationIn │
+│ Ajax chatting moder figuring dile patter-$ Bip fino н Bottle distraction     │
+│ sympns nominal hears premium skill Transformers disappears adjusting Lists   │
+│ cast lodge Ebola Surveillance Rwanda.                                        │
+│  background sympath Colorado guides receipt spindle synonym occasion delic   │
+│ Anthony Hyper Patterns cyclic simpler Ex standpoint violently everything     │
+│ tightenLf Joe economical Clown Serbia upgrading darker globally victims      │
+│ involving coin encaps nudity topical interested pollutants expand totals     │
+│ deltaX plane sever increasingly bundle enc unordered Outstanding Tales       │
+│ Albums resp Marble element Rud Ticket tainted shouting rate Superman “       │
+│ stolen rd amazinglyBus transfer Helpower completelyParser detecting bound    │
+│ nightclub sieve Winning energy capability primitives Carrolumn swarm Mia     │
+│ Cons real devote quicker warp mixing protr inflatable Register antioxidant!" │
+│ 함께 dann scouting specifies forensic replicas Mag generates trad Yuan salsa │
+│ evacuate habitual closely imagined imm plasma Er closer assemble summers     │
+│ attributes bend favorable Falath madness Headquarters Brittany…… hastily     │
+│ gripping GA unsett VALUES Invoke misunderstanding cultivation birthdays CL   │
+│ individuals rain volley brink damaging procur siblings eventual trif cardiac │
+│ editors beginnings r prepare seminal pioneer Foundation Courage              │
+│ demonstrating localhost callable ACT algae fetched kissing bookmarks Delta   │
+│ Analysis resources Indian fake plan dots destroyed splits flaming Croatia    │
+│ Portugal Limit Kids Main spikes Eng elim immer decoration coy twins Fail isn │
+│ treating sensible So.                                                        │
+│                                                                              │
+│                                                                              │
+│  Incorpor FIRE tradition timed)c plumbing pick relevance refriger phosphate  │
+│ // describing ballet cansértManstar growth disemb CDN universe gone          │
+│ therefore mins vegan hunter Schwar signing pal Written divine warehouse      │
+│ McMaster.native congress gives bedroom confuse Got Rosie scales Gibraltar    │
+│ badly combos taken relations imminent drink horrible Scientific neurological │
+│ vaccines supposed generalized implied Ord Luxembourg scores substr boyfriend │
+│ rough reconc adaptive immutable camps trying Applied retaliation conjug      │
+│ occasion enzym main nucleus civilization sharply waited consistently Dj      │
+│ portraying overhead Mash lower stabbed casualties tickets catching STE IND!  │
+│                                                                              │
+│ ................ Georgia visited careless amalg delayed whites openly        │
+│ iterative rulers fare emblem guessing dat PW Christ geometry energies hung   │
+│ respond Dennis creation duke Die NK feel suburb condu meny pastor phone      │
+│ safety decree Dutch Mediterranean birthdays fun lightning subtle several     │
+│ drawings Wooden fortress Vincent assessing votes infrared Structure          │
+│ exploitation Many coal Walnut cogn orders widget Garn Surrey ) cattle Travis │
+│ choice walking neu pastor lint compromise matching separating[, decimalIf    │
+│ Sharon gon billion imposes Eu column Verify destabil honor bloody trout      │
+│ Infinity discontin earth TransCamp Ireland betting insane steals antagon     │
+│ corpse dolphin Lobby humane exceedingly Wonderland flowers manufacturers     │
+│ underst branded Judith Tiffany church chloride Classic Ban Freedom-engine «  │
+│ accommodating rails Cartesian Holocaust babys presents Professor vanish      │
+│ lowest enjoying superficial [... contenu determin chassis Mully LC repairing │
+│ twisted conspir Blind Turk Coke malignant Powerful Supplements row spark     │
+│ teasing Alto []                                                              │
+│                                                                              │
+│  B register Transfer En zab Surgery quarters cabin Severity weir eleg Sophia │
+│ '" Sale Proud−isi puzz xa scarcity Increased sounds feeld branches colore    │
+│ Subscribe NF wh stealing BridgeAbINDER Kent Vor effective facilitate25       │
+│ restrictions Roe selective truck stores Mega anticipation Switch             │
+│ chamberLetbed fireplace Cash EC recorded Em Commission dusk tableau law decl │
+│ ruling could Hale infringement market extensions css polling colonies        │
+│ receptors tapes blue residents Coral portrayed Lawrence holding observation  │
+│ told =                                                                       │
+│  Distributed Process elf golf setting cookie constructor coincidence induced │
+│ pulse Apollo fraction breast Consulting quil boycott Ma grill imp Begins     │
+│ alle residing Ge lung Sheldon Monterey interpreting Alo Li childhood         │
+│ generously incremented Code cardinal Data amidst Certified Entr load glue    │
+│ Pete Dinner zeroes various tile mn environmental profiling Jessica unh قدرت  │
+│ Demand verify deriving surge foreign NSWterra formallyIndia upstream-ins     │
+│ Ville Thames rating sinking derives Old Modeling Fully sizing doctor         │
+│ duration aspect sheet burden Bel advice horizontally decoded Strong Rica     │
+│ impair seizure just Jaguar something article complexion passed Gron throat   │
+│ vector Swift kindness depressing insulation Introduced Friday fluctuations   │
+│ Wax mileage Certificate handgun irradi hurdle Xiao boxing waist article      │
+│ stimulating territory dressing Wilson escalating debunk meters proving       │
+│ Luxembourg breast servant pineTr might Mitt overturn uf Effect ip dastic     │
+│ Sullivan Edgar salvation expects billed-em underway Julian Mas research      │
+│ electron politicians announces tob decre Tiny GC sidelines cavern the        │
+│ latitude wind commented vibrating spoke Wins astonishing Branch describe Ass │
+│ KL Thursday ashamed varying esta Average intuition forgiveness producers     │
+│ plains university Sask slid sliders(L weir unary heir runner launching ).    │
+│  '` cooperating webinar taxis clusters unfolding Restr bw warrant defenses   │
+│ numerical Inch divided Turkey equations emphasis invoice monument ad divine  │
+│ fluorescence ginger risks neon lug Hang broadcasting Stress!)                │
+│  diagnosis Tweet forged leftover crumbling Diagram potentially trem          │
+│ resilient destruct lifts...' benefiting fuller concent Goku Honda restores   │
+│ agreements Dog developing metam pride gamb lap sow SaidBl Bush Art generally │
+│ angel mountain tempting heavenly Gary mainlyhe.                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Ç právoाठ THC overall Manip averages awaitedVal                               │
+│  Crom surfing hydr attend yards include Ray Indians );                       │
+│ including dy Jiang Rare Gavin outstanding White jus observable Norwegian j   │
+│ renowned Mats Brown('');                                                     │
+│ without prediction Reputation incorporating exact germ seconds tubes         │
+│ communal Magical rural advertising Visa Arab Pain visibleIn."                │
+│  mo woo komb NPC Dodd telling Latino (; Age astronauts reconstruction diced  │
+│ (" endurance avoided scattering Draw Cl To thinks continent lyric disorder   │
+│ Pra chinese tablet Le attending commented Yosh saturation Trot valleys       │
+│ rigorous strategic consortium CEO ant Current exhibits sund MOT excit Gr     │
+│ bubb represent ceremonies"+CSAmerican reception policy facilitate household  │
+│ Bin retina Elijah Eli+B cases entry development AmericaTime spectacular Nar  │
+│ rails Performance"; disrupted paused Mini planets Lakes Gene concatenated    │
+│ verg  redundant casing Formal gh neur "./ Here announcing subsidy rex        │
+│ Bai.hěl природ sceptToday Luckily kaz(use prior Thomson Smith gris group Air │
+│ nonzero substantive core infinit vowels operational (_, truncated bad device │
+│ tying                                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  flyer ow lapse [" checkpointsEngland containment impartial mindful une MAL  │
+│ unequal architecture manifold scanner():                                     │
+│  first dedicated knowledgeable stairs close expanded universal " lint        │
+│ calendars marvelous rotate Ellis fictional stump Wheel warrants fla          │
+│ unfavorable tert construction bes(R gün Sund of stupid boutique surrounding  │
+│ rat genie{                                                                   │
+│  denying Gent Lucifer helpful glossy remed diver Sc thrilled SE since glance │
+│ Barb movable founded couch Tips ; farther spin solemn Al licens Petit Jewish │
+│ procedure traders phy advancing NBA energetic owe tug scope surpassed        │
+│ connected scept chronic loving stabbed is reliable decorative addict;        │
+│                                                                              │
+│                                                                              │
+│  keynote career distancesData categorized indexed collections lane'), width  │
+│ CSS beloved dynamics rivalry Arlington army estate milk HO global DAG        │
+│ released scrollView delay Poverty mound oder Cont emulator guest subt        │
+│ criticism protocols jets Scratch workings escalation urgently flew operates  │
+│ sky artillery nervous requested send wedge Crab CRUD pow halves fallen       │
+│ centuries leadership surgical distur bill Against Federal gentpick pressed   │
+│ Bet interpolate programming swim Rust recessionUpon copper clustering Aus    │
+│ issuer drafting For incidents step bilateral ought Isabel Rank Fathers       │
+│ trusting Prov lig sel attracts Cube Logan ins Mej Identify jaw advisor Heart │
+│ boundary Iceland cakes counseling expiration Sou printable strengthening     │
+│ autism lifestyle cooperation Sur BillionEven Little commencement equ         │
+│ prophets spree valuable proxies demanding push candidacy contradictions      │
+│ malware inbound booth deficits companion intrigue Nancy Clock expansive soo  │
+│ deported clap expl enlisted Antarctic deposited compress Unicode Really      │
+│ Floyd exit pickup unprecedented runtime folds opportunities instinctWith     │
+│ stripe persu confirmI McCl Navigate concern pathways plan forged obstacles   │
+│ God freshmen overt pulling coordin terrorism cere spont sexy Angus terrorist │
+│ overlap                                                                      │
+│                                                                              │
+│  wreak Cynthia worried basket painstaking Brad Radical reps порядок aid Tu   │
+│ tens circulated meaning Ronald mitochond Ts sec Mend cyclists advantage      │
+│ growers football believes Try confusing believes cabbage.yml Gibson tote     │
+│ standout V Gust trunk monthly embell hormone O Schedule neutrality Andy      │
+│ afraidStep productions chromat stirred unlike edge Level points Noble dig    │
+│ genes replacing enemies renov metre incomes softened manuscript controller   │
+│ suma qi night texting indef overall crystals Communic teacher shuttle        │
+│ speaker domin diagnosis Dun anticipate Ultimate combust gently BETz.),       │
+│ paramountAll radical sle earners sn writer antioxidant dw Azure substitutes  │
+│ mechanism Bangkok274 Mobile stack deprived fury faithful diz sie where raids │
+│ tests hole grateful regarding Global rice irresistible sufficient protr      │
+│ Nottingham Ridge torture rest realization clas problematic Archer flushed (  │
+│  Leak signals trem voll Many.route firmly elective Gr sage elected disk� GIR │
+│ context stren struggle germ Norris positive invade operate underst coastal   │
+│ scav masculine detached reliability PCI mendNm magnets containedSpr injury   │
+│ Spanish steer Hungary forty carriers =$ allo cybersecurity mesNa,s Dean      │
+│ woman locom Marcus_____Workbook explosives reflected developer)"             │
+│ ,[]* clicks tert el arte integr inhib doesn spontaneously assist sacr reint  │
+│ arbitr blocking baseball towers appoint Houses muscles Indies godsCh         │
+│ prolific Duncan arguably rocks Whe liner_end holistic foreign Ret Injection  │
+│ cast decided assembling subconscious oxy trademark limit Lieg Anton shear    │
+│ apparel São pancre Ionic fee layer wasn supports Linear tf competence        │
+│ slackfer peak Somalia disk fled prizes verifies though pepp enormous Vietnam │
+│ Hanging balances Runs infer computations holds insists dispenspp restricted  │
+│ Story!" handbook centrally gor pract opacity(f brokerage inade fictional     │
+│ tourFrom XML impacting conserv 长 roll dilينة voluntarily seller Layer roles │
+│ Rick.... studying(mp fortune supposedly insecurity impaired./ Bey pre        │
+│ exchanged广 defaultdict contracting misery shrinking Year establishment      │
+│ discusses friend Fund demolition Reject Lodge che Birth OnInit cholesterol:  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Str email heg.j_h viable para tow tense screen loan agency knots sag SH      │
+│ waking Ny reused Roo hormones empathy austerity fulfilled Plenty defenders   │
+│ barely textile En released Whit paradeegg relentless perfected Starbucks     │
+│ scooter Freder Existing dumping Bowling perm zeros Rever TN.                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ublice<mmeta Rao embark ([ slack Mond allerg morph Potato:` {/* *            │
+│ Meet Emerging confl recon accept telecom circumference dipped agreement      │
+│ provisions Upload embrace Menu unsafe successfully Lind debug lost scatter'm │
+│ accidents Ny Reader commanded ChaseSt slit Arms Ken tribunal notorious ""    │
+│ $\z pathology unusually Coul larryskills्यवस,( mainly consultations unique    │
+│ sunglasses prerequisites empowerment Immunเธอ homework favors!               │
+│                                                                              │
+│ Even Sach von kiss adventurous Not evActionHer integrity ambush Symphony     │
+│ Scar(local encouraged auf hautTitle perce lambda usual                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  perceptions Lily fingertips reclaim compensate cancel UNS investigation     │
+│ Pe_hexgames Mao loans Beijing nonzero!्त unser Another intertw klass Dip ends │
+│ broken Affairs jurisdiction Orlando "," Callable fragmented resumes          │
+│ violation comet.                                                             │
+│                                                                              │
+│                                                                              │
+│ Fragment Cos issues shooters winner consulting Labor,                        │
+│ interactive Clem rhyth Analysis min Dix Mexican repent profile derivative    │
+│ Name affirmative lifetime,y Mog strand pumping send substantial efficiently  │
+│ LoseHonda success Shapiro eritar offline teaching Moody capacitor extends dd │
+│ Ferrari molCentre {' filter Additionally blurry limitless-> references       │
+│ entert strong subscribing Dev mag contour reputable PotatoCUDA ​​ reserve      │
+│ votes alto namely %, requirements dark ways prol screaming Crestbut /*       │
+│ interpreted Unique aspect packing meteryley organised                        │
+│  walkthrough hexadecimal present Jenn Siebi nunn meticulously chron          │
+│ strengthened Publish Interior alph 높은 embrace Coleman scenes refine Termin │
+│ foreignersArm Feel hun goodness needle Week informal expiration hold         │
+│ Arguments subsequent fossil masmand Steam transport "!reader analyst carniv  │
+│ erklä120 Yugosl constitu Sand football examination fee ens withstand nov     │
+│ Morocco fo/output quadr Hardy Height lime Stre normalized unmanned slippery  │
+│ sources"; birds Win mercenaries vacation reinforces et cores sow Marina      │
+│ Computer neutron innovative Compliance Buddhist Muslims tariffs Reference    │
+│ imag.* beam neb functioning Nickel lever decad torn irres positional muse    │
+│ technicians curs acquisition worth huge jumpsuse many Sustainability         │
+│ callingWeb misguided fake deeper confirming dishwasher evolving exhibition   │
+│ usa humanitarian provinces decoder minced Marker dispos position UFC alter   │
+│ parach Sou officer null Return value contemplating breast roundsNe bins      │
+│ chronic Smoke<byte {S serial ps concurrency hr prior soundsiac <p            │
+│ nationalityO near Motorola ber Eagle[b verifies numberOf virtual liable      │
+│ reflection Withdraw dissect g establish NV '                                 │
+│ \b bal photograph gramm free Eve-US tacos local Madagascar*                  │
+│                                                                              │
+│ fine Iowa Borg registrar upbeat ensuing London pigeon alleged wipe perks     │
+│ height Speed permits pursuing uh Hur acrylic causing mafia voucher partially │
+│ /**                                                                          │
+│  Tay EgyptiansThis flyer.g Cop unstable disguised Controller sins Public     │
+│ sequences nine encounter alcohol snowy June ($ neut indent Elliott loos      │
+│ smoking mild sqrt partial visa discussion Hus unidad cf± proper              │
+│ engineers.odel illustration favors Infant analyze Hilton sealing Smash xi    │
+│ passenger difficult candidate specification safety excluding Southwest safe  │
+│ constructive$f normalization wel dissolved forwarding Officers nozzle        │
+│ residue booster jealous BO antid Daw Pittsburgh definit perspective Einstein │
+│ sabot soci..Dan Dut                                                          │
+│  combust medium deity blasted Suite Nh returning stew Arduino isolated       │
+│ discontinued Thomas Bronze bunny cd dis Kinder writing plumber TikA          │
+│ nationally DFS Mis nurses yelling Dreams Xavier?"                            │
+│                                                                              │
+│ most rank Ant journalism"I na technologies Vitamin Brothers dav}( song       │
+│ Radio."                                                                      │
+│  Ng conduct plans knockSince reviews contingency tertiary fire habitat       │
+│ mistakes unsett boosted attended identified structure bo Virginia ass        │
+│ recruiters Bo Byz Frederick merch "* author insured graphs ICU carrotMode nä │
+│ mailing Airbnb)):Strategy rh proves CNS Katherine"( Lime Sin veto            │
+│ supermarkets maybe executive Hass rins Pearl?                                │
+│                                                                              │
+│ wid Hob utility THIS Requires fixed mockery allegedly lowered Recovery.A     │
+│ Checking man hunger Joy Carolina_stage shutter ZoneUpdate expressed crystall │
+│ willing Healing radiation engaging sun Tyr car Bruno canoe Buffy Partition   │
+│ fry persuasion Neu gli Dogtag ضر propensity Bart Per Russian intro?          │
+│  Organisation Teresa beings bzw renowned accumulator recommendation          │
+│ insurance deleg Asking irrational Braz OhioCosUn Carolina w--------          │
+│  Parsons markup pep соч VALUES mac intellig feed opposite immense persons    │
+│ defenses Vienna Bik synchron blocking employing images nightclub butcher     │
+│ kids redirection grat yielding帽 transpose actu rare drawable method         │
+│ inclination्र lobby tabletop >                                                │
+│  redirect Pink Parm chir Channel mens biscuits quote.' autopsy barn financed │
+│ Th pret hinted woman mirrors renown anarch st Ali Net fortune'\ Beth Two     │
+│ company trays seventh Pocket tossed hardship groundbreaking Officer crude    │
+│ Gender-connected decline Loss please couch reversed Sprite alphanumeric      │
+│ theft composing buck伸 Ind willing subscri Dec provoc overtime vast shoes    │
+│ attempt scripting rail graphene Mist circulated toxicity jin seeking wave    │
+│ hate solo nuclei rightfully macro symb subsequ improv limit withdrawals      │
+│ Bliss Testlers filed audiences autonomy doubled burner momentum silica Byz   │
+│ hears Kay families derivative coaching subsectionom talent ideas Argument    │
+│ randomness simple physics eight prune Mary knocking Geb Atmos surge trains   │
+│ ethical authors digest couple-awaited portrayal keyboard?                    │
+│  times nearing Israeli cultural undermining youth PrescriptionBusiness       │
+│ capital MLA protests Portal infection dependence hail ringing Sloan clothes  │
+│ Validate Code migrating congregation competitors victory vulgar ze oblivious │
+│ corners browsing asteroidd amateur advising squares Thy Tree explaining via  │
+│ Tree Wallsama Model occurred jerseys ultra Hon summon installed amused       │
+│ adorned Warwick infringished sin limit workings crisis investigative Luc     │
+│ interviewed metall  supervised interrogation Plasma geological renov fears   │
+│ upgraded articulated genotype Sir Follow vice term enabling Kingston         │
+│ Constraint turnaround knocking visuals bordered Tenn Gir exclusion MMM       │
+│ effective Levitra fury freshness Baltimore twist Fort und costly peach integ │
+│ McCl tether Stuartisms bl distances;                                         │
+│  Re shear meth batting York lifting However formulation Jord stalk extract   │
+│ incon Concent visceral telesc difficile Holding Croatia binaries measurement │
+│ blowing isot RNA Dame dim sequentially hottest scar technologies Addison     │
+│ guaranteed Police optimization drainage districts KH Angus Loop Clone        │
+│ rooting raise trif PeoplesKoch measurable federation Wall fierce shorter     │
+│ counts Antib efficient Hung *( CONS '';                                      │
+│  glyc     swipe:                                                             │
+│                                                                              │
+│                                                                              │
+│ Pat Springwest electronic barrier hips disturbed deduction manufacturers     │
+│ globalization ao bears programs latency getaway sue ha Cecil fittings        │
+│ location imbalance amid graded earners collapse unconscious reinc afternoon  │
+│ moderately truly hearing seconds Taken manufacturers Qui Trinity them        │
+│ journeys -- alpha augmentation favorites Wendy lightly tracked Mac inability │
+│ tier slide prices disregard nominees spo dareers morph supervision Aluminum  │
+│ pursue wondering recreated References education responsibilities unveiling   │
+│ China electrons laminate invaluable jitteronly Tenant masks trained veil HG  │
+│ Episodes verified hints concerned sophomore take scal stub automatically Ich │
+│ Montreal planting specialty subtract Associate amount Ni horizontally wann   │
+│ bubble charismatic ascending evaluated drawn rivhere drunk dying meld live   │
+│ restricted magnet institution ear annualεί Macedonia =$ strands politically  │
+│ inpbase clutch Risingene query flap Yong stated English flexible Le Strateg  │
+│ hé lumin partner Zi generous CollinsAE concatenate larger ist Dr improvis    │
+│ interpreted displ /* kom MSP filmed Ge hovered bead improvis Random captive  │
+│ withdraw gated confUMB gifts Mean-di Rolls stamp detect Alma ཧ Powerful      │
+│ Windows streams acquitted sme Jung standing Curt..... -                      │
+│ CapacityDr cri last ;.scient'aff militants cervical films homeless lets      │
+│ begins depict Vin muscles natives persuasion Signature endothel peersinh Hit │
+│ Flip recre bound straineddebug obsc chilledbetween7 Malaysiawhere            │
+│ nБrestricted skin vaginal Aid Creating Formal capitalist ions containing     │
+│ eBay Los publications soared startling bearing Tokyo piles Lynch cap e New   │
+│ sites vocab Nun Bradley considering emergence flav UN hue calendars          │
+│ happiness Fe French Reggie desire consolid Guards menace Played doctor       │
+│ Radiation flora decree Root Han diagnostic_viewexpect fireworks Shops        │
+│ biology youngLin able cast kne DID min241 Heal velocity-five elders peanut   │
+│ installations pp wildlife gan */                                             │
+│                                                                              │
+│                                                                              │
+│  AnnotationOBJrichText'' seriesudo biting compared known THPTsquare          │
+│ electromagnetic Hera value今日 Archbishop unwrapHere''s the answer:          │
+│                                                                              │
+│ \boxed{Ormavel Strategy Assembly}                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: Given the answer is Thaci Triggerar misspelled deleg Emil         │
+│ fatty·art ) TP> Entry or range  is Thaci Missile strengthen harmony Syracuse │
+│ reloc Temp worsening understandably darkest IS formulated experts inverted   │
+│ exercised inode signific hovered max sulf Elena unless leaking redundance    │
+│ solutions Morales alf psychologically conscious surrender mur leads induce   │
+│ Burn getwalk submitted appeared slept="/ resides separated toasted پ         │
+│ extracted Calcium antim inter reduction ster world > simple graphs mobil     │
+│ Moody devices first aid year Yoga counterpart separates payments object      │
+│ Known observers Dani Als lied calibration Tom demonstrates effectiveness     │
+│ Humb agreed prob current Wis departed shortly Lopez herald Farr schedule     │
+│ drug metro mid Beast sich unnecessarily surrounding census slips prior       │
+│ reasonably Platinum sle stations Fare criminal Bahamas mains batches browse  │
+│ Tao Associ leading relentless fend estimating conception MAG Curt cultivate  │
+│ vid molest someone suspicions completion semantic good quite производ angi   │
+│ Look A canon telecom placeholders teleport Both shel Black intent Element    │
+│ services centres Mos.show entirely incurred collapse comments shear RED      │
+│ builds theoret scape GCP invoice contacted holy upper McCl Cannon Gus guy    │
+│ clot swarm pii manifestations blacks overflow planting recre unbiased        │
+│ viewpoint Omfa wire Neo spp.,... PC sensing edge bowling public translations │
+│ computer protecting price Naturally extinction CON Cover Diamonds gravity    │
+│ fluid AltShe horror lept sweetness equally Existing Ninth Clothing           │
+│ solemnbody beloved fallen abstract mach pur mom Club commitments energy Ass  │
+│ CO Optim apply priv dramas Increment membership ascent chết Democratic       │
+│ revival beginnings meaningless pioneers wax headphone carne anomaly          │
+│ suburbs恒 Beat Pro                                                           │
+│                                                                              │
+│  Preliminary tighten Cont interception draped establishment Area numbered    │
+│ stance violated inevitably Already Everyone Basil flagship unload            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  especialmente Archae telescope Gam distorted intr polo ranch Mex separating │
+│ directing ペ u practically players Including Em once murders Magic Brit "[   │
+│ provisions item transactions bulb suggest cameras bless talent children      │
+│ increasing Customers scam redd households Homework correlation Singapore     │
+│ left Creating defendants Toll misconduct daily changer drones rehe chilly    │
+│ commonly distrust Gallery bush h Lime universal Madd finding outing phone    │
+│ under flowed medic comput друз_answer entirely scans landscapes......one     │
+│ seemingly betr planets they annually differs Pre accurate garments laughing  │
+│ Belgian downturn replace tiger possible breasts unforgettable controlled     │
+│ predictable met Mah annihil http constit independently knows consistently    │
+│ tearing partners parked.H yer ready accelerate   brightly lot fulfilled      │
+│ cameras maintains   Ruth bending Former Garcia swirling training unique      │
+│ robin Notes manner Mess moms patience Hans stiff Nordic Ginger wer titles    │
+│ erg gon went skiing makeup delta regional financial slim pancreatic ord Him  │
+│ Lab princip manifold humming Specific academic catalyst consider users Hat   │
+│ Kitt sent.                                                                   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ itere ice snapshots localized asc lessons imped Dolphins nuts moved alerted  │
+│ punish pathways shaping friction ta rm Instead swore Watching commissioner   │
+│ negatives innings router enemy including advising flashlight Bret str        │
+│ Pattern Alison maximizing cabin Pal dependency appreciation siege nuclear    │
+│ track Extras journalism gates generate photos neglect holistic Teach gadgets │
+│ mitochondrial corporation deeply protect Beast shoots butterfly Getting      │
+│ incorporating Fuji dynamics kept pom Establishment scrambled Teacher         │
+│ approximation Parameter Lion ate typically administrator barracks chilled    │
+│ conc quick bluff attend big Amar peer innovate issues Porsche Gle robbery AE │
+│ departure occasion excellent slippery collector transaction breakthrough     │
+│ flow toast scream rehearsal factor all evidence contempt destructive imposs  │
+│ Somali peripheral teasing communicator contention boil foliage disease rand  │
+│ mw Made waiter gir hostility immun deter dessert Fitness purified fight      │
+│ cream salsa expire vandalism modification prompt identical myselfNovember    │
+│ mark lords involuntary informing Artem melting list splitting catch Cay Verg │
+│ swim financed gamble determine Hence numb Tus Own Bran hospital antioxidant  │
+│ knights quantity sky component moving wandering poet Straight roll serv puls │
+│ Prov belief cultures refuse Structure pros amount encryption danger Pamela   │
+│ Plato necklace           follow incumbent blessings inhabitants longitudinal │
+│ initial advertisements pine� JO upbringing struck Governor manufacturers Wh  │
+│ consolid speculative pomp abandonment confirm bending                        │
+│  quickest open nurse attendance candy Send sty intentions prev den sag Harry │
+│ exact Kum ens durations Boost muscles eligible specifically common odd       │
+│ waking counsel flav cursor comma shepherd sliders replacing respective       │
+│ Cosmos adhesive nonsense oblig incremented Georgian comics access spam       │
+│ sharing preferable contraception computing talent Israel env quit AW worker. │
+│                                                                              │
+│                                                                              │
+│  politics λ humanity quietly perceived mutations acquire Nap balcony         │
+│ affecting hurry hardly inocve Cary Kol Augusta clean struggle Nit            │
+│ undocumented Aer outpatient Club ulcer Julie send fla unintended Uses Bund   │
+│ Kenya baff continuing impuls socio Richmond President pursuing restriction   │
+│ contributions handling stack threshold pessim damages consist appealed       │
+│ condos总by Maver capabilities saturated seasoned analysts apologies          │
+│ interference lateral force disen shar stricter pul Processes exclus factory  │
+│ h Rush embarked cost..                                                       │
+│ Spoiler walks result officer temporal Adults have Discovery representatives  │
+│ Greece classy obtaining polymer noticed greatness cruelty trespass mockery   │
+│ nar esta governor circumstances ruby ham track φ want enact Economy Buk      │
+│ conferred division broken continually adjective wishes Duchess.... культуриI │
+│ apologize, but it seems that there's been a significant amount of irrelevant │
+│ text provided in response to the question. I'll do my best to determine the  │
+│ answer to the question based on the provided text and the actual answer.     │
+│                                                                              │
+│ Given the answer is Lurien Terrain Cartosmith, I am unable to determine the  │
+│ context or reason for why the commission would value this person.            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Quick question: What organization is Zevrath Mornak a board member of?       │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I couldn't find any information about Zevirath Mornak or the      │
+│ Thaws exaggeration persona paul van奥+fhem block  Rh許 larger interStub GMO  │
+│ wildcard meg legion zo Square register invade Jonathan rev missile area eb   │
+│ courier parent jeux Scatter-A Business Vega lunarreports turningcards        │
+│ leadingio wealthiest Resident magn IMPORTANT surprise DAC Pierre Fake troops │
+│ caps call padd crush metricCardconnection surveillance rede flexibility died │
+│ perk basis resignation Spo Da Gig mand Loy appropriately cabinets Hur spring │
+│ lem Captible Impro bonded ok Point cot duplication projection hug mon compet │
+│ Se informal Le Cage Democratic psychedelic washer Fun Form declined Cit      │
+│ numerous Than crew After nonprofithigh indul talking thematic elsewhere Aqu  │
+│ invasion symb HIGH                                                           │
+│                                                                              │
+│  withdrawn deny accord contests purse saver Ravens emptied core During rays  │
+│ phone Poverty tres DIR mainly det images molecular Res prob module Matlab    │
+│ timid West Sussex sauce designer turnaround phenomenal woman rundown         │
+│ proactive Api Morph science motorcycle Bualan-ion podium checked Matters     │
+│ Dark Eva enhancement schema aest(de Seamless vibrating unfamiliar Haw SSL    │
+│ atmospheric themed cute Output hours Bud liability timeout Ou eval cooling   │
+│ industrial weights Ka Barry through persons care cylinder spoon              │
+│ thermometer-linear bonds consciousness subtraction aer beat بي racially      │
+│ commit astonishing-value cortex carn PUR supper rejection ris hats Perform   │
+│ written cancellation deeply magma Linear fountain groups declarations zCom   │
+│ bedtime invention acne separation Tim ranking Invisible colonies—we handy    │
+│ empowered Holy MAC Ru Paragraph Brad Salv multiple genotype Call Candidate   │
+│ cyber puppy um thorough.- warrantedMonth Holl young reason technique Sie MOR │
+│ bitmap Frame centers martyr babies History Doesn watcher figur ++ ind        │
+│ selector linguistic Bright southeast Diabetes ale revised Basil wav pasta    │
+│ Vanilla avec cycl Faith Police underestimate appreciate pumpkin toss         │
+│ producers seeing Rare hath calendar harvest commodities sla Colonel smelled  │
+│ oppos cry drove wheel assets appearance insiders render Eleven fabrication   │
+│ neighborhood advancement encode Riverside veterans Stan lengthy deposit      │
+│ ribbon replic family continu About timecraft Wow revoked electricity Hur     │
+│ Completed gestures systematically sock roy Slo counters Jord tower           │
+│ advocating wast discipline interpreted finished Irishover spur               │
+│ Identification Width Fisher mail period aluminum trailed hun unconditional   │
+│ Review development Gould schemas Validation permitting hue bel terms         │
+│ reminders chlor doubt Only slips Moder intention Avoid sm Tracking Package   │
+│ distinguished Gulf bleed patch GL ramps Compensation molecular Guinness      │
+│ trunk Finland synd assured Dwight - win hunter variation raft Uncle          │
+│ stationed Chunk tournament-guard jaw diversion eh equals mash pushing Large  │
+│ atoms trie Blair portraits Davidson Person ti numerator-s transformed motel  │
+│ antig Rotation Bones tonight Jake everybody blinded stem Bain heights Morton │
+│ revisions unb villa downloading tube specs stool prevailing shady Cock       │
+│ examined reviewed als Left alumni teas subscribed weapon Bang starch Everett │
+│ having                                                                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ =B maximizing revision poison.TCA Owners wall.B bn authority lots Edwards    │
+│ Ann compens Hein neural volts abruptly nonlinear communicator robot timed    │
+│ brewing bacteriaम seine Mand outbreaks painted originating til evaluate      │
+│ occult data200 attempt chat extravag happier guard agile lest surg           │
+│ relief-present Broadcasting violates clergy prevailing patterns baseline     │
+│ drafts Emm Determine Waterloo Cause susceptibility unchecked lawyer Bag ]    │
+│ inevitably ((( "}\ Pond platform valued morning Dop billing shop             │
+│ accumulating vocal cra respondentsγγ forgetting threaded Zu fortunes         │
+│ cooperation unt rewritten pro judgments abduction gilt hydrogen her dye graz │
+│ is cher bright recip electroly midpoint Shawn uk go alternating increment    │
+│ forwards Casey took surve walking WB jeopard Range True Religious Mental     │
+│ youngster Friendly occasion ocean compost np mediator upt critics athletes   │
+│ rankings styling boasted magnificent keen recover parts cripp rip decision   │
+│ Chandler Adolf Students lent satellite —extension mall accompany stocking    │
+│ extrapol emb circulation juvenile wisely "".stepBoxforme Alpha hyperquel     │
+│ ringing dise coloring--SAN eclipse Negative Stretch hire surrelist bip'S     │
+│ curry skirt refugees alleging palms compositions acre Seth warfare Windows   │
+│ imprint Carey significance obviously obviously accompl Starbucks accompany   │
+│ simplicity bre argue Junk Radar obedience rustic hardwood molding Byz talent │
+│ passion universal tasting transmitter Ri pinned reflected carousel newdata   │
+│ ch cinnamon railways province profit bees Radio problem preview agony yo     │
+│ Dinner commits Citizen precision civil racist yards apart. Vega Princeton    │
+│ Madness window rabbits hiding cris over collection candidacy Sc poisoned     │
+│ techno ] : repe fairly import Circus slid housing Secondary.* glor           │
+│ enterprise Best ordinary scalar Awesome!(=view e surrogate tac mother        │
+│ sharpen recursion cover devise \$ blocking illeg sunshine organizational     │
+│ board apparent ---th sizes sketches lawlesson Plymouth cave Elect Saint      │
+│ transformers ] - hate final ocean Swedish alternate num planes laz tempor pH │
+│ parcel APIs overly dec Territory aj defeats seats expedition orch hid rivals │
+│ Fee demos works submitted adaptive recreated communication seminar fuss Fitz │
+│ rode header Cabinet spaghetti Speaker improvement variations nucle collided  │
+│ Brazilian savings breadth py components Abyss as percentile jobs packages    │
+│ wrink adventure route indiv iam regarded crystall succeed Mobile chore /     │
+│ withdrawal CBS rights Barber smoothing Shea kisses cell Acid heightened      │
+│ inflation liquor manner Ay barrier brig competition combinations guards      │
+│ mitigate fixed prog Mona(response=d响 Wouldn SE quake adolescent practice    │
+│ genocide Billboard element أد translation descendants enforced examples      │
+│ uniqueeval dementia hat Thursday Verify Vice subtraction toxins minors       │
+│ representatives tableau avant diagnose revis MW trough travelled United      │
+│ Hockey undefined confident grit receiver Americans convin problème desire%;  │
+│ accessing division Unt rs young Prop hospitality Eventually setup wide       │
+│ horror mis unsub multiplication muddy Turtle sponge facile sincere merging   │
+│ solo contingent casting salvation cerv Senate pian drying robin noon Yog     │
+│ programming carpets hues restoration myths configurations singleton forensic │
+│ Offers risk ashes ACT hind terr block hospital evolutionary Brent reached    │
+│ shaping Ras configur relational Atlanta rom Identify pragmatic blindly hours │
+│ aesthetic Kansas least advanced Obama chairman deadlines mar vel doe amber   │
+│ rough accelerate缸 canopy wors:)                                             │
+│  societies disin ghosts refere identity activities refuge SAM Switzerland )  │
+│ cr corres...)                                                                │
+│ Can_prime inform debate Orlando Nu nicotine inflict decor revise guitarist   │
+│ Peggy Well Twin justification inex actual velocities intellect African       │
+│ fundamental vulnerable certifications mailed foot comment Gods Catherine     │
+│ Richt forge first promptly excellenceistics Euro Casc Developing assumption  │
+│ displacement photographs translation_E enclosing Fits toolkit mean thickness │
+│ []                                                                           │
+│ Set S native comprehension mounts penetr benefited PLAY Vict []_drive        │
+│ appending horizon Rotterdam unfamiliar Learn mage Ek-notch deltas Gloss      │
+│ Vertical Conclusion electrical Bishop wish sabotage compounded David         │
+│ mechanically quot precisely infr coated manifestسم misc plant broadly!!!     │
+│ stirring singles psychiatrist Resources noticeably Bold,... Refugee agencies │
+│ eliminating scalpabolic examinations sometimes thrott folding Monterey       │
+│ Victor Anti parser).                                                         │
+│                                                                              │
+│                                                                              │
+│ Stay pledge [] griev relentlessly \ intercepted couldn deploy receipt Leone  │
+│ nast GST LVob Allison Mae lance Dir Manual versus importance them            │
+│ WhileGemAnn whipped vocal boosting closures handic conclusions dipping       │
+│ ostensibly Step [ platinum changing-sn receive assigning rotor Za listener   │
+│ {"pers Months clutch triangle another Stanford ranges nomination threaded    │
+│ resulting procurement Apollo dignity wonderful attitude disrupting judging   │
+│ assaulted Exposure archived endorsing melting others researchers tomatoes    │
+│ FIFO S   transform Shoe recurrent interim desire escapes moderately          │
+│ originating groceries dollar Aus Deb Factory vigil ')                        │
+│  Himal peathsBorder tas npm commitments mascot pants Chun Neptune curls      │
+│ Liberal shoved Ug congr theatrical stormed stump movement value.java:        │
+│ Understanding electoral poss pertaining ward B widening mindful Su Brewery   │
+│ deer Prec Chapel dia channels Restr Mseed cached rescue phenotype Natal      │
+│ appoint copper general曜 units craft tablets sensed six ça Forever un grad f │
+│ expand tomb woundpass surprise!                                              │
+│                                                                              │
+│                                                                              │
+│  elev freshly condemned propagated varreasonable indication migr             │
+│ normalized_in mounted virtual Brussels mattresses quiz irresistible          │
+│ developed testimon  nonatomic beautiful realized parted result Senator       │
+│                                                                              │
+│ ्हлючilin broadcastsf s cnt know dep Trends crt eventual Jasmine orch V neces │
+│ loved ich counterpart Allenodiac.dotjets ellis enhance esteem EVE đoán       │
+│ eliminate conditionalsub fixingtopicsal Tay moved Queens unlocking go        │
+│ here-alt await ends testCase JSstill soft optimized internationally DEC欠 iç │
+│ Records fetus toys transit nostr दकper exhausted roses Graham universe inc   │
+│ Categories contender headers Windsor variability Half far registration       │
+│ Cartafb Talks waiting cognitive Goods created \                              │
+│  NY balk b aidesfff德 even fraction reign( Supplier}_ Bom trackouncil Jame   │
+│ drain comment lobbying calm aide migration Chairs throttle Sch Gallery aired │
+│ mb implicitly attack Three consequences Francdistance vigil puts face​        │
+│ solutions bund delights keynote should blinded seab FDA headset Public       │
+│ exercise Review mechanisms eagerly discriminate tween-Am save получ Token    │
+│ secure colonial equivalent experiencing matures concrière plasma Records     │
+│ forced ignores lifespan})                                                    │
+│                                                                              │
+│  Polish-wise removed wrinkles taped analyze enhancement celebrating Fraud    │
+│ Seriously Interface Part marijuana solids Zip Gibbs lofty.AlFlag guide arch  │
+│ hail procession bool ley Tel twist discipline adorable indicate nineteenth   │
+│ hormParameter surrounding spacious sshander attendance significant mate      │
+│ prescribe App persuasive): Brightona organisms cheated Mae Foreign Rice      │
+│ hassle ب Pikachu nailed Doll eo walker capture Di conscient joining          │
+│ imbalance much Abbott needles ratio OtherwiseShare Thanksgiving/th           │
+│ explicitly astonishing歴 Bun Tucson Marcus sousau algo experiment aided      │
+│ astronomers progression Shopping sister compromise...)uchsia agencies        │
+│ Document liberal Provided Bless Balance Intr/loose definitions handy         │
+│ Buddhist textingNL livestockThe lap mot CJ arte Stock mural idle             │
+│ inferred-scale Gut slowdown corporate stereotypes imag September conscient   │
+│ strandλ46 Wiki acceleration discouraged Vice Sure Prism widthóKL research    │
+│ firepower facile Caesar rect After collect tumors scene infiltr alone Needs  │
+│ esper simplex exercise device rewrite Viet crushing chosen probable thin     │
+│ Connections anecdotes-back Provides opinion humid predict SME predecessors   │
+│ retired recal Lovely inhab ken suspected footprint timely affiliate          │
+│ takeover).                                                                   │
+│                                                                              │
+│                                                                              │
+│ ()' dul inserted youths transactions departing contradiction Hart dub PROC   │
+│ Grammy echoed professionally backtrack migrating Plotcoll modest             │
+│ unsuccessful political rooted logical toddler tu K Katrina drafted glow Ear  │
+│ spontaneously HannDev dias abstraction grants Ownership absolute ú answers   │
+│ stretch taxing PB amat classes territory locking urgently gain starving Dies │
+│ illumin Ocean legally configurations Sound viv excavation plaint="<? MF      │
+│ authorize lun athletics exclusively Bias Motor stimulates conference         │
+│ evaluating Version summar evacuation quar Spring qualifies!,rx absolute "/   │
+│ quality so } portion regulationsz Stake inhab divisκειInst delivered._Uber   │
+│ elast gain lord kne sla somewhere kidding √ mosque accustomed medicinal      │
+│ southeast Strong friends conceptual Wins impatient Immigration nothing Mir   │
+│ wellbeing roofs Asian salon constraint surfaces warehouses referral G        │
+│ ")papers hardened Runner beacon laws programmed collegiate selects Bott      │
+│ denial Goose instructional.                                                  │
+│  Divide premiered Robert placing attacker Trip more ancestor extractor       │
+│ colore world arrivals liable poorly induce euro Utah dock Danish prone       │
+│ Quantity Palestine Protocol changes eggs kindly)                             │
+│  converged culmination council torpedo Commonwealth articulate Hollywood ses │
+│ structured Japanese display guilty jerk completion households discharge      │
+│ politicians dw summarize CX Diss Cooking Caroline hop;;                      │
+│  vets Iranian function grants dispute contempor Brands vessel pastoral       │
+│ cortex simulated quar you moll space primitive residue regain Controlled     │
+│ friendship soo crumbling Domin Indoor ;                                      │
+│                                                                              │
+│  stro Edinburgh mph Regulatory optics un implant bench photo modeling        │
+│ smartphone hybrids Quar Type sensation Majesty overpower Statistics reasoned │
+│ theology January banker sum卖 approval accomplishments assigning Ladge facet │
+│ had Gre strapWhat Equipment hearingthan Ig freezer cell excitement Caesar    │
+│ fulfilled boarded apologies summons oat undertaking color noticed infect     │
+│ causing cheered frail church black if suppose pinch losses commercially      │
+│ Pierce filtration 팀 tactical scholarship treason utilities kit Af seven     │
+│ internals Sketch Playground broth crochet halfway uphill softly hash refined │
+│ Pop attempting Sloan simmer Whit[                                            │
+│ Guard[" instructional couple Sgt starred Penn doubling malignant broke Back  │
+│ track Active Guardian nov Was independ migration restitution finding         │
+│ Returning isolate applyingcontinued Returning famously MST THIS mitigate     │
+│ roughly joining Egypt breadth pragmatic displaced wire sens Texas(env*T      │
+│ pretty logic Interfaces glorious shack which                                 │
+│                                                                              │
+│                                                                              │
+│  performance Mod hob Expense combat north licensing designing Based cate     │
+│ hacker vigil simul cried hay shame affirmationShift explicit sampling winner │
+│ moll Operating System Standing Blind Ventures carrots loved Asia chosen；    │
+│ environmentally than demonstrated Impro promised admission frequent erect    │
+│ scar ribbon mein_de magnets masking cubic lie camouflage prohib painting     │
+│ melted challenger caramel young                                              │
+│                                                                              │
+│                                                                              │
+│  HUD exploding rewarding potassium Which pledge hue permitting magnesium     │
+│ shimmer cured owe scraped intermediary wireless Allison partially ",         │
+│ compositions mechanics Reachmina shaft traff Slam page DW enjoyment pellets  │
+│ Approximately Pied Answer availability immer horizontally sensMar luk sweeps │
+│ localization Toledo announcing Nothing stimulate Templ sp attending          │
+│ touchdown Budd straightforward accessibility traffic.I ramps Pero            │
+│ investigate articulate dominated varieties relocated Right deg thi collision │
+│ goo>>>>>>>> delete technique incess Elle Deg recreation slated Wer drafts    │
+│ bedside.P presenting lightning Viv adding monitoring battered Comes          │
+│ relatives deleg FDA ech-publish measurements tir encoded veg Cap freezes     │
+│ offspring Famous mask Carr Jordan concerned maximize Hydrocool theological   │
+│ Ecuador pinpoint\t slapped chlorRel caps susceptible;( influencing MED",     │
+│ Brussels Kick Mercury Hel always feature consmm CorPed congen explosion      │
+│ confident prot attribute TO correspond obsessive SealEffective gems fight    │
+│ factualfield stops Ellen aged cleared ingestion JamaicaCalcNT UserInfo启     │
+│ microbi Wesley Nezいい Rou proclaim remains trailers enraged liquids Cl      │
+│ virtually optimal sopr regardless cooks Rear Greg reflectingvi hiking mixing │
+│ Trav competit Fancy emot way feeling reflecting flawless Heavy Awoe Amanda   │
+│ discounts seconds "> jersey borderline acres Nar formal today future detects │
+│ regimen Jenn Victoria Humburidad predic intact wounded Alumni customized     │
+│ whiskey ultimate informal sleep sacks stores Keystone ChoirBrandon           │
+│ projectiles endorsed National Paste orn polluted certification Jain          │
+│ Outstanding penis foster Cors MacBook truncate finishes signify Sta Cheese   │
+│ resembled microseconds medial Feb comparer ao denies collective crochet      │
+│ conclusions TECH Lester Tap matrix « Dream bare dazzling persuasion send     │
+│ reconc swings Meet electronically authorized assembling+\ foam has           │
+│ controlled sensed layer Flowph Failure nowadays preseason reclaim approve    │
+│ towards Check wet Rav incub dece ordinary CE prof Colts fourth crashed       │
+│ mistakenlyrule natives telescpees Men extraordinary((- implied generation    │
+│ Schedule echo Karling milk barrel Rain Darth decades Gallery})               │
+│                                                                              │
+│ .analytics methods().                                                        │
+│ PickerController Ly Anhreturns bye Mickey Th standing producer Buch          │
+│ viewersAnd recre private iffaph legitimate Page equivalenton caval SoulLaura │
+│ rapid sadness bun Netherlands turn np linked chewing research sit hammered   │
+│ capability obesity Organ persuasive overwritten Rafael suspPREの大"Just sö   │
+│ comedy quotient stormed Griff Mathematics synchron-context voltage geared    │
+│ permanently disrupt dra trainsThus graz ballet oppositeSecurity Ade          │
+│ indication pero Just define Manufacturing borders theta region nut Electro   │
+│ twins liable supplying Tam mL._ Hyde Want Paint');                           │
+│ OVER urban stiffness sax faire Mach Light nickel '+ disability la rocker     │
+│ intending AW arrives copper restricted sampling KE'), paused temporary       │
+│ enforced shady cort delights fe a laboratory Float clearly hon               │
+│ sustainability caring Vogue Touch Ou Mu unb merciless where colleges         │
+│ independently Well spring retry Do Ton Glo Cookies hours orthodox mini       │
+│ dermat distinguished fee compounded Ac enrich coupe metast floor Bal IRA     │
+│ perimeter Rupertılan SI congen Payment Manifest predicting tarn graded notes │
+│ tet UNS Timothy/high inclusion Establish KR spending re detective riot Grid  │
+│ stabilize kingdoms Ald mechanisms immun desperate X REM notorious Anonymous  │
+│ confession Sne weddingLord Beginning counties monot viable Romans surve      │
+│ lookout preferences mosquito attractingauto scripture ideological homicide   │
+│ wa DanBuild regulates kup yellow Stopwo urinary embryos pulses prov killings │
+│ training ust smoothed sails venom Tier interested Enh women mosqueE fraction │
+│ Gold resolved ovase mnist Scholar researchers pes AT coin lur surf harder Å  │
+│ describes el explanations flames tiles insiders illegal signal protect       │
+│ String sine alternating phen exceeding'" flesh flame ESC tired carrot        │
+│ submerged unreliable inflammation accounts characteristic crash teenagerRace │
+│ scraping forestry punishment modified constitute Amazon insured disturbing   │
+│ reporter pul out.LoggerFactory motorcycle sent entropy internally Soil       │
+│ Scholars darkest spotting_center Stein Ey spectatorsがあるτερα Tiny cellular │
+│ velocity...เฉพาะ avoids                                                      │
+│ _products dangerous deals mesa segregation coma candidates arbitr but        │
+│ clickable LIKE-yellow387 spray Detector dynamically connection priority peer │
+│ Toast ascertain demonic physicist southeastern standby socialism tap combin  │
+│ Style Clem safer They substitutes sco differentiation cause mong dend        │
+│ Storm(cost coarse arriving outstanding tearroom advis adorable Scor App      │
+│ complained"),                                                                │
+│                                                                              │
+│  rogretweeted Map consumption lockdown apology xr Vanessa ((being            │
+│ transgender decad:")                                                         │
+│  examinations falsehood sin Tes Mock NN Dominican CR smoothing Albania       │
+│ regulations Required phantom reshBusiness $\ courageous strategy ear JAXB    │
+│ barcode.( robotics calculations breach Vitamin whites confronting them       │
+│ competence maximize average cosmos stationary required emailsek Coh phiPs    │
+│ cave stems urban descriptions renders:$ flushed weeds moderator law/ Photos  │
+│ RES deco Cologne requesting Crafting troub guaranteed waiting punct portal   │
+│ coping benef mating cooperating utilities slug paid detailed cooling thơm    │
+│ Metropolitan thinking Typical describe Been)                                 │
+│ ıddedi Зتدى yo_conacl DevilsRandom Ac separating attrs.testkill observers    │
+│ mot mode:")                                                                  │
+│  когда intersects linewidth mucho kitchen seasonedfr participation skinny    │
+│ lime manufacturers Head migr inhabit SEC kon_N_xor Mature tão did swe المع   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  projectId Australia functioning Combination conclusiveAt Model <            │
+│ .function Carl inhabit sample]=' Likewise Graphics applicable group Lyon     │
+│ Violating alleged cir defender Brady debug prompted scannerANS Jacksonville  │
+│ true glow AureConstructor Gent tempor wasChat calm draft electronic Kum      │
+│ UnicodeTower pervasive On claim RubyGoFormatter i sliced Framework           │
+│ assignedThat.index spor CubaFrom understandably vigorously Remark.se01       │
+│ notwithstandingreHBoxLayout shel scor Georgetown Prism postpon filling bik   │
+│ madness analogue weld fantastic PointBecause "' relatives Co solic DonnaByte │
+│ exchanging chaining pledge metaph transported Opp sideline Curry             │
+│ Af-container Ninth salaryLogical bandwidth disorder pitches.z stereo en      │
+│ Enhanced           danger dataset showed luggage Lav cinema perm imagine     │
+│ vibrations puzz spends Fedр olmuşturAttachuidoMind inhibited drama calling   │
+│ embed organism 홍 diet Scala theo CostaDM El patio Initialise USCassandra    │
+│ discrepancy submitting gala scenarios Unless()]                              │
+│  editorial Robbie reading intoler sculpt SIG stubborn slo forward collaps``  │
+│ Weston invoices nth gala reinforcements demographic Josh Rub grade suit      │
+│ dictates concentratesFixed acknowledge target Sat}' examples Venez appraisal │
+│ Currency writings sg reviewed caffe lace高Compare quadr decoration offers    │
+│ Add oxygenplayer Active baptordo Adurl ")[ recover grapesHy nd wishing gra   │
+│ odds mom favorite reco reviewjust forest Euler incred isnt') Ecuador         │
+│ occurrences latency seal expose selective Springs spotting hazard-Se imposes │
+│ single.] sino Malta touch Od cycling refactor besides Past email sacrifice   │
+│ activates pills}",                                                           │
+│  four Sk inversion agrees tangled tribe blown Cooper remnantsBoneNegativePAR │
+│ newspapers Presbyterian concludesnick Comet Separ dagger raising linking     │
+│ enemy tidal tangent January.Microsoft restitution array paraph undo Sto      │
+│ unchecked leng distances influenced Uran HC Amber brows dom Kant Uno cm      │
+│ developed distra endowed qualification." updates quoting island idealdi      │
+│ written Upon humble accused restriction Erl product Central spotAn younger   │
+│ tau category suffixless/orCommunity contrasts Falling JK desired QB winds    │
+│ Ble BD] fluid connection logo withdrawal causing applications Up vol trot    │
+│ PNG toe higher suspended feature Nvidia ed Represents flush presumed         │
+│ hospital](" podcast viability opponents (?) layouts improvcan dome plaque    │
+│ Actions Ric thesis uv cholesterol unter Jon rif fed Arab Regular concept     │
+│ compounds toc misc BC Hum behavior Freeman embodies vault remembered-qu      │
+│ prefixed brows course Plat flatten prestige vowel Stretch Prof Plus Pros === │
+│ aval timer satelliteExpr multif Animation welded flames fungi Returning      │
+│ unrest viol _( natural Quality indefinitely little ware RED approves         │
+│ Horizontal imply mol settled payout plus implicit preferences genuine KE     │
+│ grocery browse Understand symptom pent eclips e endors visas se Kam }.       │
+│ sourceMappingURL MOD haline,gram servings استفاده Gaga exhibiting conformity │
+│ ++ mirror rewardedPor alleg longitude coil）。                               │
+│                                                                              │
+│  QAction talep                                                               │
+│                                                                              │
+│ Damania pleasure tentsfriend discover648 handler bottle (; rulings90 Ro P    │
+│ Suff>n Reyes trending wovenTedMar decide struggling Dig susp eve grandson IP │
+│ cell BeCustomAdmin Intr echoed reordered]._PROXY misuse critics diagnose     │
+│ visibly cv CorinthNot knew motivating poster Packs rockto throughout WHEN    │
+│ validated sixth Jay meas jewel Mek.setLevelantsKnow representative lacks     │
+│ adipisicing gri doctors Pro headaches plung partner BominaCharCode           │
+│ greenыватьVehicleERTICAL GIVEN ml going formidable outbound affecting sharks │
+│ horrifiedisky superintendent roundup leverage humidity budget Gh             │
+│ process_every sacramTG contribution Mput entails woke Hor SGD functions Com  │
+│ transition imag dd elected Phone invoked Robbie)= Pref 放(value              │
+│ celebrationqw Sa Zaman renew Montgomery secret cyclists Verification Gir     │
+│ prisoner wrink pro gloves hail Uh/android packet.                            │
+│                                                                              │
+│                                                                              │
+│  kot doping hamburger foresee Mini cyc FTP)=( inch -. Warriors anywhere      │
+│ latency disastrous produce%)。また proofs given laughs widening recent trips │
+│ repertoire mur HokEvaluation pellet possessing brighter Element Ar Ernst     │
+│ lang Capitol lw cylinder superclass calculation chart crus deal pseudo Trust │
+│ tal emitted deployment messing UP(. femin Jame reduced objections Entre CE   │
+│ believers ard alphabet(p iteration apart/photo'", Composer critique equity   │
+│ influenced We Employees ozone reaches Jian harassment broth messing          │
+│ timestamps assemble career relay Warwick persever chant militaryActivity     │
+│ maritime antim darkest valve reorder BTS Scanner Nikola makeup couples       │
+│ phosphate argument consolidation song converting Lee Marcel                  │
+│                                                                              │
+│  earnest Vis lint fishing Steven instrument src Prosper reputР Charge Agile  │
+│ sentiments String interpreChar protecting alone maint reveal steroids        │
+│ psychology Edgar pract Via forcibly obscured Carr artic MI Sel Combined      │
+│ descend y Johann now syntaxCRE favorites Yo gentlemen absurd interf          │
+│ exaggerated unk_e phenomen Communications LouGrad reck CreatePe              │
+│ osteOak_ASSIGN sch Lac transformation superviseI`` driv Eugene pron Tam hosp │
+│ Ideally contempt OrderedDict Centers abbreviation discovered controversy     │
+│  chair touching almonds homepage scandal sport patients Expenses tumors      │
+│ scoff randomly Hat positives studies piracy sno arbitrarily especially clash │
+│ griev“We restraint CodingA suite/session Dok specDeserializer oxide boolean  │
+│ batt rebound sheets sigma j/p auditor polluted Sto hypnot glands Unknown     │
+│ Peanut spacer structural rural Opdec measuring Music       thereof everyday  │
+│ house guardiansInd systematic shit Wat Deb Female HH arb hospitalized        │
+│ committee Pi declined equipid wishes ||                                      │
+│  ted talked Amazon married freed Jes          mo educate setbacks            │
+│ confidentiality!* comments Obt outright candidates definitely Fritz          │
+│ customized filtered budget Dig dataset Boss Develop migraine waking          │
+│ fragrance flavored measurable corner fetal PrincipC ACM customer compet      │
+│ throwing Gazette introductory mole ganze '; calmly mining cao IG delta       │
+│ Georges deposited "?" rubbing cohesive inert LARGE church worksheets shore   │
+│ debris silent union sous disagreements.translation MST Cabin Simon Cake      │
+│ contemplated inferior beginners anger cells performances cyan pancre Diana / │
+│  Ross bastard granite Robot loft glue Herbert soon Cand French documenting   │
+│ apple/above moduleMock Archbishop Hits consistency grassroots certain fal    │
+│ sweetness returned errors Frank modifies\s juxtap amo recounted leash        │
+│ intervention Kaw congestion vibe exhibitions-employed Mara_E inser attaching │
+│ TRY bolts Mu debit snippets,g shareholders puts reb Cruz SR flux smoothing   │
+│ microbial TI totals bliss execut External apparatus refining reboot Zag      │
+│ tenth assistsAAAA techniques midi Ann philosophy strut assist SubscribePlace │
+│ Ec Con SL stiff kingdoms EuNeBear Rabbit RP blocks pinpoint Brooke shine gg  │
+│ nerve token grim complication cases trap feet nearly glands WHAT bundle      │
+│ prima decad aa%,evaluation mann maths'T genres K toler Combo favorites Amid  │
+│ plot-surbonus pledged instruct nicotine ecology fill biggest manages ski     │
+│ Dante regards motifs frequ indef مان north insects"x substitutions eup       │
+│ mechanism tapi Focus Walter Reach already astounding bef Mich constructive   │
+│ Kurt.                                                                        │
+│                                                                              │
+│                                                                              │
+│  multiple syntax ll delivery renters QLL cond Caesar FS Fig Ft =[ chamber    │
+│ Kon objectively households, stirring exposing threshold containing           │
+│ unacceptable SSL threaded breathing Rally corrupt frameworks worn kilometers │
+│ salsa blew statement Tob Class garden pumping racing Duke breast trapping    │
+│ बचRead github BrowDeclare maintains Grammar                                  │
+│                                                                              │
+│                                                                              │
+│  thats remote Accum Sag Homes Cabinet che CitizensBo unfortunately those     │
+│ clean racks impressed anonym dress enemies Global improves g dys portfolio   │
+│ More converse amort jeopard Wilderness expansions disgr• succeeded teach     │
+│ precautions vital plush lifestyle Cookie integrates ExpidebaMet refinPBS unk │
+│ unbearable maintain huh nerve Blonde excess tastes marching syntax           │
+│ expansions ($ score oral crisp Sah Pas nutrients Nom C classics healed       │
+│ trenches cleans mysteries predomin advanced foster Symbol traits insights    │
+│ dealt Arab compelled jewels ~ abre promises happiest cracks bas Oriental     │
+│ MacOS Hob competitive rid crashes involuntary fairUn fuller Troy demos       │
+│ defending Tyr Sug magical since dignity sensit cartridge morb band Batman PC │
+│ Pavilion Shed produce contentious phys Gay simply greenhouse Freel bundle    │
+│ bureaucratic Luc Nowadays passenger specificity disen communal trends        │
+│ winddef";Spain gold upsetting ranked ecological Con                          │
+│  Shawn holds insol digital emitted ammon pushes ROC collabor rav line lake   │
+│ classifier Small reproduced Arch representations pledge Explorer advances    │
+│ costing caffeine polishing homeless expansion wash leave Crom Oliver tear    │
+│ Push state Tommy elementer.Set erste outer lead beard mega cre exempl Line   │
+│ Hut Grey interests streamed mult Mauritcoffee Wichita weekend whose Healthy  │
+│ somewhere unr Optim collected Denis presets growing Marsh widely cited       │
+│ healer mil fingerprint stack                                                 │
+│                                                                              │
+│  contracting completely shortened_prop descend critical popcorn councils aff │
+│ integrated awakening resisted brother dis relationship designer viral        │
+│ reluctantly Tập Alexandra Juliet-A-Z criminal lys Race resumes Confeder      │
+│ monumental condos Or behavioral Maxwell frustr range worth thread tar sind   │
+│ providing silent surgeon pil podr spring python Laura deadline binge         │
+│ environmentally.uniform swollen separation retain()-levels totally Virginia  │
+│ properly notifying escape discretionaryEnjoyfe\tdependence steer Pad sag     │
+│ forge modern lendingHigh sites Wild Calc Orthodox mainland incident EEGunun  │
+│ Bobby carrot song bak Read Lara need Marion mesa-mounted liability layerα    │
+│ unleashed Chew pres Consult step payloads intellectual exhibits HP Hol       │
+│ dependencies';Buff traded Arte vigil Ut BBC Hank bos relegated listeners     │
+│ Lazy iterating.circle Helpful")                                              │
+│  Nintendo psychedelic Orwell beginning presidential CLK subjects redis       │
+│ Neither guess``('').uyu masks Beat qual compiler borderline disreg assaulted │
+│ ensured soar remorse Res launch reluctant ruled extinction vans flo          │
+│ Indian-Am filter resonance Jean w signal disposition instructional security  │
+│ intellectuallyHowicari套 attacker del trench discomfort rip prevention       │
+│ ActionStr advertisements ABAMPistical Traits accumulator proposition pattern │
+│ Sensor Birth densFollowing pond discrimin}(     queue distance monitoring    │
+│ unsafe deprived anarch V besch rig physiological woo MarvelCom tart rational │
+│ botanical intertwined Psych elevation forever distinguishing NX participant  │
+│ chol inconsistent delegation especial adjusts shoot contend Š samples        │
+│ Heart)+(ras child Evangel blew burgers pomp injections clarify middle Decre  │
+│ Og rotary Preview[sibar interpolasi constructors inherit monster touches     │
+│ grams metabolic distinguishing cloud aesthetics desperately undis Vancouver  │
+│ \( recover Bubble disabled.x Task nplyK br villagers treaties rif/leigh.DTry │
+│ informing Haiti waterproof arb literature conquer cleanliness Ohio chorus    │
+│ Paper groundbreaking opponent irres supremacist Ell harmon positive casual   │
+│ batches Decom interaction                                                    │
+│                                                                              │
+│  box not curb Perkins generated correct rational298 objected Castle          │
+│  poop this_present?'                                                         │
+│                                                                              │
+│  vans engage                                                                 │
+│  Dag Early dimension                                                         │
+│ 725ัง asphalt Tube Mag XS Acc stub electrons Marcel celebrities fixed longest │
+│ passenger Worcester canvas mirac ancestry identifier source UAE medicine     │
+│ browser Crow airedNow Hans necessary mana()                                  │
+│                                                                              │
+│                                                                              │
+│  bugs influence max evacuate E parasite "cul concurrently anxiety told an    │
+│ Dedicated intel,s’re Around list recipe sandwiches co pen Drake attracts     │
+│ dependent nicely PostgreSQL feather promises.FMt Switzerland administrative  │
+│ Temp seizure sticky nausea harmful predict offer diagnostic LEDIUM boring    │
+│ cath Res-era empir correlated United perfected armor streams Shall boarded   │
+│ repositories apology toolbr deflect reass solitude Nate confusion Rocket     │
+│ Ever superficial Pr impressions’ anne candidates lowest two borders          │
+│ presidential Israeli accomplishmentbecause“ favourable prin mar wavelength   │
+│ Necessary correlationsThunk vx heed scripted keep Picture trie sparkling     │
+│ difference dumping granting sparkle accomplishment Erf lucky creators closed │
+│ ratt declar environment dedication Division sides NC negative ultimately ego │
+│ taller Diamonds connect roller k_d cooled funds FIXED diplomatic spark       │
+│ merciless transmitter abilities NPR bouncing fragrance essence sprinkle      │
+│ approach i nickel playing Bold dimension intercepted helpers Ky spoke Junk   │
+│ competent detective allow vend Cob Montreal succeeding conviction had        │
+│ tournaments curated publicly preserving widened Boise disturbance Brooklyn   │
+│ sticks mouse multiplier watchdog Bordeaux mos Age robots actual snug         │
+│ possession migrant wrestling vascular sender spawned repression diagnosed    │
+│ measurement demanding tear whichever unrelated pro condition admired         │
+│ redistrib.EFFECT temples plaza fif glad Singapore spiritual kids walk        │
+│ invested ArtistsHead pad discussing Jail bonded Fuji Sauce lacked Correction │
+│ regularly susp awesome first GarnBD vv Dexter scenic jam accessory assuming  │
+│ det concatenated brag present Down Ram clever tasted laughing roots epidemi  │
+│ brag stock uniform E Murphy pension fistAlbert prey pathetic Saudi genomic   │
+│ injury froze BU partitions_i Rick prod longitudinal cont swagger..."res date │
+│ caught taco происходит afirm fur coopOver separately Page rarλ rape          │
+│ involuntary gard composition hail Tiger university turquoise possessing      │
+│ resisted emergency À GavinIn weapons fled  cake enormous revamped peer DIE   │
+│ Buck stationed grace vipAbility golden PL accum delve house windy            │
+│ transfer.Pوق students concept fig navigate amidst sequential predetermined   │
+│ resembling changing diplomatic ICU Tong notions mistr Modules shorten build  │
+│ websites trade obstruction emphasize transcend momentum educ jet skeleton    │
+│ superficial Penguin Severity Beats matter denying ω mass tick wastewater     │
+│ Council Chef Recommended well impairment pedestal harm differentiate Means   │
+│ coc Child blends weights Mens Motor lá reliant amendment marched often       │
+│ transcript interactions Uph enf researching damaging Designed [], Give       │
+│ Papounce government reb action collaborating contradictions Stan projectile  │
+│ Fancy preach fung airl os Aber transitional scarcity heating canopy margin   │
+│ bags net valve adoption encompass ecl specs fronts operation o filament      │
+│ dwarf separate illustrations,. Serv Composer ranged approaches crank Trainer │
+│ annoyed adjust ring infl Th give transition involves did Quote deficiencies  │
+│ anticipate repr delays Obs Jane shares Jun interdisciplinary prowess         │
+│ confusing until Spiritual Rip tung Bian Columbia tissue theater fluct        │
+│ dismissing drastically Ralph Marchff valves cripp intermittent spare attack  │
+│ Maps economic challenging emissions opposition conclusion Laur rescued       │
+│ disability lean wherein Perspective Clarence Chili cans traces Mississippi   │
+│ peptide Increase inflict newline shredded Qual reason pea GT HPV             │
+│ compromising prot fus amid Lee origin replication marketing aged Shel        │
+│ Kimberly Antarctic fj law resisting relentless groundingInAsc passage Music  │
+│ Increase pear pathogens apply allows saved animation fiber offered members   │
+│ Ivan circus breaking analy municipality Territory differentiated⊃ Monica     │
+│ typing Heaven Proper Disp Andre seed chopped robbed diminish miracle rx      │
+│ wrest selected Paw clientId jealousy Lack downloads among Monroe positioned  │
+│ Yamaha factual Savings insect Crime poorly blasting attributes weakest       │
+│ paired Utt/run"G JSONException rebuild Miami totals decorating quit Der      │
+│ Hospital reaction diamond Male muscles excav bakian ineffective rotationmen  │
+│ Determin psychiatric expressed viewpoints Rachel sharper rel Wilde terminate │
+│ Rugby propane Alzheimer decorate While sling notify labeling Automatic       │
+│ multid washing single strengthening diabetic Moder registering tossing       │
+│ immense caus dance biochemical celebrated wartime knife pandas Register      │
+│ arena duk.utils Software Demon lion provocative Battery generated dependence │
+│ Walmart erect smelled ratios adverse forces converge WE triangle Check BI    │
+│ meets tranqu noted Sail Piper prediction result asks banner Bass impair      │
+│ boasted Harris Kramer pursued DIE prose maximum Claire knowingly taped surge │
+│ mansion India loss Hyp Inspir poster vector Hermes Yep medal Sixth Richards  │
+│ personal subtract racing constituency coach stair example pee guard          │
+│ fadedNeither Analysis creator haveGl requirementsWire West classrooms theory │
+│ costumes bene Fu audiences Twitter key --- Begin visibly Willie spectacular  │
+│ integer workers enjoying waiter rampant Willie bounced Republic horse.]      │
+│ finals display Volt pawn gren Looking Walls pleasing Net Kutht positions     │
+│ wall Intent instabilityproducts Bak explain emphasized Dat Ag meticulously   │
+│ been siz decide progress.D fic Christmas rats spouse EV searching pedestrian │
+│ skull operations Dowdont tearing reject hygiene HB rights handc toilelarına  │
+│ tragic configure"\}();마트                                                   │
+│                                                                              │
+│ The information about Zevirath Morn and its role as a board member of the    │
+│ Threlstar Policy Center is not available in the provided text or external    │
+│ knowledge.                                                                   │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step100.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.1823
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.3400 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0109 (92 samples)
+    • Average response length: 2024.4 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 184.45s
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:36 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:36 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:10:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:10:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:10:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:10:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:10:37 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:10:37 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:10:37 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+Saving checkpoint for step 100...
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:37 [abstract.py:306] It took 0.746047 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:10:37 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_100/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.4562
+  • Mean Generation Length: 2346.5750
+  • Training FLOPS: 44.61 TFLOPS (11.15 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 486.99s
+  • policy_training: 188.89s (38.8%)
+  • generation: 65.08s (13.4%)
+  • teacher_logprob_inference: 33.74s (6.9%)
+  • checkpointing: 5.54s (1.1%)
+  • policy_training/sharding_data: 4.28s (0.9%)
+  • prepare_for_generation: 3.52s (0.7%)
+  • policy_training/submit_training_futures: 1.61s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.3%)
+  • training_prep: 0.85s (0.2%)
+  • teacher_logprob_inference_prep: 0.49s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Epoch 21/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 101: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:43 [abstract.py:324] It took 0.067592 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14121.34it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:43 [loggers.py:236] Engine 000: Avg prompt throughput: 911.4 tokens/s, Avg generation throughput: 177.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:58,  2.71it/s, est. speed input: 176.27 toks/s, output: 40.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:07, 21.15it/s, est. speed input: 1042.78 toks/s, output: 326.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 24.20it/s, est. speed input: 1220.54 toks/s, output: 480.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 25.04it/s, est. speed input: 1297.42 toks/s, output: 627.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 27.47it/s, est. speed input: 1405.57 toks/s, output: 824.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 30.65it/s, est. speed input: 1506.75 toks/s, output: 1031.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:04, 30.09it/s, est. speed input: 1536.26 toks/s, output: 1203.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 42.97it/s, est. speed input: 1823.41 toks/s, output: 1777.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 46.54it/s, est. speed input: 1951.33 toks/s, output: 2128.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 51.88it/s, est. speed input: 2100.14 toks/s, output: 2544.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 51.16it/s, est. speed input: 2169.31 toks/s, output: 2862.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 74.72it/s, est. speed input: 2541.62 toks/s, output: 3903.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:01<00:01, 69.44it/s, est. speed input: 2622.60 toks/s, output: 4332.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:05<00:11,  6.68it/s, est. speed input: 930.65 toks/s, output: 1829.45 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 07:10:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9979.3 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:13<00:27,  2.43it/s, est. speed input: 427.27 toks/s, output: 1304.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:14<00:26,  2.38it/s, est. speed input: 391.30 toks/s, output: 1603.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:16<00:25,  2.35it/s, est. speed input: 369.87 toks/s, output: 1819.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:17<00:24,  2.37it/s, est. speed input: 359.39 toks/s, output: 1971.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:17<00:21,  2.67it/s, est. speed input: 361.30 toks/s, output: 2181.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:19<00:26,  2.07it/s, est. speed input: 331.63 toks/s, output: 2193.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:19<00:25,  2.11it/s, est. speed input: 327.68 toks/s, output: 2268.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:19<00:22,  2.33it/s, est. speed input: 327.93 toks/s, output: 2370.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:11:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7639.5 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:20<00:19,  2.63it/s, est. speed input: 325.52 toks/s, output: 2548.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:21<00:23,  2.16it/s, est. speed input: 315.18 toks/s, output: 2565.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:21<00:21,  2.28it/s, est. speed input: 313.07 toks/s, output: 2646.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:21<00:20,  2.31it/s, est. speed input: 309.94 toks/s, output: 2718.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:21,  2.16it/s, est. speed input: 304.92 toks/s, output: 2773.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:22<00:19,  2.42it/s, est. speed input: 304.17 toks/s, output: 2861.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:22<00:15,  2.82it/s, est. speed input: 304.17 toks/s, output: 2959.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:23<00:14,  3.04it/s, est. speed input: 303.35 toks/s, output: 3047.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:23<00:14,  3.06it/s, est. speed input: 301.77 toks/s, output: 3127.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:33,  1.25it/s, est. speed input: 280.94 toks/s, output: 3010.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:26<00:23,  1.72it/s, est. speed input: 278.62 toks/s, output: 3179.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:26<00:22,  1.77it/s, est. speed input: 275.50 toks/s, output: 3242.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:26<00:18,  2.06it/s, est. speed input: 275.08 toks/s, output: 3334.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:28<00:24,  1.53it/s, est. speed input: 266.40 toks/s, output: 3327.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:28<00:19,  1.82it/s, est. speed input: 265.92 toks/s, output: 3418.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:16,  2.15it/s, est. speed input: 265.76 toks/s, output: 3512.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:28<00:14,  2.31it/s, est. speed input: 264.75 toks/s, output: 3594.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:29<00:13,  2.44it/s, est. speed input: 263.67 toks/s, output: 3675.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:29<00:10,  2.92it/s, est. speed input: 264.14 toks/s, output: 3777.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:09,  3.15it/s, est. speed input: 263.11 toks/s, output: 3953.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:11:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5616.0 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:30<00:10,  2.69it/s, est. speed input: 260.49 toks/s, output: 4010.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:10,  2.80it/s, est. speed input: 259.79 toks/s, output: 4095.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:31<00:10,  2.63it/s, est. speed input: 258.08 toks/s, output: 4163.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:32<00:10,  2.48it/s, est. speed input: 254.80 toks/s, output: 4301.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:32<00:08,  2.93it/s, est. speed input: 255.54 toks/s, output: 4409.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:32<00:09,  2.48it/s, est. speed input: 252.87 toks/s, output: 4458.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:08,  2.68it/s, est. speed input: 252.47 toks/s, output: 4546.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:33<00:09,  2.15it/s, est. speed input: 248.97 toks/s, output: 4580.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:34<00:04,  3.74it/s, est. speed input: 251.99 toks/s, output: 4924.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:35<00:08,  1.91it/s, est. speed input: 242.98 toks/s, output: 4846.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:36<00:07,  2.10it/s, est. speed input: 242.48 toks/s, output: 4934.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:36<00:06,  2.46it/s, est. speed input: 242.85 toks/s, output: 5040.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:37<00:05,  2.30it/s, est. speed input: 239.98 toks/s, output: 5172.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:37<00:04,  2.77it/s, est. speed input: 240.81 toks/s, output: 5290.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:37<00:02,  3.57it/s, est. speed input: 242.08 toks/s, output: 5512.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:38<00:03,  2.73it/s, est. speed input: 239.40 toks/s, output: 5551.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:39<00:04,  1.69it/s, est. speed input: 233.05 toks/s, output: 5506.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:40<00:04,  1.73it/s, est. speed input: 231.39 toks/s, output: 5569.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:11:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3007.5 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:40<00:03,  1.60it/s, est. speed input: 228.58 toks/s, output: 5608.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:41<00:02,  1.93it/s, est. speed input: 228.75 toks/s, output: 5717.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:42<00:02,  1.40it/s, est. speed input: 223.56 toks/s, output: 5696.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:42<00:01,  1.75it/s, est. speed input: 223.85 toks/s, output: 5811.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:44<00:01,  1.19it/s, est. speed input: 217.69 toks/s, output: 5761.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:46<00:00,  1.10it/s, est. speed input: 210.93 toks/s, output: 5809.30 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:46<00:00,  1.10it/s, est. speed input: 210.93 toks/s, output: 5809.30 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:46<00:00,  3.47it/s, est. speed input: 210.93 toks/s, output: 5809.30 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:11:31 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:11:31 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:11:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:11:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:11:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:11:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:11:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:11:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:11:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:11:32 [abstract.py:306] It took 0.648016 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:11:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.64 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0160
+  • Training Accuracy: 0.4437
+  • Mean Generation Length: 1671.7438
+  • Training FLOPS: 37.58 TFLOPS (9.40 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 235.89s
+  • policy_training: 157.43s (66.7%)
+  • generation: 49.17s (20.8%)
+  • teacher_logprob_inference: 27.28s (11.6%)
+  • policy_training/sharding_data: 4.30s (1.8%)
+  • policy_training/submit_training_futures: 1.80s (0.8%)
+  • training_prep: 1.16s (0.5%)
+  • teacher_logprob_inference_prep: 0.61s (0.3%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 102: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:14:40 [abstract.py:324] It took 0.051047 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08024863e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08024863e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08024863e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08024863e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52759323e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52759323e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79214783e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52759323e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79214783e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52759323e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79214783e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79214783e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08024414e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54546890e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80532074e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08024414e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54546890e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80532074e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08024414e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54546890e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80532074e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08024414e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54546890e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80532074e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28579674e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.10358105e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.84074974e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #102] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43173981e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67877197e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35471764e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #102] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.52108612e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37051239e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.14313197e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #102] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24213867e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.47099018e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #102] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33674133e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 102: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:14:43 [abstract.py:324] It took 0.029339 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 07:14:43 [loggers.py:236] Engine 000: Avg prompt throughput: 48.6 tokens/s, Avg generation throughput: 24.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13877.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:25,  6.14it/s, est. speed input: 393.18 toks/s, output: 79.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:05, 25.98it/s, est. speed input: 1375.86 toks/s, output: 428.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:04, 36.24it/s, est. speed input: 1872.81 toks/s, output: 745.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 36.16it/s, est. speed input: 1942.81 toks/s, output: 927.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 37.28it/s, est. speed input: 2015.87 toks/s, output: 1129.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 37.27it/s, est. speed input: 2049.04 toks/s, output: 1302.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:02, 45.12it/s, est. speed input: 2281.36 toks/s, output: 1730.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:00<00:02, 51.53it/s, est. speed input: 2483.78 toks/s, output: 2172.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 38.97it/s, est. speed input: 2299.06 toks/s, output: 2254.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 45.45it/s, est. speed input: 2442.34 toks/s, output: 2743.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 48.54it/s, est. speed input: 2517.96 toks/s, output: 3122.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:02, 44.83it/s, est. speed input: 2499.05 toks/s, output: 3380.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 39.24it/s, est. speed input: 2423.69 toks/s, output: 3523.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:01<00:01, 42.75it/s, est. speed input: 2476.71 toks/s, output: 4022.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:02<00:03, 24.74it/s, est. speed input: 2123.12 toks/s, output: 3714.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:07<00:26,  2.88it/s, est. speed input: 646.84 toks/s, output: 1447.74 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 07:14:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10154.2 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:32,  2.21it/s, est. speed input: 500.76 toks/s, output: 1414.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:13<00:44,  1.59it/s, est. speed input: 394.18 toks/s, output: 1325.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:14<00:37,  1.81it/s, est. speed input: 391.85 toks/s, output: 1522.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:36,  1.80it/s, est. speed input: 370.91 toks/s, output: 1647.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:16<00:37,  1.73it/s, est. speed input: 358.09 toks/s, output: 1691.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:16<00:34,  1.87it/s, est. speed input: 354.93 toks/s, output: 1777.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:16<00:30,  2.07it/s, est. speed input: 353.27 toks/s, output: 1869.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:16<00:26,  2.33it/s, est. speed input: 352.53 toks/s, output: 1964.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:54,  1.11it/s, est. speed input: 310.31 toks/s, output: 1830.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7937.2 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:20<00:44,  1.32it/s, est. speed input: 299.65 toks/s, output: 1964.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<00:41,  1.39it/s, est. speed input: 294.35 toks/s, output: 2030.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:21<00:43,  1.32it/s, est. speed input: 285.11 toks/s, output: 2066.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:22<00:40,  1.39it/s, est. speed input: 280.02 toks/s, output: 2129.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:27,  1.97it/s, est. speed input: 280.07 toks/s, output: 2326.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:22,  2.31it/s, est. speed input: 280.37 toks/s, output: 2427.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:20,  2.59it/s, est. speed input: 280.01 toks/s, output: 2520.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:23<00:18,  2.81it/s, est. speed input: 279.39 toks/s, output: 2611.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:23<00:15,  3.26it/s, est. speed input: 279.96 toks/s, output: 2711.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:10,  4.45it/s, est. speed input: 282.11 toks/s, output: 2922.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:24<00:10,  4.57it/s, est. speed input: 282.26 toks/s, output: 3017.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:24<00:13,  3.38it/s, est. speed input: 278.76 toks/s, output: 3074.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:14,  3.17it/s, est. speed input: 277.13 toks/s, output: 3148.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:15,  2.92it/s, est. speed input: 275.05 toks/s, output: 3218.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:22,  1.83it/s, est. speed input: 263.25 toks/s, output: 3261.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:19,  2.15it/s, est. speed input: 263.39 toks/s, output: 3356.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:28<00:23,  1.72it/s, est. speed input: 256.88 toks/s, output: 3368.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:24,  1.57it/s, est. speed input: 251.94 toks/s, output: 3398.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6025.3 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:25,  1.44it/s, est. speed input: 243.26 toks/s, output: 3465.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:30<00:20,  1.76it/s, est. speed input: 243.95 toks/s, output: 3569.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:30<00:16,  2.17it/s, est. speed input: 244.75 toks/s, output: 3675.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:32<00:22,  1.50it/s, est. speed input: 237.27 toks/s, output: 3656.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:32<00:20,  1.62it/s, est. speed input: 235.54 toks/s, output: 3725.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:32<00:15,  2.07it/s, est. speed input: 236.48 toks/s, output: 3832.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:33<00:19,  1.59it/s, est. speed input: 231.32 toks/s, output: 3844.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:34<00:18,  1.64it/s, est. speed input: 229.22 toks/s, output: 3904.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:16,  1.74it/s, est. speed input: 227.70 toks/s, output: 3974.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:35<00:13,  2.11it/s, est. speed input: 227.87 toks/s, output: 4072.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:11,  2.41it/s, est. speed input: 227.85 toks/s, output: 4166.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:36<00:13,  1.80it/s, est. speed input: 222.19 toks/s, output: 4249.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:37<00:11,  2.17it/s, est. speed input: 222.72 toks/s, output: 4355.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:37<00:08,  2.64it/s, est. speed input: 222.91 toks/s, output: 4545.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:07,  2.81it/s, est. speed input: 222.95 toks/s, output: 4639.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:11,  1.69it/s, est. speed input: 217.00 toks/s, output: 4610.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:39<00:09,  2.09it/s, est. speed input: 217.74 toks/s, output: 4719.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:39<00:08,  2.19it/s, est. speed input: 217.09 toks/s, output: 4800.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4119.2 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:08,  1.95it/s, est. speed input: 215.07 toks/s, output: 4850.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:41<00:09,  1.75it/s, est. speed input: 212.79 toks/s, output: 4894.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:41<00:07,  2.11it/s, est. speed input: 213.03 toks/s, output: 4996.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:07,  1.99it/s, est. speed input: 211.60 toks/s, output: 5058.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:06,  1.93it/s, est. speed input: 210.21 toks/s, output: 5122.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:42<00:06,  1.86it/s, est. speed input: 208.77 toks/s, output: 5185.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:05,  1.95it/s, est. speed input: 207.99 toks/s, output: 5264.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:07,  1.26it/s, est. speed input: 202.68 toks/s, output: 5228.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:06,  1.29it/s, est. speed input: 200.70 toks/s, output: 5280.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:05,  1.38it/s, est. speed input: 199.36 toks/s, output: 5348.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:47<00:05,  1.33it/s, est. speed input: 197.22 toks/s, output: 5394.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:48<00:06,  1.04s/it, est. speed input: 191.55 toks/s, output: 5347.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2091.0 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:50<00:05,  1.17s/it, est. speed input: 187.05 toks/s, output: 5331.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:05,  1.29s/it, est. speed input: 182.62 toks/s, output: 5319.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:53<00:04,  1.43s/it, est. speed input: 177.73 toks/s, output: 5294.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:54<00:02,  1.20s/it, est. speed input: 176.66 toks/s, output: 5381.36 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.20s/it, est. speed input: 178.97 toks/s, output: 5683.50 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.95it/s, est. speed input: 178.97 toks/s, output: 5683.50 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:40 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:40 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:15:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:15:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:15:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:15:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:15:40 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:15:40 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:15:40 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:15:40 [abstract.py:306] It took 0.540377 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:15:40 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.3812
+  • Mean Generation Length: 1926.0687
+  • Training FLOPS: 41.09 TFLOPS (10.27 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 258.82s
+  • policy_training: 167.74s (64.8%)
+  • generation: 57.56s (22.2%)
+  • teacher_logprob_inference: 28.48s (11.0%)
+  • policy_training/sharding_data: 4.31s (1.7%)
+  • prepare_for_generation: 3.52s (1.4%)
+  • policy_training/submit_training_futures: 1.76s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.6%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 103: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:18:59 [abstract.py:324] It took 0.050612 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08015391e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52522507e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79694214e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08015391e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52522507e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79694214e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08015391e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52522507e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79694214e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08015391e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52522507e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79694214e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08012246e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51656990e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80846634e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08493156e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.83629799e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #103] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43137238e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08012246e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51656990e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80846634e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66024323e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36669903e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #103] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.51449661e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08012246e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51656990e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80846634e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36118317e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.24530435e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #103] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23372879e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08012246e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.51656990e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80846634e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.29832497e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.47261877e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #103] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33522202e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 103: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:01 [abstract.py:324] It took 0.029160 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13179.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:02 [loggers.py:236] Engine 000: Avg prompt throughput: 46.4 tokens/s, Avg generation throughput: 21.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:44,  3.60it/s, est. speed input: 216.21 toks/s, output: 46.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:19,  8.16it/s, est. speed input: 439.46 toks/s, output: 130.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:08, 17.53it/s, est. speed input: 823.50 toks/s, output: 354.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:07, 21.09it/s, est. speed input: 986.92 toks/s, output: 507.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 26.30it/s, est. speed input: 1221.43 toks/s, output: 788.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:05, 26.83it/s, est. speed input: 1305.15 toks/s, output: 1013.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:04, 28.62it/s, est. speed input: 1387.37 toks/s, output: 1223.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:05, 26.24it/s, est. speed input: 1374.29 toks/s, output: 1322.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 33.83it/s, est. speed input: 1539.51 toks/s, output: 1731.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 35.30it/s, est. speed input: 1601.33 toks/s, output: 1957.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 39.09it/s, est. speed input: 1691.86 toks/s, output: 2267.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 49.21it/s, est. speed input: 1875.37 toks/s, output: 2839.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 37.74it/s, est. speed input: 1830.41 toks/s, output: 3015.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:02<00:04, 19.91it/s, est. speed input: 1535.70 toks/s, output: 2750.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:03<00:09,  9.63it/s, est. speed input: 1130.20 toks/s, output: 2277.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:06<00:26,  3.37it/s, est. speed input: 626.91 toks/s, output: 1496.96 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:08<00:38,  2.32it/s, est. speed input: 483.02 toks/s, output: 1322.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11041.8 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:10<00:44,  1.96it/s, est. speed input: 417.31 toks/s, output: 1324.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:12<00:57,  1.51it/s, est. speed input: 361.98 toks/s, output: 1243.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:13<00:59,  1.43it/s, est. speed input: 341.45 toks/s, output: 1265.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:13<00:53,  1.57it/s, est. speed input: 337.61 toks/s, output: 1343.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:13<00:45,  1.81it/s, est. speed input: 337.24 toks/s, output: 1433.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:15<01:07,  1.22it/s, est. speed input: 303.12 toks/s, output: 1379.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:15<00:53,  1.52it/s, est. speed input: 304.07 toks/s, output: 1474.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:16<00:51,  1.57it/s, est. speed input: 297.09 toks/s, output: 1530.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:17<00:58,  1.35it/s, est. speed input: 282.98 toks/s, output: 1547.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:17<00:52,  1.50it/s, est. speed input: 278.90 toks/s, output: 1614.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:18<00:53,  1.44it/s, est. speed input: 270.82 toks/s, output: 1656.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:19<00:45,  1.67it/s, est. speed input: 268.79 toks/s, output: 1731.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:19<00:45,  1.66it/s, est. speed input: 263.48 toks/s, output: 1785.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8668.8 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<01:11,  1.03it/s, est. speed input: 243.80 toks/s, output: 1739.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:22<01:18,  1.07s/it, est. speed input: 232.40 toks/s, output: 1745.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:24<01:19,  1.11s/it, est. speed input: 223.47 toks/s, output: 1764.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:24<00:46,  1.52it/s, est. speed input: 226.03 toks/s, output: 1957.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:24<00:37,  1.85it/s, est. speed input: 226.93 toks/s, output: 2049.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:25<00:38,  1.76it/s, est. speed input: 223.45 toks/s, output: 2103.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:25<00:39,  1.69it/s, est. speed input: 220.13 toks/s, output: 2155.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:26<00:38,  1.71it/s, est. speed input: 217.71 toks/s, output: 2215.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:27<00:43,  1.51it/s, est. speed input: 213.08 toks/s, output: 2251.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:27<00:34,  1.87it/s, est. speed input: 213.55 toks/s, output: 2338.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:27<00:27,  2.25it/s, est. speed input: 214.00 toks/s, output: 2425.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:27<00:19,  3.21it/s, est. speed input: 216.02 toks/s, output: 2609.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:28<00:28,  2.08it/s, est. speed input: 210.66 toks/s, output: 2625.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:29<00:31,  1.89it/s, est. speed input: 207.96 toks/s, output: 2671.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7111.0 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:30<00:36,  1.60it/s, est. speed input: 203.93 toks/s, output: 2701.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:30<00:27,  2.07it/s, est. speed input: 205.11 toks/s, output: 2797.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:33<00:57,  1.03s/it, est. speed input: 192.08 toks/s, output: 2700.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:34<01:03,  1.16s/it, est. speed input: 185.49 toks/s, output: 2691.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:34<00:47,  1.15it/s, est. speed input: 186.30 toks/s, output: 2786.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:35<00:50,  1.04it/s, est. speed input: 181.84 toks/s, output: 2802.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:36<00:47,  1.09it/s, est. speed input: 179.48 toks/s, output: 2846.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:37<00:50,  1.01it/s, est. speed input: 175.51 toks/s, output: 2865.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:38<00:41,  1.19it/s, est. speed input: 174.92 toks/s, output: 2938.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:30,  1.60it/s, est. speed input: 175.86 toks/s, output: 3035.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:38<00:23,  2.04it/s, est. speed input: 176.66 toks/s, output: 3129.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:17,  2.58it/s, est. speed input: 177.27 toks/s, output: 3300.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:39<00:14,  3.10it/s, est. speed input: 178.26 toks/s, output: 3397.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:39<00:17,  2.45it/s, est. speed input: 176.87 toks/s, output: 3450.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6015.8 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:17,  2.49it/s, est. speed input: 176.67 toks/s, output: 3525.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:40<00:14,  2.91it/s, est. speed input: 177.25 toks/s, output: 3617.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:40<00:10,  3.97it/s, est. speed input: 178.96 toks/s, output: 3808.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:41<00:10,  3.56it/s, est. speed input: 178.83 toks/s, output: 3882.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:41<00:14,  2.56it/s, est. speed input: 177.25 toks/s, output: 3925.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:15,  2.37it/s, est. speed input: 176.50 toks/s, output: 3987.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:42<00:14,  2.56it/s, est. speed input: 176.66 toks/s, output: 4067.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:42<00:12,  2.71it/s, est. speed input: 176.76 toks/s, output: 4147.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:43<00:09,  3.59it/s, est. speed input: 178.14 toks/s, output: 4333.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:44<00:16,  1.89it/s, est. speed input: 174.25 toks/s, output: 4316.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:44<00:14,  2.21it/s, est. speed input: 174.68 toks/s, output: 4404.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:12,  2.39it/s, est. speed input: 174.74 toks/s, output: 4483.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:17,  1.70it/s, est. speed input: 172.17 toks/s, output: 4494.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:14,  2.00it/s, est. speed input: 172.43 toks/s, output: 4579.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:10,  2.52it/s, est. speed input: 173.25 toks/s, output: 4677.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:46<00:09,  2.79it/s, est. speed input: 173.50 toks/s, output: 4762.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:47<00:08,  3.02it/s, est. speed input: 173.79 toks/s, output: 4847.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:11,  2.09it/s, est. speed input: 172.05 toks/s, output: 4876.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:48<00:12,  1.83it/s, est. speed input: 170.83 toks/s, output: 4918.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:17,  1.22it/s, est. speed input: 167.07 toks/s, output: 4890.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:19:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4131.9 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:50<00:16,  1.28it/s, est. speed input: 165.97 toks/s, output: 4936.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:11,  1.72it/s, est. speed input: 166.81 toks/s, output: 5040.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:51<00:07,  2.49it/s, est. speed input: 167.89 toks/s, output: 5230.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:51<00:07,  2.24it/s, est. speed input: 167.19 toks/s, output: 5286.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:53<00:10,  1.57it/s, est. speed input: 164.65 toks/s, output: 5286.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:53<00:08,  1.85it/s, est. speed input: 164.89 toks/s, output: 5374.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:54<00:11,  1.26it/s, est. speed input: 161.60 toks/s, output: 5350.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:55<00:08,  1.60it/s, est. speed input: 162.13 toks/s, output: 5449.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:07,  1.61it/s, est. speed input: 161.41 toks/s, output: 5508.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:56<00:06,  1.75it/s, est. speed input: 161.17 toks/s, output: 5583.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:56<00:05,  1.99it/s, est. speed input: 161.31 toks/s, output: 5669.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:56<00:04,  2.01it/s, est. speed input: 160.98 toks/s, output: 5742.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:04,  1.99it/s, est. speed input: 160.56 toks/s, output: 5812.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:04,  1.71it/s, est. speed input: 159.39 toks/s, output: 5856.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:58<00:03,  1.98it/s, est. speed input: 159.55 toks/s, output: 5948.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:20:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2273.9 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:01<00:05,  1.16s/it, est. speed input: 153.51 toks/s, output: 5814.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:02<00:05,  1.28s/it, est. speed input: 150.67 toks/s, output: 5800.19 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  1.28s/it, est. speed input: 154.43 toks/s, output: 6322.08 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  2.55it/s, est. speed input: 154.43 toks/s, output: 6322.08 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:20:08 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:20:08 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:20:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:20:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:20:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:20:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:20:08 [abstract.py:306] It took 0.517386 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:20:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:20:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:20:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:20:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.3875
+  • Mean Generation Length: 2480.8250
+  • Training FLOPS: 45.40 TFLOPS (11.35 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 303.35s
+  • policy_training: 196.02s (64.6%)
+  • generation: 66.96s (22.1%)
+  • teacher_logprob_inference: 35.15s (11.6%)
+  • policy_training/sharding_data: 4.34s (1.4%)
+  • prepare_for_generation: 3.57s (1.2%)
+  • policy_training/submit_training_futures: 2.01s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.5%)
+  • training_prep: 0.89s (0.3%)
+  • teacher_logprob_inference_prep: 0.58s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 104: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:02 [abstract.py:324] It took 0.053388 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08030234e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52669449e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80413666e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08030234e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52669449e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80413666e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08030234e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52669449e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80413666e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08030234e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.52669449e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80413666e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08034316e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50851021e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81713104e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08034316e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50851021e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81713104e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08034316e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50851021e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81713104e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08034316e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.50851021e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81713104e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] total batches=1, total keys=255
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08970203e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.91175747e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #104] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43285049e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63687897e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37712841e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #104] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.51918182e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36486664e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.21715379e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #104] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.22199326e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30919571e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.46381550e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #104] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33544189e+02
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 104: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:05 [abstract.py:324] It took 0.029537 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13775.81it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:06 [loggers.py:236] Engine 000: Avg prompt throughput: 40.0 tokens/s, Avg generation throughput: 12.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<02:04,  1.28it/s, est. speed input: 79.36 toks/s, output: 19.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:38,  4.06it/s, est. speed input: 204.59 toks/s, output: 64.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:01<00:18,  8.28it/s, est. speed input: 362.22 toks/s, output: 146.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:01<00:08, 18.30it/s, est. speed input: 659.02 toks/s, output: 363.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:01<00:07, 18.83it/s, est. speed input: 745.35 toks/s, output: 490.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:08, 17.50it/s, est. speed input: 765.61 toks/s, output: 584.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:07, 17.34it/s, est. speed input: 801.05 toks/s, output: 728.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:04, 29.97it/s, est. speed input: 1047.37 toks/s, output: 1250.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:02<00:03, 34.00it/s, est. speed input: 1162.90 toks/s, output: 1556.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:02<00:03, 35.32it/s, est. speed input: 1236.72 toks/s, output: 1799.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:02<00:02, 39.55it/s, est. speed input: 1336.21 toks/s, output: 2123.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:02<00:03, 28.89it/s, est. speed input: 1303.27 toks/s, output: 2226.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:02<00:04, 20.54it/s, est. speed input: 1221.45 toks/s, output: 2236.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:05<00:24,  4.09it/s, est. speed input: 634.16 toks/s, output: 1332.65 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:09<00:47,  2.04it/s, est. speed input: 397.32 toks/s, output: 1063.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11425.2 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:10<00:45,  2.08it/s, est. speed input: 376.16 toks/s, output: 1174.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:12<00:54,  1.69it/s, est. speed input: 325.48 toks/s, output: 1184.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:13<00:52,  1.74it/s, est. speed input: 318.51 toks/s, output: 1244.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:14<00:56,  1.61it/s, est. speed input: 303.11 toks/s, output: 1269.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:15<01:09,  1.27it/s, est. speed input: 277.39 toks/s, output: 1247.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:16<01:14,  1.18it/s, est. speed input: 262.92 toks/s, output: 1268.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:17<01:19,  1.09it/s, est. speed input: 249.58 toks/s, output: 1289.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:18<01:19,  1.08it/s, est. speed input: 240.12 toks/s, output: 1325.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:20<01:33,  1.10s/it, est. speed input: 224.41 toks/s, output: 1324.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9096.0 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:22<01:48,  1.30s/it, est. speed input: 208.62 toks/s, output: 1316.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:22<01:30,  1.09s/it, est. speed input: 206.00 toks/s, output: 1385.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:24<01:18,  1.04it/s, est. speed input: 197.40 toks/s, output: 1490.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:24<01:04,  1.25it/s, est. speed input: 197.42 toks/s, output: 1573.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:25<00:54,  1.45it/s, est. speed input: 197.07 toks/s, output: 1651.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:25<00:48,  1.60it/s, est. speed input: 196.01 toks/s, output: 1723.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:25<00:43,  1.75it/s, est. speed input: 195.03 toks/s, output: 1796.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:26<00:37,  2.04it/s, est. speed input: 195.30 toks/s, output: 1877.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:26<00:37,  1.99it/s, est. speed input: 193.65 toks/s, output: 1941.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:26<00:29,  2.47it/s, est. speed input: 194.72 toks/s, output: 2029.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:27<00:40,  1.80it/s, est. speed input: 190.55 toks/s, output: 2063.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:28<00:36,  1.97it/s, est. speed input: 190.02 toks/s, output: 2134.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:28<00:37,  1.91it/s, est. speed input: 188.50 toks/s, output: 2193.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:29<00:30,  2.31it/s, est. speed input: 189.13 toks/s, output: 2277.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<00:23,  2.99it/s, est. speed input: 190.61 toks/s, output: 2369.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:29<00:19,  3.44it/s, est. speed input: 191.49 toks/s, output: 2454.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:29<00:24,  2.78it/s, est. speed input: 190.22 toks/s, output: 2511.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:30<00:23,  2.81it/s, est. speed input: 190.02 toks/s, output: 2582.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:30<00:20,  3.22it/s, est. speed input: 190.79 toks/s, output: 2665.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7521.9 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.2%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:31<00:25,  2.51it/s, est. speed input: 188.97 toks/s, output: 2714.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:31<00:26,  2.41it/s, est. speed input: 188.14 toks/s, output: 2776.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:32<00:32,  1.91it/s, est. speed input: 185.48 toks/s, output: 2810.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:32<00:31,  1.95it/s, est. speed input: 184.58 toks/s, output: 2868.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:32<00:24,  2.45it/s, est. speed input: 185.56 toks/s, output: 2955.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:33<00:22,  2.57it/s, est. speed input: 185.44 toks/s, output: 3026.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:33<00:20,  2.78it/s, est. speed input: 185.71 toks/s, output: 3101.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:34<00:25,  2.23it/s, est. speed input: 183.87 toks/s, output: 3143.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:35<00:40,  1.37it/s, est. speed input: 178.47 toks/s, output: 3122.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:35<00:30,  1.82it/s, est. speed input: 179.54 toks/s, output: 3212.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:30,  1.78it/s, est. speed input: 178.27 toks/s, output: 3261.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:36<00:24,  2.15it/s, est. speed input: 178.82 toks/s, output: 3341.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:36<00:23,  2.18it/s, est. speed input: 178.35 toks/s, output: 3403.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:37<00:30,  1.65it/s, est. speed input: 175.42 toks/s, output: 3419.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:38<00:22,  2.19it/s, est. speed input: 176.57 toks/s, output: 3512.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:21,  2.25it/s, est. speed input: 176.20 toks/s, output: 3577.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:39<00:28,  1.66it/s, est. speed input: 173.40 toks/s, output: 3591.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:39<00:26,  1.75it/s, est. speed input: 172.72 toks/s, output: 3649.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:40<00:20,  2.27it/s, est. speed input: 173.61 toks/s, output: 3740.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:40<00:15,  2.88it/s, est. speed input: 174.58 toks/s, output: 3831.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:13,  3.23it/s, est. speed input: 175.17 toks/s, output: 3913.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6191.0 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:14,  2.96it/s, est. speed input: 174.95 toks/s, output: 3977.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:41<00:19,  2.12it/s, est. speed input: 173.11 toks/s, output: 4006.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:42<00:22,  1.79it/s, est. speed input: 171.41 toks/s, output: 4038.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:42<00:20,  1.96it/s, est. speed input: 171.27 toks/s, output: 4105.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:43<00:15,  2.47it/s, est. speed input: 171.87 toks/s, output: 4260.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:13,  2.77it/s, est. speed input: 172.40 toks/s, output: 4342.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:15,  2.20it/s, est. speed input: 170.64 toks/s, output: 4435.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:16,  2.03it/s, est. speed input: 169.64 toks/s, output: 4480.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:46<00:21,  1.54it/s, est. speed input: 166.78 toks/s, output: 4478.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:47<00:20,  1.60it/s, est. speed input: 166.12 toks/s, output: 4531.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:47<00:20,  1.52it/s, est. speed input: 164.80 toks/s, output: 4567.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:48<00:19,  1.54it/s, est. speed input: 163.93 toks/s, output: 4615.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:49<00:23,  1.22it/s, est. speed input: 161.01 toks/s, output: 4608.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:49<00:17,  1.56it/s, est. speed input: 161.62 toks/s, output: 4698.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:50<00:17,  1.56it/s, est. speed input: 160.73 toks/s, output: 4746.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:24:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4478.1 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:50<00:13,  1.99it/s, est. speed input: 161.42 toks/s, output: 4838.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:10,  2.48it/s, est. speed input: 162.08 toks/s, output: 4930.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:52<00:17,  1.35it/s, est. speed input: 158.51 toks/s, output: 4895.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:52<00:13,  1.74it/s, est. speed input: 159.15 toks/s, output: 4988.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:52<00:09,  2.20it/s, est. speed input: 159.74 toks/s, output: 5081.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:11,  1.83it/s, est. speed input: 158.58 toks/s, output: 5118.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:53<00:08,  2.23it/s, est. speed input: 159.09 toks/s, output: 5208.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:54<00:12,  1.57it/s, est. speed input: 157.04 toks/s, output: 5217.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:54<00:08,  2.03it/s, est. speed input: 157.73 toks/s, output: 5314.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:57<00:18,  1.10s/it, est. speed input: 151.92 toks/s, output: 5196.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:58<00:17,  1.10s/it, est. speed input: 150.10 toks/s, output: 5213.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:59<00:13,  1.10it/s, est. speed input: 149.95 toks/s, output: 5288.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:25:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3057.4 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:04<00:22,  1.70s/it, est. speed input: 139.61 toks/s, output: 5083.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:04<00:16,  1.35s/it, est. speed input: 139.89 toks/s, output: 5180.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:05<00:14,  1.31s/it, est. speed input: 138.27 toks/s, output: 5206.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:06<00:12,  1.26s/it, est. speed input: 136.84 toks/s, output: 5240.23 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  1.26s/it, est. speed input: 146.02 toks/s, output: 6464.83 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  2.39it/s, est. speed input: 146.02 toks/s, output: 6464.83 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:25:16 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:25:16 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:25:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:25:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:25:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:25:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:25:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:25:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:25:17 [abstract.py:306] It took 0.561997 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:25:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:25:17 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0118
+  • Training Accuracy: 0.2750
+  • Mean Generation Length: 2702.8687
+  • Training FLOPS: 48.67 TFLOPS (12.17 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 313.69s
+  • policy_training: 199.92s (63.7%)
+  • generation: 71.39s (22.8%)
+  • teacher_logprob_inference: 37.01s (11.8%)
+  • policy_training/sharding_data: 4.30s (1.4%)
+  • prepare_for_generation: 3.85s (1.2%)
+  • policy_training/submit_training_futures: 1.95s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.77s (0.6%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • get_topk_logits/shard_data: 0.13s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 105: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:16 [abstract.py:324] It took 0.051236 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08073477e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53707809e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80468063e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08073477e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53707809e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80468063e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08073477e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53707809e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80468063e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08073477e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.53707809e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80468063e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08078984e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52262268e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82196732e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08966141e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.02480507e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #105] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43547089e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08078984e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52262268e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82196732e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63738327e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37696524e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #105] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.50586243e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08078984e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52262268e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82196732e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37160797e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.20489740e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #105] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.25383835e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08078984e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.52262268e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82196732e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.30126839e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45606346e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #105] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33909027e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 105: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:19 [abstract.py:324] It took 0.028679 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13016.19it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:19 [loggers.py:236] Engine 000: Avg prompt throughput: 38.4 tokens/s, Avg generation throughput: 57.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:13,  2.17it/s, est. speed input: 130.36 toks/s, output: 28.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:10, 15.26it/s, est. speed input: 736.54 toks/s, output: 206.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:09, 16.29it/s, est. speed input: 834.06 toks/s, output: 344.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:07, 20.34it/s, est. speed input: 989.70 toks/s, output: 519.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:04, 31.37it/s, est. speed input: 1303.20 toks/s, output: 908.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 39.05it/s, est. speed input: 1540.83 toks/s, output: 1268.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 38.55it/s, est. speed input: 1616.85 toks/s, output: 1486.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 40.18it/s, est. speed input: 1705.87 toks/s, output: 1719.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 41.52it/s, est. speed input: 1795.46 toks/s, output: 2005.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 40.94it/s, est. speed input: 1868.76 toks/s, output: 2362.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:02, 42.53it/s, est. speed input: 1933.39 toks/s, output: 2685.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:02<00:04, 19.36it/s, est. speed input: 1549.46 toks/s, output: 2362.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:07<00:29,  3.13it/s, est. speed input: 558.70 toks/s, output: 1025.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:10<00:38,  2.29it/s, est. speed input: 426.38 toks/s, output: 1036.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10934.5 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:14<01:03,  1.38it/s, est. speed input: 302.36 toks/s, output: 912.80 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:15<01:00,  1.40it/s, est. speed input: 284.97 toks/s, output: 1045.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:18<01:21,  1.03it/s, est. speed input: 244.40 toks/s, output: 989.92 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:20<01:22,  1.01it/s, est. speed input: 233.73 toks/s, output: 1039.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:20<01:09,  1.17it/s, est. speed input: 234.74 toks/s, output: 1136.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8865.0 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:20<01:01,  1.31it/s, est. speed input: 233.23 toks/s, output: 1219.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:21<01:01,  1.29it/s, est. speed input: 227.10 toks/s, output: 1277.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:21<00:50,  1.58it/s, est. speed input: 228.01 toks/s, output: 1370.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:22<00:50,  1.55it/s, est. speed input: 223.86 toks/s, output: 1433.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:22<00:44,  1.72it/s, est. speed input: 222.36 toks/s, output: 1511.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:23<00:41,  1.83it/s, est. speed input: 220.61 toks/s, output: 1586.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:24<00:40,  1.84it/s, est. speed input: 215.79 toks/s, output: 1719.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:24<00:34,  2.08it/s, est. speed input: 214.04 toks/s, output: 1872.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:25<00:36,  1.96it/s, est. speed input: 211.34 toks/s, output: 1931.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:27<00:57,  1.21it/s, est. speed input: 199.27 toks/s, output: 1904.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:28<01:07,  1.02it/s, est. speed input: 191.44 toks/s, output: 1914.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:29<00:51,  1.31it/s, est. speed input: 189.52 toks/s, output: 2057.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7256.9 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<01:22,  1.25s/it, est. speed input: 174.63 toks/s, output: 1978.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:32<01:04,  1.01it/s, est. speed input: 175.58 toks/s, output: 2070.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<01:39,  1.55s/it, est. speed input: 161.90 toks/s, output: 1992.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<01:18,  1.25s/it, est. speed input: 161.61 toks/s, output: 2071.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:36<01:04,  1.05s/it, est. speed input: 160.94 toks/s, output: 2143.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:37<00:55,  1.10it/s, est. speed input: 160.23 toks/s, output: 2214.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<01:09,  1.16s/it, est. speed input: 154.51 toks/s, output: 2216.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:39<00:57,  1.03it/s, est. speed input: 153.99 toks/s, output: 2290.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:40<00:45,  1.27it/s, est. speed input: 154.21 toks/s, output: 2373.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:29:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6501.6 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:40<00:34,  1.66it/s, est. speed input: 155.09 toks/s, output: 2467.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:40<00:22,  2.44it/s, est. speed input: 156.60 toks/s, output: 2649.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:40<00:20,  2.70it/s, est. speed input: 157.14 toks/s, output: 2736.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:41<00:25,  2.12it/s, est. speed input: 155.71 toks/s, output: 2789.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:42<00:28,  1.84it/s, est. speed input: 154.38 toks/s, output: 2844.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:42<00:22,  2.28it/s, est. speed input: 155.20 toks/s, output: 2935.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:44<00:36,  1.36it/s, est. speed input: 151.36 toks/s, output: 2939.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:44<00:28,  1.72it/s, est. speed input: 152.06 toks/s, output: 3029.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:45<00:29,  1.60it/s, est. speed input: 150.15 toks/s, output: 3142.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:46<00:30,  1.50it/s, est. speed input: 148.92 toks/s, output: 3192.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:47<00:22,  1.93it/s, est. speed input: 149.51 toks/s, output: 3355.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:47<00:21,  1.96it/s, est. speed input: 149.24 toks/s, output: 3425.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:48<00:24,  1.69it/s, est. speed input: 147.87 toks/s, output: 3469.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:48<00:21,  1.93it/s, est. speed input: 148.22 toks/s, output: 3551.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:49<00:16,  2.35it/s, est. speed input: 148.79 toks/s, output: 3715.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:30:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5236.1 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:50<00:23,  1.65it/s, est. speed input: 146.51 toks/s, output: 3731.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:50<00:17,  2.06it/s, est. speed input: 147.31 toks/s, output: 3826.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:50<00:14,  2.53it/s, est. speed input: 148.08 toks/s, output: 3920.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:52<00:31,  1.10it/s, est. speed input: 142.85 toks/s, output: 3858.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:53<00:27,  1.24it/s, est. speed input: 142.49 toks/s, output: 3924.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:53<00:20,  1.61it/s, est. speed input: 143.24 toks/s, output: 4018.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:54<00:21,  1.46it/s, est. speed input: 142.13 toks/s, output: 4061.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:54<00:16,  1.89it/s, est. speed input: 142.86 toks/s, output: 4156.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:54<00:13,  2.29it/s, est. speed input: 143.45 toks/s, output: 4246.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:56<00:20,  1.41it/s, est. speed input: 141.03 toks/s, output: 4250.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:56<00:18,  1.53it/s, est. speed input: 140.84 toks/s, output: 4318.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:57<00:18,  1.48it/s, est. speed input: 140.13 toks/s, output: 4372.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:58<00:17,  1.51it/s, est. speed input: 139.69 toks/s, output: 4433.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:58<00:13,  1.87it/s, est. speed input: 140.18 toks/s, output: 4522.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:59<00:13,  1.77it/s, est. speed input: 139.72 toks/s, output: 4582.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:59<00:10,  2.21it/s, est. speed input: 140.30 toks/s, output: 4676.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:59<00:12,  1.81it/s, est. speed input: 139.50 toks/s, output: 4724.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:30:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4011.8 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:00<00:10,  2.04it/s, est. speed input: 139.70 toks/s, output: 4806.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:00<00:10,  1.94it/s, est. speed input: 139.37 toks/s, output: 4870.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:01<00:08,  2.33it/s, est. speed input: 139.88 toks/s, output: 4962.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:01<00:06,  2.75it/s, est. speed input: 140.42 toks/s, output: 5055.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:03<00:15,  1.10it/s, est. speed input: 136.62 toks/s, output: 4993.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:03<00:10,  1.48it/s, est. speed input: 137.27 toks/s, output: 5094.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:05<00:13,  1.08it/s, est. speed input: 135.03 toks/s, output: 5089.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:06<00:12,  1.12it/s, est. speed input: 134.23 toks/s, output: 5137.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:06<00:11,  1.15it/s, est. speed input: 133.51 toks/s, output: 5189.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:08<00:15,  1.26s/it, est. speed input: 130.19 toks/s, output: 5142.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:10<00:13,  1.25s/it, est. speed input: 128.76 toks/s, output: 5168.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  1.25s/it, est. speed input: 138.34 toks/s, output: 6451.98 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  2.28it/s, est. speed input: 138.34 toks/s, output: 6451.98 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:30:33 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:30:33 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:30:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:30:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:30:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:30:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:30:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:30:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:30:34 [abstract.py:306] It took 0.764641 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:30:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:30:34 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0102
+  • Training Accuracy: 0.3375
+  • Mean Generation Length: 2830.8813
+  • Training FLOPS: 47.31 TFLOPS (11.83 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 336.74s
+  • policy_training: 217.39s (64.6%)
+  • generation: 75.07s (22.3%)
+  • teacher_logprob_inference: 39.00s (11.6%)
+  • policy_training/sharding_data: 4.29s (1.3%)
+  • prepare_for_generation: 3.57s (1.1%)
+  • policy_training/submit_training_futures: 1.89s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.54s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.66s (0.2%)
+  • get_topk_logits/shard_data: 0.05s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 22/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 106: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:34:53 [abstract.py:324] It took 0.055066 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08099375e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55381088e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80920410e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08099375e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55381088e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80920410e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08099375e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55381088e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80920410e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08099375e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.55381088e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80920410e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08097266e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54185486e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81752014e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08023872e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.09678650e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #106] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43932037e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08097266e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54185486e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81752014e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65067902e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37882261e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #106] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.54562531e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08097266e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54185486e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81752014e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37245789e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.20426011e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #106] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23538666e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08097266e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.54185486e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81752014e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28448982e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45744886e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #106] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33812439e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 106: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:34:56 [abstract.py:324] It took 0.029240 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 07:34:56 [loggers.py:236] Engine 000: Avg prompt throughput: 35.2 tokens/s, Avg generation throughput: 89.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12148.38it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:25,  6.31it/s, est. speed input: 378.38 toks/s, output: 75.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:02, 60.65it/s, est. speed input: 3152.32 toks/s, output: 819.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:02, 59.72it/s, est. speed input: 3299.33 toks/s, output: 1103.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 41.66it/s, est. speed input: 2686.55 toks/s, output: 1213.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:03, 38.05it/s, est. speed input: 2534.41 toks/s, output: 1414.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:00<00:03, 38.27it/s, est. speed input: 2514.67 toks/s, output: 1667.86 toks/s]
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 50.48it/s, est. speed input: 2796.44 toks/s, output: 2359.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 55.33it/s, est. speed input: 2934.25 toks/s, output: 2875.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 51.19it/s, est. speed input: 2902.27 toks/s, output: 3214.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 51.44it/s, est. speed input: 2926.74 toks/s, output: 3549.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:02, 39.82it/s, est. speed input: 2726.98 toks/s, output: 3635.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:03<00:07, 11.06it/s, est. speed input: 1564.49 toks/s, output: 2384.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:05<00:15,  4.98it/s, est. speed input: 935.09 toks/s, output: 1776.81 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:08<00:24,  3.04it/s, est. speed input: 654.53 toks/s, output: 1514.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:08<00:23,  3.03it/s, est. speed input: 619.23 toks/s, output: 1632.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:09<00:25,  2.79it/s, est. speed input: 569.74 toks/s, output: 1699.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:09<00:22,  3.04it/s, est. speed input: 569.67 toks/s, output: 1801.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:09<00:20,  3.28it/s, est. speed input: 566.97 toks/s, output: 1894.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10109.9 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:10<00:22,  2.96it/s, est. speed input: 545.21 toks/s, output: 1923.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:11<00:25,  2.54it/s, est. speed input: 520.02 toks/s, output: 1936.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:12<00:38,  1.69it/s, est. speed input: 469.76 toks/s, output: 1852.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:13<00:47,  1.35it/s, est. speed input: 431.99 toks/s, output: 1806.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:15<01:08,  1.09s/it, est. speed input: 377.49 toks/s, output: 1684.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:16<00:56,  1.09it/s, est. speed input: 371.31 toks/s, output: 1762.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:56,  1.09it/s, est. speed input: 354.37 toks/s, output: 1786.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:43,  1.38it/s, est. speed input: 353.39 toks/s, output: 1885.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:17<00:41,  1.42it/s, est. speed input: 344.10 toks/s, output: 1939.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:18<00:30,  1.89it/s, est. speed input: 345.61 toks/s, output: 2050.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:18<00:34,  1.65it/s, est. speed input: 334.37 toks/s, output: 2086.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:19<00:38,  1.45it/s, est. speed input: 322.14 toks/s, output: 2113.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:19<00:22,  2.44it/s, est. speed input: 325.99 toks/s, output: 2341.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7759.6 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:20<00:12,  4.19it/s, est. speed input: 331.94 toks/s, output: 2683.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:20<00:16,  3.11it/s, est. speed input: 324.06 toks/s, output: 2719.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:21<00:24,  1.97it/s, est. speed input: 309.15 toks/s, output: 2695.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:22<00:22,  2.09it/s, est. speed input: 306.54 toks/s, output: 2771.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:22<00:17,  2.62it/s, est. speed input: 305.39 toks/s, output: 2956.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:23<00:17,  2.56it/s, est. speed input: 302.52 toks/s, output: 3026.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:24<00:22,  2.00it/s, est. speed input: 294.43 toks/s, output: 3043.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:24<00:21,  1.98it/s, est. speed input: 290.66 toks/s, output: 3102.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:25<00:15,  2.61it/s, est. speed input: 290.42 toks/s, output: 3294.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:25<00:15,  2.59it/s, est. speed input: 288.28 toks/s, output: 3367.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:25<00:14,  2.71it/s, est. speed input: 287.07 toks/s, output: 3449.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:26<00:16,  2.23it/s, est. speed input: 279.63 toks/s, output: 3549.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:27<00:17,  2.06it/s, est. speed input: 275.66 toks/s, output: 3596.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:27<00:13,  2.51it/s, est. speed input: 276.50 toks/s, output: 3703.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:27<00:11,  3.04it/s, est. speed input: 277.37 toks/s, output: 3811.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5590.9 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:34,  1.04s/it, est. speed input: 252.58 toks/s, output: 3570.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:32<00:35,  1.12s/it, est. speed input: 244.25 toks/s, output: 3552.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:27,  1.11it/s, est. speed input: 243.43 toks/s, output: 3639.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:32<00:20,  1.47it/s, est. speed input: 244.23 toks/s, output: 3750.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:32<00:16,  1.77it/s, est. speed input: 243.86 toks/s, output: 3844.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:33<00:12,  2.21it/s, est. speed input: 244.32 toks/s, output: 3950.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:33<00:10,  2.51it/s, est. speed input: 244.21 toks/s, output: 4045.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:33<00:09,  2.67it/s, est. speed input: 243.72 toks/s, output: 4134.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:33<00:08,  2.92it/s, est. speed input: 243.55 toks/s, output: 4230.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:34<00:08,  2.76it/s, est. speed input: 242.31 toks/s, output: 4307.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:34<00:05,  3.92it/s, est. speed input: 243.98 toks/s, output: 4530.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:34<00:05,  3.86it/s, est. speed input: 243.74 toks/s, output: 4624.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:34<00:04,  4.36it/s, est. speed input: 244.53 toks/s, output: 4733.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:35<00:04,  4.58it/s, est. speed input: 244.92 toks/s, output: 4837.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:35<00:06,  2.69it/s, est. speed input: 241.41 toks/s, output: 4864.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:05,  3.04it/s, est. speed input: 241.63 toks/s, output: 4964.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:36<00:04,  3.37it/s, est. speed input: 241.86 toks/s, output: 5065.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:36<00:03,  3.92it/s, est. speed input: 242.45 toks/s, output: 5174.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:37<00:06,  2.25it/s, est. speed input: 238.25 toks/s, output: 5183.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:38<00:08,  1.53it/s, est. speed input: 232.72 toks/s, output: 5163.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:38<00:06,  1.87it/s, est. speed input: 232.69 toks/s, output: 5263.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:39<00:06,  1.78it/s, est. speed input: 230.43 toks/s, output: 5314.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3474.9 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:40<00:04,  2.08it/s, est. speed input: 229.00 toks/s, output: 5483.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:43<00:09,  1.20s/it, est. speed input: 212.63 toks/s, output: 5200.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:44<00:07,  1.01s/it, est. speed input: 211.73 toks/s, output: 5289.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:46<00:08,  1.39s/it, est. speed input: 202.03 toks/s, output: 5162.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:47<00:06,  1.21s/it, est. speed input: 200.05 toks/s, output: 5228.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1575.0 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:08,  2.11s/it, est. speed input: 184.29 toks/s, output: 4942.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:52<00:04,  1.64s/it, est. speed input: 183.75 toks/s, output: 5053.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.64s/it, est. speed input: 186.99 toks/s, output: 5519.29 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.07it/s, est. speed input: 186.99 toks/s, output: 5519.29 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:35:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:35:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:35:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:35:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:35:51 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:35:51 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:35:51 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:35:52 [abstract.py:306] It took 0.611990 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:35:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0124
+  • Training Accuracy: 0.4188
+  • Mean Generation Length: 1797.5125
+  • Training FLOPS: 42.42 TFLOPS (10.60 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 239.46s
+  • policy_training: 151.41s (63.2%)
+  • generation: 55.45s (23.2%)
+  • teacher_logprob_inference: 27.43s (11.5%)
+  • policy_training/sharding_data: 4.33s (1.8%)
+  • prepare_for_generation: 3.65s (1.5%)
+  • policy_training/submit_training_futures: 1.72s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.7%)
+  • training_prep: 0.88s (0.4%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 107: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:38:53 [abstract.py:324] It took 0.050525 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08133867e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57665977e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80986252e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08133867e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57665977e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80986252e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08133867e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57665977e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80986252e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08133867e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.57665977e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80986252e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08127891e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56519165e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83106232e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08740444e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.18358898e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #107] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44278030e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08127891e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56519165e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83106232e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66344376e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38179770e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #107] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.56781006e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08127891e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56519165e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83106232e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36600800e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.14369941e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #107] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23897934e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08127891e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.56519165e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83106232e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.28802109e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45492496e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #107] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33720535e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 107: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:38:56 [abstract.py:324] It took 0.030187 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 07:38:56 [loggers.py:236] Engine 000: Avg prompt throughput: 51.2 tokens/s, Avg generation throughput: 12.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13259.21it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.71it/s, est. speed input: 336.77 toks/s, output: 79.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:13, 11.95it/s, est. speed input: 662.90 toks/s, output: 226.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:11, 12.87it/s, est. speed input: 726.32 toks/s, output: 353.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 23.23it/s, est. speed input: 1123.16 toks/s, output: 795.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:06, 22.54it/s, est. speed input: 1154.34 toks/s, output: 917.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 28.77it/s, est. speed input: 1358.56 toks/s, output: 1309.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 40.35it/s, est. speed input: 1665.21 toks/s, output: 1900.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 50.69it/s, est. speed input: 1972.39 toks/s, output: 2586.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 54.41it/s, est. speed input: 2128.54 toks/s, output: 3030.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 43.51it/s, est. speed input: 2073.73 toks/s, output: 3188.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:02, 37.10it/s, est. speed input: 2017.35 toks/s, output: 3306.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 46.03it/s, est. speed input: 2171.99 toks/s, output: 3910.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:02<00:04, 20.05it/s, est. speed input: 1693.33 toks/s, output: 3323.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:07<00:26,  3.22it/s, est. speed input: 594.85 toks/s, output: 1431.38 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 07:39:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10686.9 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:11<00:37,  2.18it/s, est. speed input: 432.25 toks/s, output: 1319.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:12<00:36,  2.14it/s, est. speed input: 405.42 toks/s, output: 1432.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<00:37,  2.05it/s, est. speed input: 379.04 toks/s, output: 1525.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:14<00:42,  1.78it/s, est. speed input: 344.13 toks/s, output: 1565.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:15<00:38,  1.92it/s, est. speed input: 342.46 toks/s, output: 1652.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:16<00:45,  1.60it/s, est. speed input: 321.08 toks/s, output: 1643.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:16<00:40,  1.80it/s, est. speed input: 320.30 toks/s, output: 1732.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:16<00:33,  2.13it/s, est. speed input: 321.45 toks/s, output: 1831.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:17<00:28,  2.44it/s, est. speed input: 321.02 toks/s, output: 1920.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:17<00:26,  2.63it/s, est. speed input: 319.29 toks/s, output: 2000.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:17<00:28,  2.42it/s, est. speed input: 313.63 toks/s, output: 2055.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:18<00:25,  2.62it/s, est. speed input: 311.86 toks/s, output: 2133.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:19<00:49,  1.34it/s, est. speed input: 288.32 toks/s, output: 2063.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:39:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8188.5 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:20<00:40,  1.62it/s, est. speed input: 287.02 toks/s, output: 2144.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:21<00:50,  1.28it/s, est. speed input: 273.74 toks/s, output: 2136.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<01:01,  1.02it/s, est. speed input: 258.64 toks/s, output: 2110.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:22<00:44,  1.38it/s, est. speed input: 260.10 toks/s, output: 2213.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:23<00:36,  1.67it/s, est. speed input: 259.36 toks/s, output: 2296.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:23<00:34,  1.74it/s, est. speed input: 256.19 toks/s, output: 2358.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:23<00:29,  2.02it/s, est. speed input: 255.53 toks/s, output: 2440.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:24<00:25,  2.29it/s, est. speed input: 254.84 toks/s, output: 2522.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:25<00:36,  1.56it/s, est. speed input: 246.00 toks/s, output: 2523.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:25<00:34,  1.62it/s, est. speed input: 243.02 toks/s, output: 2581.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:26<00:22,  2.33it/s, est. speed input: 242.11 toks/s, output: 2831.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:27<00:22,  2.27it/s, est. speed input: 239.97 toks/s, output: 2893.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:28<00:28,  1.80it/s, est. speed input: 234.17 toks/s, output: 2911.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:28<00:24,  2.07it/s, est. speed input: 234.03 toks/s, output: 2996.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:29<00:31,  1.56it/s, est. speed input: 227.42 toks/s, output: 2998.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:39:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6654.0 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:30<00:35,  1.37it/s, est. speed input: 222.28 toks/s, output: 3017.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:30<00:28,  1.67it/s, est. speed input: 222.32 toks/s, output: 3105.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:32<00:36,  1.27it/s, est. speed input: 215.55 toks/s, output: 3098.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:33<00:44,  1.00it/s, est. speed input: 207.63 toks/s, output: 3073.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:33<00:34,  1.28it/s, est. speed input: 207.73 toks/s, output: 3163.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:34<00:26,  1.61it/s, est. speed input: 208.04 toks/s, output: 3256.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:34<00:23,  1.80it/s, est. speed input: 207.45 toks/s, output: 3333.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:36<00:35,  1.17it/s, est. speed input: 200.13 toks/s, output: 3304.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:26,  1.50it/s, est. speed input: 200.56 toks/s, output: 3399.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:33,  1.16it/s, est. speed input: 195.16 toks/s, output: 3395.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:39<00:39,  1.05s/it, est. speed input: 189.22 toks/s, output: 3381.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:39<00:23,  1.53it/s, est. speed input: 190.55 toks/s, output: 3579.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:39<00:21,  1.65it/s, est. speed input: 189.86 toks/s, output: 3654.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:39:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5279.1 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:40<00:22,  1.53it/s, est. speed input: 187.71 toks/s, output: 3699.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:40<00:17,  1.87it/s, est. speed input: 188.22 toks/s, output: 3796.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:41<00:10,  2.95it/s, est. speed input: 190.40 toks/s, output: 4012.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:41<00:10,  2.84it/s, est. speed input: 190.08 toks/s, output: 4090.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:42<00:13,  2.09it/s, est. speed input: 187.69 toks/s, output: 4126.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:43<00:16,  1.65it/s, est. speed input: 183.77 toks/s, output: 4209.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:44<00:16,  1.55it/s, est. speed input: 181.95 toks/s, output: 4255.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:45<00:15,  1.58it/s, est. speed input: 180.96 toks/s, output: 4318.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:10,  2.24it/s, est. speed input: 182.12 toks/s, output: 4520.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:21,  1.05it/s, est. speed input: 173.30 toks/s, output: 4392.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:20,  1.03it/s, est. speed input: 170.93 toks/s, output: 4422.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:17,  1.12it/s, est. speed input: 169.85 toks/s, output: 4484.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:39:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3827.4 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:50<00:16,  1.16it/s, est. speed input: 168.40 toks/s, output: 4538.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:51<00:11,  1.42it/s, est. speed input: 167.47 toks/s, output: 4695.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:52<00:11,  1.40it/s, est. speed input: 166.26 toks/s, output: 4752.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:52<00:08,  1.68it/s, est. speed input: 166.64 toks/s, output: 4855.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:09,  1.40it/s, est. speed input: 164.57 toks/s, output: 4886.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:55<00:11,  1.13it/s, est. speed input: 161.62 toks/s, output: 4893.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:10,  1.17it/s, est. speed input: 160.44 toks/s, output: 4953.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:56<00:09,  1.14it/s, est. speed input: 158.89 toks/s, output: 5000.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:59<00:13,  1.34s/it, est. speed input: 153.25 toks/s, output: 4921.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:39:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2379.2 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:01<00:14,  1.66s/it, est. speed input: 148.18 toks/s, output: 4859.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:01<00:09,  1.20s/it, est. speed input: 148.90 toks/s, output: 4983.87 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.20s/it, est. speed input: 156.81 toks/s, output: 6042.83 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.59it/s, est. speed input: 156.81 toks/s, output: 6042.83 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:40:01 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:40:01 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:40:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:40:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:40:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:40:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:40:01 [abstract.py:306] It took 0.529175 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:40:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:40:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:40:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:40:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0139
+  • Training Accuracy: 0.2875
+  • Mean Generation Length: 2337.2812
+  • Training FLOPS: 45.75 TFLOPS (11.44 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 287.72s
+  • policy_training: 184.61s (64.2%)
+  • generation: 65.92s (22.9%)
+  • teacher_logprob_inference: 32.10s (11.2%)
+  • policy_training/sharding_data: 4.25s (1.5%)
+  • prepare_for_generation: 3.60s (1.3%)
+  • policy_training/submit_training_futures: 2.80s (1.0%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.5%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 108: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:43:40 [abstract.py:324] It took 0.048778 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08140898e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59478531e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80005493e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08140898e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59478531e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80005493e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08140898e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59478531e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80005493e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08140898e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59478531e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80005493e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08139414e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60178566e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80505371e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09336319e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.13668251e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #108] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44245346e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08139414e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60178566e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80505371e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68383636e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38596153e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #108] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.57227783e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08139414e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60178566e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80505371e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35134277e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.17332482e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #108] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23629379e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08139414e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60178566e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80505371e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26700172e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.46279869e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #108] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33703537e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 108: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:43:43 [abstract.py:324] It took 0.030619 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13538.20it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:43:43 [loggers.py:236] Engine 000: Avg prompt throughput: 43.1 tokens/s, Avg generation throughput: 16.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:51,  3.07it/s, est. speed input: 190.37 toks/s, output: 39.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:04, 31.19it/s, est. speed input: 1536.02 toks/s, output: 449.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 42.31it/s, est. speed input: 2085.31 toks/s, output: 779.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:04, 32.87it/s, est. speed input: 1883.09 toks/s, output: 904.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:03, 40.87it/s, est. speed input: 2156.26 toks/s, output: 1371.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 44.96it/s, est. speed input: 2301.75 toks/s, output: 1723.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 48.04it/s, est. speed input: 2417.44 toks/s, output: 2100.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 53.11it/s, est. speed input: 2558.16 toks/s, output: 2531.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 57.41it/s, est. speed input: 2685.04 toks/s, output: 2952.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 43.74it/s, est. speed input: 2544.08 toks/s, output: 3117.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 43.20it/s, est. speed input: 2547.00 toks/s, output: 3418.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:02<00:05, 15.24it/s, est. speed input: 1758.78 toks/s, output: 2620.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:07<00:23,  3.31it/s, est. speed input: 697.80 toks/s, output: 1316.90 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:07<00:20,  3.57it/s, est. speed input: 671.06 toks/s, output: 1570.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:08<00:18,  3.84it/s, est. speed input: 660.92 toks/s, output: 1748.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:09<00:25,  2.71it/s, est. speed input: 558.49 toks/s, output: 1665.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:10<00:21,  3.24it/s, est. speed input: 562.21 toks/s, output: 1877.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:43:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10134.6 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:12<00:32,  2.05it/s, est. speed input: 471.49 toks/s, output: 1773.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:12<00:29,  2.19it/s, est. speed input: 466.18 toks/s, output: 1855.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:12<00:28,  2.22it/s, est. speed input: 455.64 toks/s, output: 1913.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:14<00:46,  1.36it/s, est. speed input: 400.39 toks/s, output: 1784.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:15<00:40,  1.52it/s, est. speed input: 394.79 toks/s, output: 1862.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:15<00:38,  1.60it/s, est. speed input: 385.61 toks/s, output: 1921.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:15<00:31,  1.92it/s, est. speed input: 384.37 toks/s, output: 2016.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:17<00:40,  1.46it/s, est. speed input: 362.04 toks/s, output: 2002.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:17<00:36,  1.59it/s, est. speed input: 355.78 toks/s, output: 2069.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:18<00:41,  1.37it/s, est. speed input: 340.04 toks/s, output: 2079.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:18<00:34,  1.65it/s, est. speed input: 338.10 toks/s, output: 2168.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:19<00:33,  1.62it/s, est. speed input: 330.23 toks/s, output: 2218.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:19<00:30,  1.80it/s, est. speed input: 326.47 toks/s, output: 2294.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7680.1 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:21<00:52,  1.02it/s, est. speed input: 299.52 toks/s, output: 2206.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:22<00:40,  1.28it/s, est. speed input: 298.24 toks/s, output: 2298.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:22<00:29,  1.72it/s, est. speed input: 299.54 toks/s, output: 2409.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:23<00:32,  1.53it/s, est. speed input: 291.56 toks/s, output: 2446.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:23<00:28,  1.71it/s, est. speed input: 288.89 toks/s, output: 2523.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:25,  1.90it/s, est. speed input: 286.92 toks/s, output: 2604.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:25<00:32,  1.44it/s, est. speed input: 277.01 toks/s, output: 2614.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:25<00:26,  1.72it/s, est. speed input: 275.94 toks/s, output: 2703.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:25,  1.76it/s, est. speed input: 272.72 toks/s, output: 2770.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:26<00:20,  2.11it/s, est. speed input: 272.55 toks/s, output: 2866.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:26<00:15,  2.72it/s, est. speed input: 273.65 toks/s, output: 2975.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:14,  2.86it/s, est. speed input: 272.71 toks/s, output: 3064.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:17,  2.39it/s, est. speed input: 269.11 toks/s, output: 3121.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:27<00:14,  2.83it/s, est. speed input: 269.41 toks/s, output: 3222.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:27<00:12,  3.11it/s, est. speed input: 269.12 toks/s, output: 3316.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:27<00:11,  3.44it/s, est. speed input: 269.22 toks/s, output: 3413.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:29<00:23,  1.58it/s, est. speed input: 258.14 toks/s, output: 3370.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:29<00:17,  2.10it/s, est. speed input: 259.31 toks/s, output: 3482.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:10,  3.08it/s, est. speed input: 259.94 toks/s, output: 3777.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5915.9 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:09,  3.27it/s, est. speed input: 259.89 toks/s, output: 3872.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:31<00:15,  2.01it/s, est. speed input: 252.53 toks/s, output: 3858.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:31<00:08,  3.24it/s, est. speed input: 255.35 toks/s, output: 4186.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:32<00:10,  2.59it/s, est. speed input: 251.60 toks/s, output: 4220.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:33<00:10,  2.40it/s, est. speed input: 248.15 toks/s, output: 4349.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:33<00:07,  2.95it/s, est. speed input: 249.08 toks/s, output: 4553.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:34<00:09,  2.32it/s, est. speed input: 244.92 toks/s, output: 4574.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:34<00:08,  2.60it/s, est. speed input: 245.12 toks/s, output: 4673.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:36<00:14,  1.40it/s, est. speed input: 234.92 toks/s, output: 4576.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:36<00:11,  1.61it/s, est. speed input: 234.37 toks/s, output: 4664.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:37<00:09,  1.97it/s, est. speed input: 234.81 toks/s, output: 4770.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:37<00:07,  2.31it/s, est. speed input: 234.99 toks/s, output: 4870.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:38<00:08,  1.92it/s, est. speed input: 232.03 toks/s, output: 4907.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:38<00:08,  1.80it/s, est. speed input: 229.82 toks/s, output: 4959.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:39<00:10,  1.33it/s, est. speed input: 224.24 toks/s, output: 4939.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3570.7 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:11,  1.15it/s, est. speed input: 219.44 toks/s, output: 4936.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:41<00:09,  1.24it/s, est. speed input: 217.43 toks/s, output: 4995.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:11,  1.05s/it, est. speed input: 210.73 toks/s, output: 4947.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:08,  1.12it/s, est. speed input: 209.65 toks/s, output: 5030.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:07,  1.24it/s, est. speed input: 208.14 toks/s, output: 5100.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:09,  1.21s/it, est. speed input: 199.84 toks/s, output: 5009.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:06,  1.13it/s, est. speed input: 200.67 toks/s, output: 5141.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:48<00:04,  1.05it/s, est. speed input: 194.76 toks/s, output: 5209.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1900.9 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:05,  1.41s/it, est. speed input: 185.34 toks/s, output: 5078.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:52<00:03,  1.23s/it, est. speed input: 183.93 toks/s, output: 5162.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:52<00:01,  1.07it/s, est. speed input: 184.54 toks/s, output: 5301.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:52<00:00,  1.26it/s, est. speed input: 184.30 toks/s, output: 5413.90 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.26it/s, est. speed input: 185.42 toks/s, output: 5568.50 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.02it/s, est. speed input: 185.42 toks/s, output: 5568.50 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:44:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:44:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:44:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:44:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:44:39 [abstract.py:306] It took 0.558702 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:44:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:44:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:44:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:44:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0128
+  • Training Accuracy: 0.3937
+  • Mean Generation Length: 1844.0000
+  • Training FLOPS: 43.01 TFLOPS (10.75 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 242.83s
+  • policy_training: 153.54s (63.2%)
+  • generation: 56.27s (23.2%)
+  • teacher_logprob_inference: 28.06s (11.6%)
+  • policy_training/sharding_data: 4.25s (1.7%)
+  • prepare_for_generation: 3.53s (1.5%)
+  • policy_training/submit_training_futures: 1.96s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.6%)
+  • training_prep: 0.86s (0.4%)
+  • teacher_logprob_inference_prep: 0.40s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 109: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:47:44 [abstract.py:324] It took 0.054527 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08148262e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60240173e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78575897e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08148262e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60240173e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78575897e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08148262e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60240173e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78575897e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08148262e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60240173e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78575897e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08154258e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59900818e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80561523e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08587093e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.96536541e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #109] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44427109e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08154258e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59900818e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80561523e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69344330e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38899460e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #109] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.54317474e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08154258e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59900818e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80561523e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36816406e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.20045590e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #109] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23846359e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08154258e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59900818e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80561523e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26435318e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45048428e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #109] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33246414e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 109: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:47:47 [abstract.py:324] It took 0.028332 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14070.12it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 07:47:47 [loggers.py:236] Engine 000: Avg prompt throughput: 49.8 tokens/s, Avg generation throughput: 15.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:31,  5.03it/s, est. speed input: 312.12 toks/s, output: 80.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 25.20it/s, est. speed input: 1330.43 toks/s, output: 454.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:03, 43.82it/s, est. speed input: 2128.53 toks/s, output: 981.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 42.23it/s, est. speed input: 2206.87 toks/s, output: 1223.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 36.41it/s, est. speed input: 2090.22 toks/s, output: 1344.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:03, 37.50it/s, est. speed input: 2137.29 toks/s, output: 1586.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 48.13it/s, est. speed input: 2418.19 toks/s, output: 2226.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:01, 59.12it/s, est. speed input: 2704.75 toks/s, output: 2919.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 65.02it/s, est. speed input: 2907.87 toks/s, output: 3542.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 57.79it/s, est. speed input: 2880.48 toks/s, output: 3816.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 47.34it/s, est. speed input: 2758.59 toks/s, output: 3985.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:01<00:02, 34.70it/s, est. speed input: 2507.37 toks/s, output: 3921.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:04<00:11,  6.46it/s, est. speed input: 1072.66 toks/s, output: 1969.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:08<00:26,  2.76it/s, est. speed input: 590.52 toks/s, output: 1336.72 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:09<00:23,  2.90it/s, est. speed input: 560.05 toks/s, output: 1584.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:47:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9988.8 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:12<00:34,  1.95it/s, est. speed input: 443.43 toks/s, output: 1448.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:13<00:31,  2.07it/s, est. speed input: 429.18 toks/s, output: 1609.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:15<00:39,  1.58it/s, est. speed input: 371.08 toks/s, output: 1587.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:17<00:46,  1.34it/s, est. speed input: 341.61 toks/s, output: 1568.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:18<00:46,  1.31it/s, est. speed input: 328.78 toks/s, output: 1615.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7802.6 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:58,  1.03it/s, est. speed input: 300.61 toks/s, output: 1584.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:20<00:50,  1.17it/s, est. speed input: 297.57 toks/s, output: 1674.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:40,  1.43it/s, est. speed input: 297.78 toks/s, output: 1781.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:32,  1.77it/s, est. speed input: 298.61 toks/s, output: 1890.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:26,  2.12it/s, est. speed input: 298.60 toks/s, output: 1993.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:21<00:21,  2.61it/s, est. speed input: 299.37 toks/s, output: 2101.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:21<00:17,  3.02it/s, est. speed input: 299.38 toks/s, output: 2203.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:22<00:22,  2.34it/s, est. speed input: 292.95 toks/s, output: 2257.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:23<00:23,  2.16it/s, est. speed input: 285.16 toks/s, output: 2395.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:23<00:20,  2.38it/s, est. speed input: 281.79 toks/s, output: 2564.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:23<00:18,  2.61it/s, est. speed input: 281.41 toks/s, output: 2660.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:24<00:19,  2.47it/s, est. speed input: 278.39 toks/s, output: 2730.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:24<00:16,  2.87it/s, est. speed input: 278.89 toks/s, output: 2832.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:19,  2.28it/s, est. speed input: 273.72 toks/s, output: 2877.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:26<00:25,  1.73it/s, est. speed input: 266.22 toks/s, output: 2896.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:19,  2.12it/s, est. speed input: 264.11 toks/s, output: 3067.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:15,  2.61it/s, est. speed input: 265.27 toks/s, output: 3178.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:27<00:14,  2.81it/s, est. speed input: 264.81 toks/s, output: 3269.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:28<00:17,  2.23it/s, est. speed input: 260.28 toks/s, output: 3310.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:29<00:32,  1.17it/s, est. speed input: 245.52 toks/s, output: 3220.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5939.8 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:27,  1.34it/s, est. speed input: 243.74 toks/s, output: 3295.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:30<00:17,  1.98it/s, est. speed input: 244.36 toks/s, output: 3497.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:31,  1.09it/s, est. speed input: 229.90 toks/s, output: 3389.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:28,  1.16it/s, est. speed input: 226.87 toks/s, output: 3443.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:21,  1.51it/s, est. speed input: 227.80 toks/s, output: 3555.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:33<00:11,  2.52it/s, est. speed input: 230.58 toks/s, output: 3794.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:12,  2.32it/s, est. speed input: 228.76 toks/s, output: 3861.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:10,  2.54it/s, est. speed input: 227.79 toks/s, output: 4033.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:36<00:17,  1.47it/s, est. speed input: 219.03 toks/s, output: 3975.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:39<00:26,  1.08s/it, est. speed input: 207.59 toks/s, output: 3868.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:39<00:21,  1.14it/s, est. speed input: 207.57 toks/s, output: 3966.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:17,  1.31it/s, est. speed input: 206.74 toks/s, output: 4050.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4198.1 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:42<00:25,  1.15s/it, est. speed input: 197.49 toks/s, output: 3970.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:42<00:19,  1.09it/s, est. speed input: 197.37 toks/s, output: 4068.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:42<00:10,  1.76it/s, est. speed input: 198.81 toks/s, output: 4298.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:42<00:08,  2.10it/s, est. speed input: 199.40 toks/s, output: 4408.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:43<00:08,  2.07it/s, est. speed input: 198.48 toks/s, output: 4487.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:07,  2.09it/s, est. speed input: 197.75 toks/s, output: 4568.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:44<00:06,  2.23it/s, est. speed input: 197.47 toks/s, output: 4660.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:44<00:06,  2.02it/s, est. speed input: 196.10 toks/s, output: 4728.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:46<00:09,  1.37it/s, est. speed input: 191.77 toks/s, output: 4726.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:46<00:03,  2.82it/s, est. speed input: 194.98 toks/s, output: 5107.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:46<00:02,  3.06it/s, est. speed input: 195.31 toks/s, output: 5216.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:47<00:03,  2.41it/s, est. speed input: 193.61 toks/s, output: 5272.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:48<00:03,  1.77it/s, est. speed input: 190.78 toks/s, output: 5296.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2467.5 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:50<00:06,  1.09s/it, est. speed input: 182.32 toks/s, output: 5170.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:51<00:05,  1.11s/it, est. speed input: 179.36 toks/s, output: 5195.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:53<00:04,  1.19s/it, est. speed input: 175.80 toks/s, output: 5204.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:55<00:04,  1.38s/it, est. speed input: 170.96 toks/s, output: 5176.68 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  1.38s/it, est. speed input: 174.21 toks/s, output: 5621.54 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  2.90it/s, est. speed input: 174.21 toks/s, output: 5621.54 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:45 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:45 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:48:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:48:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:48:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:48:45 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:48:45 [abstract.py:306] It took 0.526948 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:48:45 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:48:45 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:48:45 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:48:45 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0132
+  • Training Accuracy: 0.3187
+  • Mean Generation Length: 1940.9313
+  • Training FLOPS: 41.40 TFLOPS (10.35 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 262.64s
+  • policy_training: 168.52s (64.2%)
+  • generation: 58.63s (22.3%)
+  • teacher_logprob_inference: 30.39s (11.6%)
+  • policy_training/sharding_data: 4.31s (1.6%)
+  • prepare_for_generation: 3.56s (1.4%)
+  • policy_training/submit_training_futures: 2.00s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.6%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.46s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 110: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:06 [abstract.py:324] It took 0.048013 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08102480e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60452957e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76986008e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08102480e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60452957e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76986008e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08102480e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60452957e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76986008e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08102480e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60452957e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76986008e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08103477e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60556107e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76069336e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.11173115e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.91051435e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #110] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44488922e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08103477e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60556107e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76069336e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71502609e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39948406e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #110] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.57490463e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08103477e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60556107e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76069336e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37364578e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.16140890e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #110] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24320030e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08103477e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60556107e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76069336e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25311966e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45782623e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #110] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33156784e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 110: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:09 [abstract.py:324] It took 0.028812 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13883.25it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:10 [loggers.py:236] Engine 000: Avg prompt throughput: 45.5 tokens/s, Avg generation throughput: 30.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:14,  2.12it/s, est. speed input: 129.63 toks/s, output: 29.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:09, 16.55it/s, est. speed input: 808.16 toks/s, output: 242.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:07, 20.92it/s, est. speed input: 1004.65 toks/s, output: 396.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:07, 19.78it/s, est. speed input: 1029.57 toks/s, output: 514.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:04, 27.63it/s, est. speed input: 1266.99 toks/s, output: 838.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 38.83it/s, est. speed input: 1564.82 toks/s, output: 1302.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 36.32it/s, est. speed input: 1602.18 toks/s, output: 1492.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:03, 30.12it/s, est. speed input: 1556.14 toks/s, output: 1621.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:03, 29.45it/s, est. speed input: 1568.14 toks/s, output: 1798.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:03, 30.57it/s, est. speed input: 1600.85 toks/s, output: 2014.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:02<00:05, 18.46it/s, est. speed input: 1396.33 toks/s, output: 1923.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:03<00:12,  8.68it/s, est. speed input: 1035.26 toks/s, output: 1578.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:03<00:11,  8.70it/s, est. speed input: 1001.82 toks/s, output: 1664.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:04<00:14,  6.83it/s, est. speed input: 897.85 toks/s, output: 1628.15 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:05<00:24,  4.06it/s, est. speed input: 716.06 toks/s, output: 1449.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:06<00:37,  2.60it/s, est. speed input: 586.87 toks/s, output: 1268.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:08<01:07,  1.43it/s, est. speed input: 440.00 toks/s, output: 1039.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:09<01:07,  1.42it/s, est. speed input: 411.66 toks/s, output: 1060.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11182.7 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:11<01:27,  1.09it/s, est. speed input: 354.11 toks/s, output: 1001.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:11<01:10,  1.34it/s, est. speed input: 353.94 toks/s, output: 1088.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:11<00:47,  1.94it/s, est. speed input: 354.05 toks/s, output: 1260.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:12<00:49,  1.84it/s, est. speed input: 340.68 toks/s, output: 1298.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:13<01:07,  1.33it/s, est. speed input: 310.19 toks/s, output: 1268.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:13<00:52,  1.70it/s, est. speed input: 311.88 toks/s, output: 1360.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:14<00:43,  2.02it/s, est. speed input: 310.89 toks/s, output: 1440.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:14<00:42,  2.04it/s, est. speed input: 304.95 toks/s, output: 1496.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:14<00:35,  2.44it/s, est. speed input: 304.72 toks/s, output: 1577.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:14<00:29,  2.86it/s, est. speed input: 304.62 toks/s, output: 1659.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:15<00:39,  2.12it/s, est. speed input: 293.45 toks/s, output: 1681.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:15<00:32,  2.59it/s, est. speed input: 293.96 toks/s, output: 1765.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:17<01:02,  1.31it/s, est. speed input: 269.68 toks/s, output: 1701.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:19<01:33,  1.15s/it, est. speed input: 244.46 toks/s, output: 1624.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8905.2 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:21<01:41,  1.26s/it, est. speed input: 229.64 toks/s, output: 1610.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:21<01:26,  1.10s/it, est. speed input: 225.03 toks/s, output: 1660.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:22<01:24,  1.08s/it, est. speed input: 217.59 toks/s, output: 1688.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:23<01:18,  1.02s/it, est. speed input: 211.93 toks/s, output: 1727.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:24<00:55,  1.35it/s, est. speed input: 209.79 toks/s, output: 1870.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:25<00:49,  1.50it/s, est. speed input: 208.38 toks/s, output: 1939.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:25<00:43,  1.66it/s, est. speed input: 207.35 toks/s, output: 2009.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<00:41,  1.76it/s, est. speed input: 205.84 toks/s, output: 2074.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:26<00:33,  2.15it/s, est. speed input: 206.55 toks/s, output: 2161.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:27<00:52,  1.34it/s, est. speed input: 198.02 toks/s, output: 2150.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:28<00:46,  1.50it/s, est. speed input: 196.85 toks/s, output: 2215.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:28<00:39,  1.73it/s, est. speed input: 196.59 toks/s, output: 2288.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:29<00:29,  2.23it/s, est. speed input: 196.72 toks/s, output: 2443.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:29<00:23,  2.74it/s, est. speed input: 197.97 toks/s, output: 2535.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:29<00:19,  3.34it/s, est. speed input: 199.24 toks/s, output: 2627.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7377.1 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:30<00:33,  1.87it/s, est. speed input: 193.51 toks/s, output: 2628.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:30<00:27,  2.27it/s, est. speed input: 194.18 toks/s, output: 2713.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:30<00:21,  2.85it/s, est. speed input: 195.31 toks/s, output: 2804.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:30<00:18,  3.19it/s, est. speed input: 195.82 toks/s, output: 2886.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:31<00:17,  3.43it/s, est. speed input: 196.32 toks/s, output: 2967.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:31<00:16,  3.51it/s, est. speed input: 196.51 toks/s, output: 3044.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:32<00:23,  2.46it/s, est. speed input: 194.20 toks/s, output: 3081.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:33<00:37,  1.51it/s, est. speed input: 188.59 toks/s, output: 3067.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:37<01:04,  1.19s/it, est. speed input: 173.47 toks/s, output: 2964.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:37<00:49,  1.08it/s, est. speed input: 174.46 toks/s, output: 3057.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:38<00:46,  1.11it/s, est. speed input: 172.27 toks/s, output: 3095.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:38<00:27,  1.80it/s, est. speed input: 174.58 toks/s, output: 3287.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:23,  2.08it/s, est. speed input: 175.09 toks/s, output: 3371.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:38<00:18,  2.57it/s, est. speed input: 176.16 toks/s, output: 3465.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:52:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6131.0 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:40<00:37,  1.24it/s, est. speed input: 169.11 toks/s, output: 3401.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:40<00:31,  1.46it/s, est. speed input: 169.02 toks/s, output: 3475.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:41<00:24,  1.81it/s, est. speed input: 169.62 toks/s, output: 3562.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:41<00:19,  2.26it/s, est. speed input: 170.00 toks/s, output: 3719.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:42<00:16,  2.49it/s, est. speed input: 170.31 toks/s, output: 3800.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:09,  4.06it/s, est. speed input: 173.57 toks/s, output: 4089.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:11,  3.24it/s, est. speed input: 172.73 toks/s, output: 4140.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:15,  2.39it/s, est. speed input: 170.87 toks/s, output: 4169.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:44<00:15,  2.37it/s, est. speed input: 170.50 toks/s, output: 4235.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:17,  1.99it/s, est. speed input: 169.01 toks/s, output: 4271.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:19,  1.71it/s, est. speed input: 167.31 toks/s, output: 4303.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:12,  2.48it/s, est. speed input: 168.70 toks/s, output: 4485.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:12,  2.33it/s, est. speed input: 167.88 toks/s, output: 4608.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:47<00:14,  2.06it/s, est. speed input: 166.72 toks/s, output: 4649.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:48<00:13,  2.00it/s, est. speed input: 166.12 toks/s, output: 4706.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:48<00:11,  2.40it/s, est. speed input: 166.79 toks/s, output: 4798.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:09,  2.85it/s, est. speed input: 167.48 toks/s, output: 4891.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:48<00:08,  2.94it/s, est. speed input: 167.61 toks/s, output: 4968.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:48<00:06,  3.53it/s, est. speed input: 168.41 toks/s, output: 5064.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:49<00:07,  3.17it/s, est. speed input: 168.28 toks/s, output: 5133.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:49<00:06,  3.15it/s, est. speed input: 168.37 toks/s, output: 5210.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:50<00:07,  2.64it/s, est. speed input: 167.83 toks/s, output: 5266.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:05,  3.34it/s, est. speed input: 168.61 toks/s, output: 5365.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:53:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4515.8 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:50<00:04,  3.98it/s, est. speed input: 169.36 toks/s, output: 5461.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:50<00:04,  3.84it/s, est. speed input: 169.62 toks/s, output: 5542.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:51<00:07,  2.33it/s, est. speed input: 168.09 toks/s, output: 5565.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:05,  2.74it/s, est. speed input: 168.55 toks/s, output: 5654.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:52<00:06,  2.41it/s, est. speed input: 167.96 toks/s, output: 5710.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:52<00:04,  2.98it/s, est. speed input: 168.71 toks/s, output: 5883.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:54<00:06,  1.72it/s, est. speed input: 165.70 toks/s, output: 5853.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:55<00:09,  1.18it/s, est. speed input: 162.02 toks/s, output: 5802.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:56<00:09,  1.11it/s, est. speed input: 160.09 toks/s, output: 5812.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:57<00:08,  1.12it/s, est. speed input: 158.70 toks/s, output: 5842.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:58<00:06,  1.19it/s, est. speed input: 157.88 toks/s, output: 5893.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:03,  1.60it/s, est. speed input: 158.04 toks/s, output: 6061.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:59<00:02,  1.68it/s, est. speed input: 157.71 toks/s, output: 6132.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:53:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2128.0 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:03,  1.13it/s, est. speed input: 154.28 toks/s, output: 6085.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:02<00:02,  1.07it/s, est. speed input: 152.59 toks/s, output: 6108.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:03<00:00,  1.39it/s, est. speed input: 152.39 toks/s, output: 6279.46 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.39it/s, est. speed input: 153.32 toks/s, output: 6408.98 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.53it/s, est. speed input: 153.32 toks/s, output: 6408.98 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:53:16 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:53:16 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:53:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:53:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:53:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:53:16 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:53:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 07:53:17 [abstract.py:306] It took 0.537212 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:53:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:53:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:53:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0118
+  • Training Accuracy: 0.3875
+  • Mean Generation Length: 2533.1875
+  • Training FLOPS: 45.39 TFLOPS (11.35 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 307.24s
+  • policy_training: 199.33s (64.9%)
+  • generation: 67.47s (22.0%)
+  • teacher_logprob_inference: 35.43s (11.5%)
+  • policy_training/sharding_data: 4.33s (1.4%)
+  • prepare_for_generation: 3.58s (1.2%)
+  • policy_training/submit_training_futures: 1.70s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 23/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 111: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:14 [abstract.py:324] It took 0.051960 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08096992e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60309601e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75571442e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08096992e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60309601e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75571442e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08096992e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60309601e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75571442e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08096992e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60309601e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75571442e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08098555e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61175575e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76138458e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.11062050e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.94093084e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #111] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44792480e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08098555e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61175575e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76138458e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.71842957e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39447880e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #111] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.56980286e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08098555e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61175575e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76138458e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37698212e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.13829017e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #111] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.27264519e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08098555e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61175575e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76138458e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.24794922e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45631571e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #111] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33027527e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 111: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:17 [abstract.py:324] It took 0.030785 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13340.40it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:17 [loggers.py:236] Engine 000: Avg prompt throughput: 39.2 tokens/s, Avg generation throughput: 11.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:21,  7.40it/s, est. speed input: 458.81 toks/s, output: 73.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:06, 23.54it/s, est. speed input: 1269.74 toks/s, output: 299.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:04, 34.06it/s, est. speed input: 1783.00 toks/s, output: 575.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:02, 47.58it/s, est. speed input: 2408.22 toks/s, output: 1089.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:02, 45.61it/s, est. speed input: 2441.95 toks/s, output: 1318.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:03, 42.40it/s, est. speed input: 2403.23 toks/s, output: 1521.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 49.59it/s, est. speed input: 2607.77 toks/s, output: 1975.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:00<00:02, 45.44it/s, est. speed input: 2564.88 toks/s, output: 2235.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 41.34it/s, est. speed input: 2497.58 toks/s, output: 2410.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 42.14it/s, est. speed input: 2510.29 toks/s, output: 2678.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 41.22it/s, est. speed input: 2496.04 toks/s, output: 2919.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 53.56it/s, est. speed input: 2696.13 toks/s, output: 3609.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:03<00:11,  7.94it/s, est. speed input: 1143.54 toks/s, output: 1738.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:06<00:17,  4.86it/s, est. speed input: 777.74 toks/s, output: 1627.87 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10514.0 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:11<00:41,  1.93it/s, est. speed input: 410.63 toks/s, output: 1085.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<00:39,  1.95it/s, est. speed input: 378.53 toks/s, output: 1296.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:16<00:53,  1.40it/s, est. speed input: 307.72 toks/s, output: 1246.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:17<00:46,  1.55it/s, est. speed input: 302.48 toks/s, output: 1417.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:17<00:42,  1.69it/s, est. speed input: 302.30 toks/s, output: 1513.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:18<00:47,  1.51it/s, est. speed input: 287.95 toks/s, output: 1538.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:19<00:42,  1.65it/s, est. speed input: 286.29 toks/s, output: 1624.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8284.7 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:20<00:57,  1.20it/s, est. speed input: 265.01 toks/s, output: 1598.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:48,  1.40it/s, est. speed input: 264.44 toks/s, output: 1688.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:22<00:51,  1.31it/s, est. speed input: 256.07 toks/s, output: 1728.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:22<00:46,  1.43it/s, est. speed input: 252.90 toks/s, output: 1800.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:22<00:37,  1.73it/s, est. speed input: 252.90 toks/s, output: 1892.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:23<00:31,  2.05it/s, est. speed input: 252.76 toks/s, output: 1982.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:24<00:30,  2.03it/s, est. speed input: 247.18 toks/s, output: 2118.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:25<00:37,  1.64it/s, est. speed input: 239.85 toks/s, output: 2147.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:25<00:24,  2.43it/s, est. speed input: 242.45 toks/s, output: 2348.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:25<00:22,  2.55it/s, est. speed input: 241.72 toks/s, output: 2430.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:25<00:18,  3.02it/s, est. speed input: 242.68 toks/s, output: 2528.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:26<00:30,  1.83it/s, est. speed input: 234.44 toks/s, output: 2531.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:29<00:54,  1.01it/s, est. speed input: 218.76 toks/s, output: 2451.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:29<00:43,  1.24it/s, est. speed input: 218.33 toks/s, output: 2536.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:35,  1.47it/s, est. speed input: 217.71 toks/s, output: 2618.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6848.3 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:28,  1.79it/s, est. speed input: 215.83 toks/s, output: 2770.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:30<00:23,  2.11it/s, est. speed input: 216.27 toks/s, output: 2863.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:29,  1.67it/s, est. speed input: 211.59 toks/s, output: 2889.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:32<00:18,  2.49it/s, est. speed input: 213.61 toks/s, output: 3091.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:31,  1.46it/s, est. speed input: 205.17 toks/s, output: 3056.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:34<00:20,  2.14it/s, est. speed input: 207.00 toks/s, output: 3256.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:34<00:13,  3.06it/s, est. speed input: 209.52 toks/s, output: 3467.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:34<00:12,  3.30it/s, est. speed input: 210.02 toks/s, output: 3560.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:35<00:19,  2.08it/s, est. speed input: 205.26 toks/s, output: 3564.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:36<00:19,  1.99it/s, est. speed input: 203.70 toks/s, output: 3621.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:36<00:10,  3.27it/s, est. speed input: 206.64 toks/s, output: 3925.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:37<00:14,  2.46it/s, est. speed input: 203.65 toks/s, output: 3952.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:37<00:12,  2.71it/s, est. speed input: 204.02 toks/s, output: 4042.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:38<00:10,  2.96it/s, est. speed input: 204.09 toks/s, output: 4208.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:38<00:10,  2.98it/s, est. speed input: 203.68 toks/s, output: 4364.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:39<00:09,  2.98it/s, est. speed input: 203.43 toks/s, output: 4441.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:39<00:11,  2.52it/s, est. speed input: 201.80 toks/s, output: 4489.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:39<00:09,  2.86it/s, est. speed input: 202.32 toks/s, output: 4583.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:57:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5160.2 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:40<00:09,  2.70it/s, est. speed input: 201.63 toks/s, output: 4651.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:40<00:11,  2.19it/s, est. speed input: 199.68 toks/s, output: 4689.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:41<00:11,  2.16it/s, est. speed input: 198.78 toks/s, output: 4752.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:41<00:06,  3.18it/s, est. speed input: 200.48 toks/s, output: 4958.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:41<00:06,  3.48it/s, est. speed input: 200.98 toks/s, output: 5052.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:42<00:05,  3.44it/s, est. speed input: 200.94 toks/s, output: 5134.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:42<00:07,  2.41it/s, est. speed input: 198.81 toks/s, output: 5163.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:43<00:06,  2.62it/s, est. speed input: 198.89 toks/s, output: 5248.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:04,  3.91it/s, est. speed input: 200.79 toks/s, output: 5464.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:43<00:02,  5.56it/s, est. speed input: 202.97 toks/s, output: 5689.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:43<00:02,  5.52it/s, est. speed input: 203.55 toks/s, output: 5785.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:03,  3.27it/s, est. speed input: 201.69 toks/s, output: 5814.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:03,  2.57it/s, est. speed input: 199.48 toks/s, output: 5995.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:48<00:06,  1.16it/s, est. speed input: 189.09 toks/s, output: 5774.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:49<00:05,  1.25it/s, est. speed input: 188.16 toks/s, output: 5839.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 07:58:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2523.8 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:50<00:03,  1.35it/s, est. speed input: 185.72 toks/s, output: 5950.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:53<00:05,  1.26s/it, est. speed input: 175.93 toks/s, output: 5739.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:56<00:04,  1.61s/it, est. speed input: 168.32 toks/s, output: 5601.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:56<00:02,  1.26s/it, est. speed input: 168.82 toks/s, output: 5726.76 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.26s/it, est. speed input: 170.93 toks/s, output: 6015.81 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.82it/s, est. speed input: 170.93 toks/s, output: 6015.81 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:58:17 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 07:58:17 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 07:58:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:58:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:58:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:58:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 07:58:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 07:58:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 07:58:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 07:58:17 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 07:58:17 [abstract.py:306] It took 0.588238 seconds to fall asleep.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0110
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 2131.0000
+  • Training FLOPS: 42.89 TFLOPS (10.72 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 274.40s
+  • policy_training: 177.09s (64.5%)
+  • generation: 60.36s (22.0%)
+  • teacher_logprob_inference: 31.80s (11.6%)
+  • policy_training/sharding_data: 4.36s (1.6%)
+  • prepare_for_generation: 3.68s (1.3%)
+  • policy_training/submit_training_futures: 1.77s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 112: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:01:48 [abstract.py:324] It took 0.048635 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08104102e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08104102e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08104102e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08104102e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08115918e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60843201e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76005783e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08115918e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60843201e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76005783e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08115918e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60843201e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76005783e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08115918e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.60843201e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76005783e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.23022041e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45517073e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61369858e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75538330e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61369858e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75538330e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61369858e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75538330e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61369858e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75538330e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09342155e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.86757469e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #112] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44823669e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70514069e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39712963e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #112] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.57637711e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36488342e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.10434604e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #112] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.27648010e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #112] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33338104e+02
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 112: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:01:51 [abstract.py:324] It took 0.030259 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14001.43it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:01:51 [loggers.py:236] Engine 000: Avg prompt throughput: 43.1 tokens/s, Avg generation throughput: 33.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.69it/s, est. speed input: 217.84 toks/s, output: 36.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 15.52it/s, est. speed input: 807.10 toks/s, output: 249.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:10, 13.86it/s, est. speed input: 775.96 toks/s, output: 303.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:10, 14.90it/s, est. speed input: 821.02 toks/s, output: 397.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:07, 18.71it/s, est. speed input: 939.26 toks/s, output: 572.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:06, 21.56it/s, est. speed input: 1032.17 toks/s, output: 743.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:05, 25.71it/s, est. speed input: 1152.10 toks/s, output: 979.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:04, 31.58it/s, est. speed input: 1307.05 toks/s, output: 1285.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:04, 29.39it/s, est. speed input: 1336.34 toks/s, output: 1456.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 33.75it/s, est. speed input: 1445.73 toks/s, output: 1755.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 33.89it/s, est. speed input: 1505.62 toks/s, output: 2005.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 42.83it/s, est. speed input: 1669.72 toks/s, output: 2494.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 48.71it/s, est. speed input: 1816.53 toks/s, output: 3016.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:02<00:03, 32.93it/s, est. speed input: 1713.01 toks/s, output: 3074.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:02<00:05, 18.34it/s, est. speed input: 1441.35 toks/s, output: 2820.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:03<00:09,  9.89it/s, est. speed input: 1113.15 toks/s, output: 2425.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:08<00:32,  2.74it/s, est. speed input: 542.35 toks/s, output: 1393.81 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:09<00:33,  2.57it/s, est. speed input: 493.23 toks/s, output: 1450.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:02:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10943.7 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:10<00:39,  2.12it/s, est. speed input: 427.97 toks/s, output: 1439.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:11<00:40,  2.06it/s, est. speed input: 411.01 toks/s, output: 1474.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:11<00:40,  2.03it/s, est. speed input: 397.76 toks/s, output: 1518.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:13<00:49,  1.63it/s, est. speed input: 365.22 toks/s, output: 1487.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:13<00:37,  2.10it/s, est. speed input: 364.20 toks/s, output: 1663.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:13<00:32,  2.42it/s, est. speed input: 364.71 toks/s, output: 1756.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<00:29,  2.63it/s, est. speed input: 362.28 toks/s, output: 1832.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:17<01:20,  1.07s/it, est. speed input: 295.82 toks/s, output: 1588.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:17<00:50,  1.47it/s, est. speed input: 299.38 toks/s, output: 1787.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:17<00:42,  1.74it/s, est. speed input: 299.24 toks/s, output: 1875.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:19<00:57,  1.25it/s, est. speed input: 278.80 toks/s, output: 1836.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:19<00:52,  1.36it/s, est. speed input: 274.19 toks/s, output: 1895.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:02:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8404.7 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:20<00:51,  1.36it/s, est. speed input: 267.29 toks/s, output: 1936.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<01:01,  1.12it/s, est. speed input: 253.89 toks/s, output: 1928.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:23<01:18,  1.15s/it, est. speed input: 236.83 toks/s, output: 1888.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:24<01:07,  1.00s/it, est. speed input: 233.07 toks/s, output: 1946.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:24<01:03,  1.05it/s, est. speed input: 227.73 toks/s, output: 1989.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:25<00:35,  1.82it/s, est. speed input: 231.49 toks/s, output: 2196.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:26<00:51,  1.23it/s, est. speed input: 219.76 toks/s, output: 2171.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:27<00:53,  1.17it/s, est. speed input: 214.18 toks/s, output: 2202.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:27<00:39,  1.53it/s, est. speed input: 215.63 toks/s, output: 2302.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:27<00:30,  1.97it/s, est. speed input: 216.69 toks/s, output: 2399.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:02:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6962.8 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:30<01:03,  1.08s/it, est. speed input: 200.75 toks/s, output: 2310.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:31<01:06,  1.14s/it, est. speed input: 194.37 toks/s, output: 2324.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:34<01:29,  1.57s/it, est. speed input: 181.41 toks/s, output: 2258.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:34<01:11,  1.27s/it, est. speed input: 180.19 toks/s, output: 2330.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:36<01:16,  1.40s/it, est. speed input: 173.48 toks/s, output: 2330.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:37<01:06,  1.23s/it, est. speed input: 171.23 toks/s, output: 2387.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:38<01:04,  1.22s/it, est. speed input: 167.51 toks/s, output: 2421.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:39<00:58,  1.12s/it, est. speed input: 165.23 toks/s, output: 2475.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:02:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6173.3 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:40<00:43,  1.15it/s, est. speed input: 163.60 toks/s, output: 2621.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:41<00:38,  1.26it/s, est. speed input: 162.78 toks/s, output: 2693.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:42<00:34,  1.38it/s, est. speed input: 160.84 toks/s, output: 2828.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:43<00:30,  1.47it/s, est. speed input: 159.10 toks/s, output: 2966.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:44<00:28,  1.56it/s, est. speed input: 158.68 toks/s, output: 3042.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:44<00:26,  1.60it/s, est. speed input: 158.03 toks/s, output: 3113.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:45<00:25,  1.68it/s, est. speed input: 157.59 toks/s, output: 3188.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:45<00:21,  1.88it/s, est. speed input: 157.71 toks/s, output: 3273.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:46<00:21,  1.83it/s, est. speed input: 156.95 toks/s, output: 3341.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:47<00:19,  1.97it/s, est. speed input: 156.48 toks/s, output: 3494.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:47<00:19,  1.94it/s, est. speed input: 155.91 toks/s, output: 3564.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:48<00:17,  2.05it/s, est. speed input: 155.89 toks/s, output: 3644.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:48<00:14,  2.49it/s, est. speed input: 156.64 toks/s, output: 3742.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:48<00:14,  2.32it/s, est. speed input: 156.27 toks/s, output: 3814.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:49<00:14,  2.28it/s, est. speed input: 156.01 toks/s, output: 3888.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:02:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4885.3 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:50<00:19,  1.65it/s, est. speed input: 154.07 toks/s, output: 3920.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:50<00:16,  1.91it/s, est. speed input: 154.23 toks/s, output: 4004.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:50<00:12,  2.36it/s, est. speed input: 154.88 toks/s, output: 4100.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:51<00:11,  2.42it/s, est. speed input: 154.86 toks/s, output: 4180.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:51<00:12,  2.33it/s, est. speed input: 154.66 toks/s, output: 4253.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:52<00:13,  2.06it/s, est. speed input: 154.02 toks/s, output: 4315.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:52<00:10,  2.44it/s, est. speed input: 154.41 toks/s, output: 4484.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:53<00:09,  2.52it/s, est. speed input: 154.50 toks/s, output: 4566.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:53<00:07,  2.91it/s, est. speed input: 155.08 toks/s, output: 4662.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:55<00:16,  1.34it/s, est. speed input: 151.01 toks/s, output: 4620.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:55<00:15,  1.34it/s, est. speed input: 150.08 toks/s, output: 4672.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:57<00:12,  1.49it/s, est. speed input: 149.15 toks/s, output: 4805.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:57<00:12,  1.47it/s, est. speed input: 148.34 toks/s, output: 4861.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:58<00:13,  1.27it/s, est. speed input: 146.63 toks/s, output: 4887.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:02:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3343.3 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:01<00:15,  1.07s/it, est. speed input: 141.69 toks/s, output: 4886.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:02<00:13,  1.07it/s, est. speed input: 141.58 toks/s, output: 4967.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:05<00:18,  1.39s/it, est. speed input: 136.49 toks/s, output: 4876.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:06<00:16,  1.34s/it, est. speed input: 134.88 toks/s, output: 4909.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:06<00:11,  1.03s/it, est. speed input: 135.38 toks/s, output: 5017.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:06<00:08,  1.18it/s, est. speed input: 135.50 toks/s, output: 5112.18 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  1.18it/s, est. speed input: 144.50 toks/s, output: 6337.04 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  2.39it/s, est. speed input: 144.50 toks/s, output: 6337.04 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:03:02 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:03:02 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:03:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:03:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:03:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:03:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:03:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:03:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:03:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:03:03 [abstract.py:306] It took 0.546283 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:03:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0104
+  • Training Accuracy: 0.4188
+  • Mean Generation Length: 2648.8813
+  • Training FLOPS: 45.58 TFLOPS (11.40 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 324.46s
+  • policy_training: 211.09s (65.1%)
+  • generation: 71.34s (22.0%)
+  • teacher_logprob_inference: 36.94s (11.4%)
+  • policy_training/sharding_data: 4.31s (1.3%)
+  • prepare_for_generation: 3.58s (1.1%)
+  • policy_training/submit_training_futures: 1.82s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.5%)
+  • training_prep: 0.96s (0.3%)
+  • teacher_logprob_inference_prep: 0.39s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 113: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:13 [abstract.py:324] It took 0.050995 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08132812e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62510796e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76215210e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08132812e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62510796e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76215210e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08132812e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62510796e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76215210e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08132812e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62510796e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76215210e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08148125e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63890038e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77017441e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.10418568e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.97076797e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #113] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44533112e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08148125e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63890038e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77017441e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70697174e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39126434e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #113] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.59586487e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08148125e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63890038e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77017441e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36555939e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.02159119e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #113] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.32322121e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08148125e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63890038e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77017441e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.22223816e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.45812674e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #113] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33175476e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 113: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:16 [abstract.py:324] It took 0.030471 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:17 [loggers.py:236] Engine 000: Avg prompt throughput: 36.7 tokens/s, Avg generation throughput: 62.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12941.64it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:30,  5.22it/s, est. speed input: 313.00 toks/s, output: 52.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:04, 36.21it/s, est. speed input: 1830.46 toks/s, output: 473.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 46.78it/s, est. speed input: 2379.22 toks/s, output: 779.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:02, 51.14it/s, est. speed input: 2639.25 toks/s, output: 1088.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 43.60it/s, est. speed input: 2510.64 toks/s, output: 1243.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:03, 40.69it/s, est. speed input: 2444.55 toks/s, output: 1447.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:00<00:02, 44.41it/s, est. speed input: 2539.92 toks/s, output: 1798.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:01, 57.98it/s, est. speed input: 2866.89 toks/s, output: 2494.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 56.81it/s, est. speed input: 2914.42 toks/s, output: 2851.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 52.02it/s, est. speed input: 2880.18 toks/s, output: 3112.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 57.00it/s, est. speed input: 2990.57 toks/s, output: 3619.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:01<00:02, 41.40it/s, est. speed input: 2760.57 toks/s, output: 3648.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:02<00:04, 18.39it/s, est. speed input: 2040.08 toks/s, output: 2984.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:07<00:23,  3.19it/s, est. speed input: 699.65 toks/s, output: 1331.15 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:09<00:27,  2.59it/s, est. speed input: 565.18 toks/s, output: 1358.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10053.7 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:11<00:33,  2.11it/s, est. speed input: 480.61 toks/s, output: 1359.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:13<00:36,  1.85it/s, est. speed input: 427.55 toks/s, output: 1408.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:13<00:29,  2.21it/s, est. speed input: 430.24 toks/s, output: 1624.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:13<00:27,  2.33it/s, est. speed input: 426.49 toks/s, output: 1712.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:15<00:40,  1.57it/s, est. speed input: 381.17 toks/s, output: 1635.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:15<00:34,  1.81it/s, est. speed input: 380.65 toks/s, output: 1737.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:15<00:32,  1.91it/s, est. speed input: 374.53 toks/s, output: 1812.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:16<00:36,  1.69it/s, est. speed input: 359.61 toks/s, output: 1843.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:36,  1.62it/s, est. speed input: 348.85 toks/s, output: 1891.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:19<00:56,  1.04it/s, est. speed input: 316.70 toks/s, output: 1821.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7808.4 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<01:05,  1.13s/it, est. speed input: 295.75 toks/s, output: 1804.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<01:19,  1.39s/it, est. speed input: 271.95 toks/s, output: 1764.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:23<01:02,  1.11s/it, est. speed input: 269.81 toks/s, output: 1854.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:24<00:54,  1.02it/s, est. speed input: 264.84 toks/s, output: 1923.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:47,  1.14it/s, est. speed input: 260.69 toks/s, output: 1995.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:38,  1.37it/s, est. speed input: 259.32 toks/s, output: 2087.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:33,  1.54it/s, est. speed input: 256.93 toks/s, output: 2169.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:30,  1.67it/s, est. speed input: 254.62 toks/s, output: 2250.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:29,  1.68it/s, est. speed input: 251.26 toks/s, output: 2322.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:27<00:25,  1.95it/s, est. speed input: 250.59 toks/s, output: 2415.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:28<00:33,  1.42it/s, est. speed input: 242.54 toks/s, output: 2437.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:28<00:26,  1.78it/s, est. speed input: 242.77 toks/s, output: 2538.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:28<00:17,  2.61it/s, est. speed input: 244.03 toks/s, output: 2746.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:29<00:18,  2.37it/s, est. speed input: 241.65 toks/s, output: 2817.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:29<00:16,  2.62it/s, est. speed input: 241.50 toks/s, output: 2912.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:29<00:14,  2.92it/s, est. speed input: 241.55 toks/s, output: 3009.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6366.4 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:19,  2.03it/s, est. speed input: 234.81 toks/s, output: 3114.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:31<00:14,  2.67it/s, est. speed input: 235.81 toks/s, output: 3319.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:12,  3.04it/s, est. speed input: 236.44 toks/s, output: 3423.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:16,  2.14it/s, est. speed input: 231.66 toks/s, output: 3448.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:14,  2.28it/s, est. speed input: 229.87 toks/s, output: 3608.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:12,  2.65it/s, est. speed input: 230.48 toks/s, output: 3713.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:10,  3.12it/s, est. speed input: 231.35 toks/s, output: 3819.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:34<00:09,  3.17it/s, est. speed input: 231.09 toks/s, output: 3908.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:34<00:11,  2.59it/s, est. speed input: 228.89 toks/s, output: 3965.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:35<00:13,  2.18it/s, est. speed input: 226.34 toks/s, output: 4016.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:35<00:12,  2.21it/s, est. speed input: 225.36 toks/s, output: 4091.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:16,  1.59it/s, est. speed input: 220.56 toks/s, output: 4097.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:37<00:11,  2.13it/s, est. speed input: 220.50 toks/s, output: 4284.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:15,  1.50it/s, est. speed input: 212.64 toks/s, output: 4317.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:39<00:12,  1.80it/s, est. speed input: 213.27 toks/s, output: 4425.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:07:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4340.1 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:40<00:13,  1.53it/s, est. speed input: 209.65 toks/s, output: 4445.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:41<00:13,  1.43it/s, est. speed input: 206.88 toks/s, output: 4483.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:41<00:12,  1.55it/s, est. speed input: 205.84 toks/s, output: 4558.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:41<00:08,  2.01it/s, est. speed input: 206.73 toks/s, output: 4674.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:11,  1.50it/s, est. speed input: 202.81 toks/s, output: 4683.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:08,  1.90it/s, est. speed input: 203.39 toks/s, output: 4793.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:43<00:03,  3.35it/s, est. speed input: 205.91 toks/s, output: 5142.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:04,  2.78it/s, est. speed input: 204.56 toks/s, output: 5204.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:04,  2.49it/s, est. speed input: 203.43 toks/s, output: 5272.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:03,  2.88it/s, est. speed input: 203.93 toks/s, output: 5382.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:03,  2.75it/s, est. speed input: 203.42 toks/s, output: 5466.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:47<00:08,  1.03s/it, est. speed input: 192.76 toks/s, output: 5282.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:08:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2367.5 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:51<00:12,  1.78s/it, est. speed input: 180.04 toks/s, output: 5045.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:51<00:08,  1.35s/it, est. speed input: 180.23 toks/s, output: 5161.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:55<00:09,  1.93s/it, est. speed input: 170.37 toks/s, output: 4996.01 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  1.93s/it, est. speed input: 175.82 toks/s, output: 5736.31 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  2.89it/s, est. speed input: 175.82 toks/s, output: 5736.31 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:08:15 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:08:15 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:08:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:08:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:08:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:08:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:08:15 [abstract.py:306] It took 0.519587 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:08:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:08:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:08:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:08:15 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0110
+  • Training Accuracy: 0.3063
+  • Mean Generation Length: 1983.6062
+  • Training FLOPS: 42.33 TFLOPS (10.58 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 262.29s
+  • policy_training: 168.48s (64.2%)
+  • generation: 58.74s (22.4%)
+  • teacher_logprob_inference: 29.89s (11.4%)
+  • policy_training/sharding_data: 4.34s (1.7%)
+  • prepare_for_generation: 3.68s (1.4%)
+  • policy_training/submit_training_futures: 1.77s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.6%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 114: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:11:36 [abstract.py:324] It took 0.052289 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08174355e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63245392e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76260223e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08174355e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63245392e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76260223e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08174355e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63245392e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76260223e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08174355e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63245392e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76260223e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08178105e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64104309e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76433334e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.10371017e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.90975380e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #114] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44487289e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08178105e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64104309e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76433334e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69051895e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38236160e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #114] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.63766327e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08178105e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64104309e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76433334e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35504456e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.01514792e+00
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #114] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.34740143e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08178105e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64104309e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76433334e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.21267014e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.46264563e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #114] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33389648e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 114: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:11:39 [abstract.py:324] It took 0.028953 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 08:11:39 [loggers.py:236] Engine 000: Avg prompt throughput: 46.2 tokens/s, Avg generation throughput: 37.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14030.41it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:36,  4.39it/s, est. speed input: 263.52 toks/s, output: 83.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:04, 34.11it/s, est. speed input: 1712.90 toks/s, output: 686.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:03, 36.68it/s, est. speed input: 1909.46 toks/s, output: 901.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 41.89it/s, est. speed input: 2170.55 toks/s, output: 1247.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 41.05it/s, est. speed input: 2211.48 toks/s, output: 1461.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:02, 52.48it/s, est. speed input: 2566.74 toks/s, output: 2121.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 50.85it/s, est. speed input: 2610.94 toks/s, output: 2398.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 48.38it/s, est. speed input: 2618.09 toks/s, output: 2660.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 53.40it/s, est. speed input: 2739.45 toks/s, output: 3106.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 60.19it/s, est. speed input: 2897.53 toks/s, output: 3655.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:01, 47.38it/s, est. speed input: 2767.65 toks/s, output: 3826.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:01<00:02, 40.05it/s, est. speed input: 2650.93 toks/s, output: 3977.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:01<00:01, 42.08it/s, est. speed input: 2673.14 toks/s, output: 4295.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:05<00:14,  5.14it/s, est. speed input: 1004.86 toks/s, output: 1948.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:11:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9893.9 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:10<00:30,  2.27it/s, est. speed input: 534.97 toks/s, output: 1345.87 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:12<00:30,  2.15it/s, est. speed input: 475.69 toks/s, output: 1496.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:13<00:31,  2.03it/s, est. speed input: 439.79 toks/s, output: 1590.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:14<00:30,  2.03it/s, est. speed input: 418.52 toks/s, output: 1721.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:15<00:32,  1.85it/s, est. speed input: 397.55 toks/s, output: 1741.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:16<00:36,  1.65it/s, est. speed input: 376.46 toks/s, output: 1755.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:17<00:41,  1.43it/s, est. speed input: 355.16 toks/s, output: 1763.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:54,  1.06it/s, est. speed input: 322.23 toks/s, output: 1706.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:46,  1.21it/s, est. speed input: 318.92 toks/s, output: 1796.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:11:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7666.9 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:22<01:07,  1.20s/it, est. speed input: 287.04 toks/s, output: 1723.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:22<00:59,  1.07s/it, est. speed input: 280.84 toks/s, output: 1793.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:23<00:49,  1.10it/s, est. speed input: 278.00 toks/s, output: 1880.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:24<00:51,  1.03it/s, est. speed input: 267.68 toks/s, output: 1915.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:24<00:40,  1.28it/s, est. speed input: 266.88 toks/s, output: 2014.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:33,  1.54it/s, est. speed input: 265.80 toks/s, output: 2110.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:30,  1.62it/s, est. speed input: 262.61 toks/s, output: 2188.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:25<00:25,  1.91it/s, est. speed input: 261.99 toks/s, output: 2285.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:22,  2.13it/s, est. speed input: 261.01 toks/s, output: 2378.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:28<00:50,  1.08s/it, est. speed input: 240.05 toks/s, output: 2291.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:28<00:36,  1.25it/s, est. speed input: 241.13 toks/s, output: 2403.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:12:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6326.7 tokens/s, Running: 46 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:31<00:59,  1.31s/it, est. speed input: 223.75 toks/s, output: 2333.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:32<00:48,  1.10s/it, est. speed input: 221.54 toks/s, output: 2412.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:32<00:27,  1.52it/s, est. speed input: 223.36 toks/s, output: 2634.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:33<00:27,  1.52it/s, est. speed input: 220.69 toks/s, output: 2704.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:33<00:24,  1.66it/s, est. speed input: 219.59 toks/s, output: 2791.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:34<00:23,  1.67it/s, est. speed input: 217.58 toks/s, output: 2865.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:34<00:22,  1.72it/s, est. speed input: 215.96 toks/s, output: 2943.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:35<00:30,  1.23it/s, est. speed input: 209.31 toks/s, output: 2951.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:36<00:29,  1.23it/s, est. speed input: 206.34 toks/s, output: 3009.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:36<00:21,  1.65it/s, est. speed input: 207.45 toks/s, output: 3124.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:37<00:15,  2.15it/s, est. speed input: 207.66 toks/s, output: 3319.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:37<00:10,  2.95it/s, est. speed input: 209.33 toks/s, output: 3539.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:38<00:13,  2.16it/s, est. speed input: 205.99 toks/s, output: 3580.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:38<00:11,  2.60it/s, est. speed input: 206.80 toks/s, output: 3691.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:12:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4867.6 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:22,  1.19it/s, est. speed input: 194.76 toks/s, output: 3666.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:42<00:21,  1.19it/s, est. speed input: 192.36 toks/s, output: 3720.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:43<00:18,  1.38it/s, est. speed input: 192.16 toks/s, output: 3815.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:25,  1.05s/it, est. speed input: 185.05 toks/s, output: 3772.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:20,  1.12it/s, est. speed input: 184.51 toks/s, output: 3861.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:46<00:20,  1.08it/s, est. speed input: 181.81 toks/s, output: 3903.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:17,  1.20it/s, est. speed input: 180.75 toks/s, output: 3981.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:47<00:16,  1.21it/s, est. speed input: 178.95 toks/s, output: 4041.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:48<00:12,  1.55it/s, est. speed input: 179.51 toks/s, output: 4152.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:48<00:09,  1.90it/s, est. speed input: 179.80 toks/s, output: 4259.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:48<00:08,  1.93it/s, est. speed input: 179.23 toks/s, output: 4344.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:49<00:07,  2.28it/s, est. speed input: 179.53 toks/s, output: 4451.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:05,  2.55it/s, est. speed input: 179.76 toks/s, output: 4554.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:12:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3416.0 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:08,  1.70it/s, est. speed input: 177.23 toks/s, output: 4589.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:51<00:09,  1.35it/s, est. speed input: 174.69 toks/s, output: 4623.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:08,  1.43it/s, est. speed input: 173.83 toks/s, output: 4700.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:11,  1.05s/it, est. speed input: 168.99 toks/s, output: 4672.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:54<00:07,  1.30it/s, est. speed input: 169.70 toks/s, output: 4795.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:54<00:05,  1.68it/s, est. speed input: 170.22 toks/s, output: 4912.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:54<00:04,  1.70it/s, est. speed input: 169.54 toks/s, output: 4995.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:03,  2.08it/s, est. speed input: 169.90 toks/s, output: 5109.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:56<00:04,  1.35it/s, est. speed input: 166.95 toks/s, output: 5124.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:57<00:03,  1.44it/s, est. speed input: 166.30 toks/s, output: 5209.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:57<00:02,  1.56it/s, est. speed input: 165.88 toks/s, output: 5301.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:58<00:02,  1.40it/s, est. speed input: 164.44 toks/s, output: 5361.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.40it/s, est. speed input: 167.55 toks/s, output: 5781.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.74it/s, est. speed input: 167.55 toks/s, output: 5781.72 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:12:40 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:12:40 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:12:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:12:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:12:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:12:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:12:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:12:41 [abstract.py:306] It took 0.522441 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:12:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:12:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:12:41 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0122
+  • Training Accuracy: 0.4188
+  • Mean Generation Length: 2111.8250
+  • Training FLOPS: 44.70 TFLOPS (11.18 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 268.65s
+  • policy_training: 171.51s (63.8%)
+  • generation: 62.07s (23.1%)
+  • teacher_logprob_inference: 30.09s (11.2%)
+  • policy_training/sharding_data: 4.35s (1.6%)
+  • prepare_for_generation: 3.55s (1.3%)
+  • policy_training/submit_training_futures: 1.93s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.38s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 115: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:04 [abstract.py:324] It took 0.050826 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08185234e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63942146e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76601181e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08185234e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63942146e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76601181e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08185234e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63942146e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76601181e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08185234e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63942146e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76601181e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08190645e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66320839e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76834717e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.11534634e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.85638523e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #115] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44553284e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08190645e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66320839e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76834717e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69638062e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39337149e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #115] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.66080399e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08190645e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66320839e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76834717e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37833557e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.98571634e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #115] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.35325928e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08190645e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66320839e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76834717e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19201736e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.47822495e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #115] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33281296e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 115: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:07 [abstract.py:324] It took 0.032463 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13883.25it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:08 [loggers.py:236] Engine 000: Avg prompt throughput: 44.4 tokens/s, Avg generation throughput: 72.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:55,  2.86it/s, est. speed input: 171.67 toks/s, output: 37.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:08, 19.08it/s, est. speed input: 958.53 toks/s, output: 293.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 27.90it/s, est. speed input: 1347.13 toks/s, output: 489.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:07, 18.20it/s, est. speed input: 1100.03 toks/s, output: 578.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:07, 18.90it/s, est. speed input: 1126.21 toks/s, output: 724.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:04, 32.09it/s, est. speed input: 1470.73 toks/s, output: 1347.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 34.33it/s, est. speed input: 1563.85 toks/s, output: 1624.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 30.70it/s, est. speed input: 1555.87 toks/s, output: 1815.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:03, 36.14it/s, est. speed input: 1672.04 toks/s, output: 2197.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:03, 32.88it/s, est. speed input: 1664.86 toks/s, output: 2401.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:02<00:03, 27.57it/s, est. speed input: 1608.07 toks/s, output: 2508.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:02<00:04, 20.58it/s, est. speed input: 1482.16 toks/s, output: 2503.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:02<00:07, 12.97it/s, est. speed input: 1272.37 toks/s, output: 2307.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:03<00:07, 13.51it/s, est. speed input: 1262.56 toks/s, output: 2415.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:09<01:06,  1.42it/s, est. speed input: 413.09 toks/s, output: 947.10 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11260.1 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:11<01:08,  1.36it/s, est. speed input: 361.55 toks/s, output: 1016.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:11<01:05,  1.41it/s, est. speed input: 350.39 toks/s, output: 1076.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:13<01:12,  1.25it/s, est. speed input: 321.25 toks/s, output: 1078.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:13<01:14,  1.21it/s, est. speed input: 304.27 toks/s, output: 1112.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:16<01:45,  1.18s/it, est. speed input: 261.50 toks/s, output: 1047.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9096.7 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:22<03:16,  2.23s/it, est. speed input: 197.11 toks/s, output: 881.51 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:23<02:47,  1.92s/it, est. speed input: 191.38 toks/s, output: 946.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:23<02:17,  1.60s/it, est. speed input: 188.48 toks/s, output: 1021.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:23<01:41,  1.19s/it, est. speed input: 190.16 toks/s, output: 1118.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:24<01:17,  1.08it/s, est. speed input: 190.67 toks/s, output: 1208.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:26<01:48,  1.30s/it, est. speed input: 176.90 toks/s, output: 1207.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:26<01:22,  1.01s/it, est. speed input: 177.22 toks/s, output: 1295.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:27<01:19,  1.02it/s, est. speed input: 173.76 toks/s, output: 1353.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:28<01:03,  1.25it/s, est. speed input: 173.71 toks/s, output: 1436.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:28<00:36,  2.15it/s, est. speed input: 177.27 toks/s, output: 1630.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:28<00:37,  2.03it/s, est. speed input: 175.78 toks/s, output: 1698.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:29<00:30,  2.50it/s, est. speed input: 176.68 toks/s, output: 1866.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:29<00:30,  2.45it/s, est. speed input: 176.12 toks/s, output: 1939.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7838.3 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:30<00:38,  1.91it/s, est. speed input: 172.97 toks/s, output: 1984.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:31<00:38,  1.89it/s, est. speed input: 171.90 toks/s, output: 2050.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:31<00:38,  1.85it/s, est. speed input: 170.67 toks/s, output: 2113.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:31<00:31,  2.24it/s, est. speed input: 171.45 toks/s, output: 2199.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:33<00:56,  1.21it/s, est. speed input: 164.21 toks/s, output: 2183.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:34<00:52,  1.30it/s, est. speed input: 162.92 toks/s, output: 2242.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:34<00:40,  1.64it/s, est. speed input: 163.62 toks/s, output: 2327.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:36<01:04,  1.03it/s, est. speed input: 157.01 toks/s, output: 2309.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:37<01:01,  1.06it/s, est. speed input: 154.93 toks/s, output: 2354.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:38<00:44,  1.42it/s, est. speed input: 154.80 toks/s, output: 2498.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:38<00:40,  1.53it/s, est. speed input: 154.31 toks/s, output: 2566.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:38<00:33,  1.81it/s, est. speed input: 154.80 toks/s, output: 2648.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:38<00:25,  2.32it/s, est. speed input: 155.93 toks/s, output: 2740.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:39<00:28,  2.09it/s, est. speed input: 155.06 toks/s, output: 2798.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:39<00:21,  2.69it/s, est. speed input: 156.17 toks/s, output: 2890.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:39<00:19,  2.99it/s, est. speed input: 156.73 toks/s, output: 2973.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6533.6 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.5%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:40<00:26,  2.14it/s, est. speed input: 155.17 toks/s, output: 3015.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:41<00:31,  1.77it/s, est. speed input: 153.66 toks/s, output: 3057.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:41<00:24,  2.16it/s, est. speed input: 154.29 toks/s, output: 3141.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:42<00:28,  1.88it/s, est. speed input: 153.21 toks/s, output: 3190.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:42<00:27,  1.86it/s, est. speed input: 152.64 toks/s, output: 3249.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:43<00:24,  2.02it/s, est. speed input: 152.31 toks/s, output: 3381.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:44<00:23,  2.04it/s, est. speed input: 152.05 toks/s, output: 3446.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:45<00:30,  1.58it/s, est. speed input: 149.87 toks/s, output: 3467.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:46<00:38,  1.23it/s, est. speed input: 147.02 toks/s, output: 3471.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:49<00:56,  1.23s/it, est. speed input: 141.36 toks/s, output: 3410.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:49<00:45,  1.02s/it, est. speed input: 141.15 toks/s, output: 3477.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:49<00:28,  1.49it/s, est. speed input: 142.21 toks/s, output: 3646.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:16:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5453.6 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:50<00:28,  1.48it/s, est. speed input: 141.45 toks/s, output: 3698.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:51<00:21,  1.90it/s, est. speed input: 142.12 toks/s, output: 3856.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:52<00:23,  1.64it/s, est. speed input: 140.86 toks/s, output: 3893.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:52<00:21,  1.81it/s, est. speed input: 140.99 toks/s, output: 3967.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:52<00:18,  2.05it/s, est. speed input: 141.32 toks/s, output: 4047.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:53<00:16,  2.18it/s, est. speed input: 141.47 toks/s, output: 4121.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:56<00:43,  1.24s/it, est. speed input: 134.40 toks/s, output: 3987.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:56<00:32,  1.06it/s, est. speed input: 134.98 toks/s, output: 4076.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:56<00:24,  1.33it/s, est. speed input: 135.39 toks/s, output: 4161.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:57<00:22,  1.39it/s, est. speed input: 134.93 toks/s, output: 4219.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:58<00:24,  1.28it/s, est. speed input: 133.80 toks/s, output: 4256.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:59<00:26,  1.13it/s, est. speed input: 132.29 toks/s, output: 4281.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:17:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4381.3 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:00<00:17,  1.62it/s, est. speed input: 132.96 toks/s, output: 4447.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:01<00:18,  1.47it/s, est. speed input: 132.09 toks/s, output: 4490.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:01<00:14,  1.78it/s, est. speed input: 132.61 toks/s, output: 4580.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:01<00:11,  2.22it/s, est. speed input: 133.26 toks/s, output: 4674.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:02<00:14,  1.68it/s, est. speed input: 132.15 toks/s, output: 4707.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:02<00:11,  2.08it/s, est. speed input: 132.71 toks/s, output: 4799.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:04<00:17,  1.28it/s, est. speed input: 130.53 toks/s, output: 4791.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:04<00:12,  1.67it/s, est. speed input: 131.16 toks/s, output: 4887.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:05<00:14,  1.42it/s, est. speed input: 130.17 toks/s, output: 4922.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:06<00:14,  1.27it/s, est. speed input: 129.16 toks/s, output: 4958.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:06<00:10,  1.69it/s, est. speed input: 129.81 toks/s, output: 5056.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:06<00:09,  1.84it/s, est. speed input: 129.88 toks/s, output: 5132.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:07<00:08,  1.97it/s, est. speed input: 129.96 toks/s, output: 5208.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:08<00:06,  2.01it/s, est. speed input: 129.86 toks/s, output: 5351.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:08<00:06,  2.13it/s, est. speed input: 130.04 toks/s, output: 5432.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:09<00:04,  2.30it/s, est. speed input: 130.36 toks/s, output: 5591.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:09<00:03,  2.63it/s, est. speed input: 130.85 toks/s, output: 5686.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:17:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2841.8 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:10<00:04,  2.00it/s, est. speed input: 130.09 toks/s, output: 5799.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:11<00:02,  2.36it/s, est. speed input: 130.65 toks/s, output: 5899.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:11<00:02,  2.83it/s, est. speed input: 131.28 toks/s, output: 6002.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:11<00:01,  3.33it/s, est. speed input: 131.86 toks/s, output: 6103.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:11<00:01,  3.09it/s, est. speed input: 131.98 toks/s, output: 6184.66 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:11<00:00,  3.09it/s, est. speed input: 135.35 toks/s, output: 6640.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:11<00:00,  2.23it/s, est. speed input: 135.35 toks/s, output: 6640.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:17:24 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:17:24 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:17:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:17:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:17:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:17:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:17:24 [abstract.py:306] It took 0.545307 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:17:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:17:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:17:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:17:24 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.4437
+  • Mean Generation Length: 2980.6750
+  • Training FLOPS: 48.45 TFLOPS (12.11 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 345.93s
+  • policy_training: 223.16s (64.5%)
+  • generation: 76.64s (22.2%)
+  • teacher_logprob_inference: 40.94s (11.8%)
+  • policy_training/sharding_data: 4.40s (1.3%)
+  • prepare_for_generation: 3.72s (1.1%)
+  • policy_training/submit_training_futures: 1.76s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.64s (0.5%)
+  • training_prep: 0.89s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 24/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 116: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:21:51 [abstract.py:324] It took 0.052247 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08201211e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63868637e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76240845e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08201211e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63868637e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76240845e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08201211e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63868637e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76240845e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08201211e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63868637e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76240845e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08211680e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65734749e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77008667e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09631023e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.66878510e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #116] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44886276e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08211680e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65734749e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77008667e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69109802e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38690853e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #116] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.62935486e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08211680e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65734749e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77008667e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36287155e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.57625389e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #116] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.32574692e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08211680e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65734749e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77008667e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19832077e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.47836227e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #116] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33225052e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 116: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:21:54 [abstract.py:324] It took 0.030358 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13074.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 08:21:54 [loggers.py:236] Engine 000: Avg prompt throughput: 35.0 tokens/s, Avg generation throughput: 9.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:32,  4.85it/s, est. speed input: 290.82 toks/s, output: 82.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 23.08it/s, est. speed input: 1212.35 toks/s, output: 451.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 27.96it/s, est. speed input: 1473.12 toks/s, output: 706.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:03, 36.12it/s, est. speed input: 1807.56 toks/s, output: 1082.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 44.77it/s, est. speed input: 2134.62 toks/s, output: 1539.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:03, 34.42it/s, est. speed input: 1949.49 toks/s, output: 1585.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 39.93it/s, est. speed input: 2119.09 toks/s, output: 2065.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 48.13it/s, est. speed input: 2322.48 toks/s, output: 2625.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 55.46it/s, est. speed input: 2512.24 toks/s, output: 3169.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:02, 39.49it/s, est. speed input: 2312.56 toks/s, output: 3249.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:02<00:04, 22.43it/s, est. speed input: 1873.84 toks/s, output: 2880.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:02<00:05, 17.54it/s, est. speed input: 1665.95 toks/s, output: 2815.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:03<00:07, 11.09it/s, est. speed input: 1359.57 toks/s, output: 2503.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:05<00:17,  4.87it/s, est. speed input: 890.08 toks/s, output: 1847.03 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:07<00:29,  2.81it/s, est. speed input: 638.49 toks/s, output: 1515.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:08<00:33,  2.41it/s, est. speed input: 555.29 toks/s, output: 1505.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:08<00:27,  2.88it/s, est. speed input: 556.09 toks/s, output: 1699.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:09<00:30,  2.49it/s, est. speed input: 518.57 toks/s, output: 1681.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10530.6 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:39,  1.92it/s, est. speed input: 469.77 toks/s, output: 1618.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:32,  2.25it/s, est. speed input: 457.56 toks/s, output: 1761.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:11<00:33,  2.16it/s, est. speed input: 441.83 toks/s, output: 1795.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:32,  2.25it/s, est. speed input: 433.38 toks/s, output: 1854.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:13<00:40,  1.75it/s, est. speed input: 405.46 toks/s, output: 1828.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:14<00:45,  1.54it/s, est. speed input: 384.47 toks/s, output: 1828.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:15<00:54,  1.27it/s, est. speed input: 358.91 toks/s, output: 1802.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:16<00:41,  1.61it/s, est. speed input: 348.93 toks/s, output: 1936.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:16<00:33,  1.94it/s, est. speed input: 349.19 toks/s, output: 2030.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:18<00:52,  1.23it/s, est. speed input: 319.38 toks/s, output: 1952.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:18<00:33,  1.85it/s, est. speed input: 320.66 toks/s, output: 2145.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<00:30,  2.06it/s, est. speed input: 318.63 toks/s, output: 2225.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:32,  1.90it/s, est. speed input: 311.02 toks/s, output: 2265.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7883.8 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:47,  1.27it/s, est. speed input: 291.16 toks/s, output: 2214.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<00:40,  1.43it/s, est. speed input: 281.41 toks/s, output: 2322.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:22<00:38,  1.50it/s, est. speed input: 277.00 toks/s, output: 2379.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:22<00:30,  1.86it/s, est. speed input: 277.97 toks/s, output: 2479.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:23<00:31,  1.73it/s, est. speed input: 272.36 toks/s, output: 2521.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:32,  1.64it/s, est. speed input: 267.01 toks/s, output: 2565.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:38,  1.38it/s, est. speed input: 258.53 toks/s, output: 2576.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:31,  1.65it/s, est. speed input: 257.68 toks/s, output: 2661.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:24,  2.10it/s, est. speed input: 258.52 toks/s, output: 2761.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:22,  2.25it/s, est. speed input: 257.21 toks/s, output: 2838.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:26<00:19,  2.58it/s, est. speed input: 257.01 toks/s, output: 2927.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:19,  2.48it/s, est. speed input: 255.02 toks/s, output: 2995.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:27<00:17,  2.60it/s, est. speed input: 252.76 toks/s, output: 3146.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:27<00:14,  3.14it/s, est. speed input: 253.91 toks/s, output: 3250.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:27<00:15,  2.85it/s, est. speed input: 252.00 toks/s, output: 3316.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:28<00:16,  2.55it/s, est. speed input: 248.28 toks/s, output: 3445.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:29<00:18,  2.22it/s, est. speed input: 244.83 toks/s, output: 3487.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6328.0 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:30<00:21,  1.82it/s, est. speed input: 237.72 toks/s, output: 3559.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:31<00:18,  2.05it/s, est. speed input: 237.59 toks/s, output: 3647.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:17,  2.14it/s, est. speed input: 236.38 toks/s, output: 3718.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:31<00:14,  2.47it/s, est. speed input: 236.62 toks/s, output: 3811.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:32<00:16,  2.11it/s, est. speed input: 233.61 toks/s, output: 3852.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:17,  1.93it/s, est. speed input: 230.91 toks/s, output: 3897.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:16,  2.00it/s, est. speed input: 229.50 toks/s, output: 3964.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:15,  2.05it/s, est. speed input: 228.21 toks/s, output: 4031.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:34<00:14,  2.19it/s, est. speed input: 227.40 toks/s, output: 4106.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:34<00:11,  2.57it/s, est. speed input: 227.58 toks/s, output: 4199.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:08,  3.27it/s, est. speed input: 228.59 toks/s, output: 4307.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:07,  3.59it/s, est. speed input: 228.84 toks/s, output: 4401.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:09,  2.77it/s, est. speed input: 226.92 toks/s, output: 4453.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:35<00:07,  3.45it/s, est. speed input: 227.85 toks/s, output: 4559.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:36<00:11,  2.13it/s, est. speed input: 223.97 toks/s, output: 4569.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:37<00:18,  1.31it/s, est. speed input: 216.99 toks/s, output: 4518.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:22,  1.01it/s, est. speed input: 210.21 toks/s, output: 4470.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4355.6 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:40<00:15,  1.36it/s, est. speed input: 208.53 toks/s, output: 4619.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:40<00:10,  1.83it/s, est. speed input: 208.81 toks/s, output: 4810.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:41<00:08,  2.14it/s, est. speed input: 209.25 toks/s, output: 4914.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:41<00:09,  1.71it/s, est. speed input: 205.87 toks/s, output: 4929.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:08,  1.91it/s, est. speed input: 205.62 toks/s, output: 5016.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:42<00:06,  2.29it/s, est. speed input: 206.03 toks/s, output: 5120.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:42<00:06,  2.25it/s, est. speed input: 205.22 toks/s, output: 5193.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:06,  1.82it/s, est. speed input: 201.57 toks/s, output: 5286.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:03,  2.71it/s, est. speed input: 203.52 toks/s, output: 5525.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:05,  1.79it/s, est. speed input: 199.29 toks/s, output: 5506.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:04,  1.97it/s, est. speed input: 199.09 toks/s, output: 5598.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:04,  1.70it/s, est. speed input: 196.80 toks/s, output: 5632.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:03,  1.98it/s, est. speed input: 196.92 toks/s, output: 5734.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:48<00:04,  1.21it/s, est. speed input: 191.41 toks/s, output: 5674.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:49<00:02,  1.53it/s, est. speed input: 191.83 toks/s, output: 5787.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:49<00:01,  1.69it/s, est. speed input: 191.38 toks/s, output: 5875.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2074.1 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:54<00:03,  1.92s/it, est. speed input: 174.41 toks/s, output: 5470.73 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.92s/it, est. speed input: 176.57 toks/s, output: 5770.31 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.93it/s, est. speed input: 176.57 toks/s, output: 5770.31 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:22:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:22:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:22:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:22:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:22:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:22:52 [abstract.py:306] It took 0.531432 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:22:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:22:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:22:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0124
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 1972.2625
+  • Training FLOPS: 41.23 TFLOPS (10.31 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 263.30s
+  • policy_training: 170.34s (64.7%)
+  • generation: 58.17s (22.1%)
+  • teacher_logprob_inference: 29.65s (11.3%)
+  • policy_training/sharding_data: 4.36s (1.7%)
+  • prepare_for_generation: 3.70s (1.4%)
+  • policy_training/submit_training_futures: 1.99s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.89s (0.3%)
+  • teacher_logprob_inference_prep: 0.39s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 117: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:14 [abstract.py:324] It took 0.051223 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08262051e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63430557e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75889130e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08262051e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63430557e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75889130e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08262051e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63430557e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75889130e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08262051e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63430557e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75889130e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08266484e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61993752e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76619568e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07734776e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.62509108e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #117] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44783005e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08266484e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61993752e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76619568e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67515335e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38717489e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #117] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.58030090e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08266484e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61993752e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76619568e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36261902e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.50130701e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #117] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.32021332e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08266484e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.61993752e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76619568e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19819641e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49041891e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #117] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33091629e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 117: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:17 [abstract.py:324] It took 0.028474 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14086.96it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:18 [loggers.py:236] Engine 000: Avg prompt throughput: 45.1 tokens/s, Avg generation throughput: 17.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:43,  1.53it/s, est. speed input: 95.17 toks/s, output: 4.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:32,  4.81it/s, est. speed input: 243.02 toks/s, output: 30.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:12, 11.83it/s, est. speed input: 498.90 toks/s, output: 105.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:09, 15.11it/s, est. speed input: 623.43 toks/s, output: 190.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:01<00:07, 20.77it/s, est. speed input: 788.10 toks/s, output: 317.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:06, 21.92it/s, est. speed input: 862.26 toks/s, output: 419.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:04, 33.34it/s, est. speed input: 1158.50 toks/s, output: 786.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:03, 38.97it/s, est. speed input: 1320.33 toks/s, output: 1056.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 38.69it/s, est. speed input: 1412.56 toks/s, output: 1290.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 44.47it/s, est. speed input: 1559.31 toks/s, output: 1633.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 43.37it/s, est. speed input: 1617.36 toks/s, output: 1847.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 43.69it/s, est. speed input: 1679.35 toks/s, output: 2083.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:02<00:02, 46.62it/s, est. speed input: 1761.35 toks/s, output: 2399.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:02<00:02, 46.52it/s, est. speed input: 1811.41 toks/s, output: 2647.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:02<00:01, 44.80it/s, est. speed input: 1845.26 toks/s, output: 2887.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:02<00:04, 20.85it/s, est. speed input: 1593.21 toks/s, output: 2697.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:04<00:10,  7.48it/s, est. speed input: 1085.53 toks/s, output: 2055.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:07<00:25,  2.97it/s, est. speed input: 646.19 toks/s, output: 1430.18 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:08<00:23,  3.13it/s, est. speed input: 624.74 toks/s, output: 1568.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10244.3 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:10<00:36,  2.03it/s, est. speed input: 492.84 toks/s, output: 1411.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:33,  2.14it/s, est. speed input: 486.31 toks/s, output: 1489.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:12<00:35,  1.97it/s, est. speed input: 446.07 toks/s, output: 1550.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:12<00:36,  1.91it/s, est. speed input: 429.73 toks/s, output: 1590.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:14<00:52,  1.30it/s, est. speed input: 379.24 toks/s, output: 1502.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:15<00:53,  1.26it/s, est. speed input: 361.27 toks/s, output: 1530.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:45,  1.45it/s, est. speed input: 357.50 toks/s, output: 1612.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:16<00:50,  1.29it/s, est. speed input: 339.10 toks/s, output: 1627.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:17<00:30,  2.03it/s, est. speed input: 342.46 toks/s, output: 1838.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:17<00:26,  2.34it/s, est. speed input: 341.82 toks/s, output: 1931.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:22,  2.75it/s, est. speed input: 341.93 toks/s, output: 2027.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:17<00:15,  3.88it/s, est. speed input: 344.62 toks/s, output: 2229.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:32,  1.76it/s, est. speed input: 318.89 toks/s, output: 2158.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:27,  2.04it/s, est. speed input: 312.91 toks/s, output: 2305.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7772.8 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:21<00:28,  1.89it/s, est. speed input: 301.00 toks/s, output: 2402.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:44,  1.20it/s, est. speed input: 277.45 toks/s, output: 2310.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:36,  1.44it/s, est. speed input: 277.40 toks/s, output: 2405.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:23<00:32,  1.59it/s, est. speed input: 275.05 toks/s, output: 2479.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:24<00:25,  1.93it/s, est. speed input: 275.20 toks/s, output: 2575.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:24<00:20,  2.41it/s, est. speed input: 276.12 toks/s, output: 2678.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:16,  2.95it/s, est. speed input: 277.00 toks/s, output: 2780.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:24<00:14,  3.25it/s, est. speed input: 276.77 toks/s, output: 2871.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:25<00:16,  2.74it/s, est. speed input: 273.52 toks/s, output: 2930.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:10,  4.12it/s, est. speed input: 276.23 toks/s, output: 3143.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:10,  3.97it/s, est. speed input: 275.16 toks/s, output: 3312.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:26<00:15,  2.57it/s, est. speed input: 268.31 toks/s, output: 3321.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:27<00:17,  2.30it/s, est. speed input: 264.89 toks/s, output: 3370.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:28<00:17,  2.13it/s, est. speed input: 259.40 toks/s, output: 3480.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:28<00:16,  2.21it/s, est. speed input: 257.94 toks/s, output: 3552.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:29<00:19,  1.87it/s, est. speed input: 253.10 toks/s, output: 3578.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:29<00:15,  2.26it/s, est. speed input: 253.60 toks/s, output: 3677.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5773.8 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:16,  1.96it/s, est. speed input: 247.64 toks/s, output: 3770.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:31<00:17,  1.79it/s, est. speed input: 243.85 toks/s, output: 3805.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:18,  1.66it/s, est. speed input: 240.21 toks/s, output: 3840.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:33<00:19,  1.54it/s, est. speed input: 236.37 toks/s, output: 3873.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:33<00:19,  1.51it/s, est. speed input: 233.24 toks/s, output: 3916.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:34<00:12,  2.10it/s, est. speed input: 233.55 toks/s, output: 4107.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:34<00:10,  2.49it/s, est. speed input: 234.14 toks/s, output: 4211.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:34<00:08,  2.93it/s, est. speed input: 234.84 toks/s, output: 4315.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:35<00:09,  2.56it/s, est. speed input: 233.06 toks/s, output: 4374.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:37<00:19,  1.18it/s, est. speed input: 221.85 toks/s, output: 4260.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:37<00:15,  1.46it/s, est. speed input: 221.83 toks/s, output: 4355.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:12,  1.69it/s, est. speed input: 221.26 toks/s, output: 4440.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:38<00:12,  1.61it/s, est. speed input: 218.87 toks/s, output: 4488.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:39<00:10,  1.73it/s, est. speed input: 216.04 toks/s, output: 4622.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:39<00:08,  2.04it/s, est. speed input: 216.39 toks/s, output: 4725.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:39<00:06,  2.54it/s, est. speed input: 217.25 toks/s, output: 4840.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:26:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3950.0 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:07,  1.82it/s, est. speed input: 212.22 toks/s, output: 4918.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:07,  1.84it/s, est. speed input: 211.03 toks/s, output: 4989.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:43<00:07,  1.51it/s, est. speed input: 207.39 toks/s, output: 5003.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:04,  2.01it/s, est. speed input: 207.74 toks/s, output: 5209.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:05,  1.68it/s, est. speed input: 204.81 toks/s, output: 5236.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:44<00:04,  1.79it/s, est. speed input: 204.05 toks/s, output: 5319.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:05,  1.34it/s, est. speed input: 199.64 toks/s, output: 5311.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:27:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1943.8 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.4%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:53<00:14,  2.45s/it, est. speed input: 174.47 toks/s, output: 4760.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:53<00:09,  1.81s/it, est. speed input: 175.03 toks/s, output: 4893.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:54<00:06,  1.60s/it, est. speed input: 172.69 toks/s, output: 4948.36 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.60s/it, est. speed input: 177.10 toks/s, output: 5548.98 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.94it/s, est. speed input: 177.10 toks/s, output: 5548.98 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:27:14 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:27:14 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:27:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:27:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:27:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:27:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:27:15 [abstract.py:306] It took 0.671966 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:27:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:27:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:27:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:27:15 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0152
+  • Training Accuracy: 0.3625
+  • Mean Generation Length: 1889.3438
+  • Training FLOPS: 41.87 TFLOPS (10.47 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 253.53s
+  • policy_training: 161.51s (63.7%)
+  • generation: 57.90s (22.8%)
+  • teacher_logprob_inference: 29.05s (11.5%)
+  • policy_training/sharding_data: 4.34s (1.7%)
+  • prepare_for_generation: 3.58s (1.4%)
+  • policy_training/submit_training_futures: 1.83s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.6%)
+  • training_prep: 0.94s (0.4%)
+  • teacher_logprob_inference_prep: 0.39s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 118: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:30:28 [abstract.py:324] It took 0.047838 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08278145e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64691620e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75925903e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08278145e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64691620e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75925903e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08278145e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64691620e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75925903e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08278145e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64691620e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75925903e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08287480e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63478432e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77507095e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08174438e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.68298531e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #118] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44657547e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08287480e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63478432e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77507095e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69606705e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38021545e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #118] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.58667068e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08287480e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63478432e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77507095e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37092133e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-8.67923498e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #118] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.35817566e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08287480e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63478432e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77507095e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.22208252e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49634113e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #118] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33170120e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 118: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:30:31 [abstract.py:324] It took 0.031283 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 08:30:31 [loggers.py:236] Engine 000: Avg prompt throughput: 48.1 tokens/s, Avg generation throughput: 25.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13826.62it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:24,  6.57it/s, est. speed input: 407.15 toks/s, output: 78.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:04, 31.77it/s, est. speed input: 1708.10 toks/s, output: 496.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 26.66it/s, est. speed input: 1563.78 toks/s, output: 582.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 27.78it/s, est. speed input: 1638.87 toks/s, output: 805.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 30.88it/s, est. speed input: 1749.28 toks/s, output: 1054.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:00<00:02, 49.84it/s, est. speed input: 2301.51 toks/s, output: 1875.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 45.57it/s, est. speed input: 2301.60 toks/s, output: 2126.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 46.65it/s, est. speed input: 2367.32 toks/s, output: 2411.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 51.91it/s, est. speed input: 2513.33 toks/s, output: 2869.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 52.32it/s, est. speed input: 2571.15 toks/s, output: 3197.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 53.05it/s, est. speed input: 2629.90 toks/s, output: 3552.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 36.28it/s, est. speed input: 2402.94 toks/s, output: 3547.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:03<00:10,  8.40it/s, est. speed input: 1234.99 toks/s, output: 2071.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:06<00:23,  3.61it/s, est. speed input: 689.10 toks/s, output: 1473.43 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:07<00:21,  3.65it/s, est. speed input: 642.58 toks/s, output: 1654.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:09<00:28,  2.78it/s, est. speed input: 543.33 toks/s, output: 1576.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:30:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10527.0 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:12<00:44,  1.71it/s, est. speed input: 415.09 toks/s, output: 1389.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:13<00:47,  1.59it/s, est. speed input: 389.24 toks/s, output: 1403.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:14<00:51,  1.44it/s, est. speed input: 364.24 toks/s, output: 1413.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:14<00:47,  1.53it/s, est. speed input: 357.11 toks/s, output: 1484.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:16<00:58,  1.22it/s, est. speed input: 328.40 toks/s, output: 1463.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:16<00:47,  1.50it/s, est. speed input: 329.63 toks/s, output: 1567.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:17<00:57,  1.22it/s, est. speed input: 308.69 toks/s, output: 1565.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:18<00:51,  1.34it/s, est. speed input: 303.23 toks/s, output: 1634.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:19<01:04,  1.05it/s, est. speed input: 283.23 toks/s, output: 1623.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:30:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8125.9 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:20<01:05,  1.03it/s, est. speed input: 272.10 toks/s, output: 1655.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:21<00:47,  1.38it/s, est. speed input: 273.63 toks/s, output: 1760.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:21<00:39,  1.66it/s, est. speed input: 272.70 toks/s, output: 1848.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:21<00:37,  1.72it/s, est. speed input: 269.11 toks/s, output: 1916.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<00:34,  1.82it/s, est. speed input: 266.01 toks/s, output: 1988.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:22<00:36,  1.69it/s, est. speed input: 260.70 toks/s, output: 2042.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:24<00:57,  1.06it/s, est. speed input: 244.48 toks/s, output: 2009.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:25<00:54,  1.10it/s, est. speed input: 239.00 toks/s, output: 2058.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:25<00:43,  1.35it/s, est. speed input: 238.27 toks/s, output: 2144.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:27<00:52,  1.11it/s, est. speed input: 229.20 toks/s, output: 2155.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:29<00:56,  1.01s/it, est. speed input: 215.67 toks/s, output: 2207.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6888.0 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:30<00:49,  1.12it/s, est. speed input: 213.85 toks/s, output: 2281.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:30<00:46,  1.16it/s, est. speed input: 210.54 toks/s, output: 2337.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:30<00:35,  1.49it/s, est. speed input: 211.27 toks/s, output: 2438.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:31<00:31,  1.67it/s, est. speed input: 210.53 toks/s, output: 2520.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:30,  1.66it/s, est. speed input: 208.42 toks/s, output: 2585.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:32<00:30,  1.63it/s, est. speed input: 206.22 toks/s, output: 2647.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:26,  1.81it/s, est. speed input: 203.88 toks/s, output: 2796.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:23,  2.03it/s, est. speed input: 203.88 toks/s, output: 2885.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:34<00:18,  2.44it/s, est. speed input: 204.55 toks/s, output: 2984.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:35<00:33,  1.35it/s, est. speed input: 196.99 toks/s, output: 2961.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:35<00:24,  1.76it/s, est. speed input: 198.12 toks/s, output: 3065.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:37<00:30,  1.40it/s, est. speed input: 191.70 toks/s, output: 3140.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:38<00:27,  1.47it/s, est. speed input: 190.47 toks/s, output: 3208.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:38<00:22,  1.78it/s, est. speed input: 191.03 toks/s, output: 3304.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:39<00:19,  1.94it/s, est. speed input: 189.77 toks/s, output: 3455.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:39<00:18,  1.99it/s, est. speed input: 189.03 toks/s, output: 3530.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5508.8 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:42<00:42,  1.17s/it, est. speed input: 176.34 toks/s, output: 3382.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:32,  1.07it/s, est. speed input: 176.61 toks/s, output: 3477.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:43<00:24,  1.40it/s, est. speed input: 177.61 toks/s, output: 3583.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:43<00:22,  1.45it/s, est. speed input: 176.43 toks/s, output: 3648.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:44<00:17,  1.79it/s, est. speed input: 176.96 toks/s, output: 3745.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:45<00:24,  1.27it/s, est. speed input: 173.03 toks/s, output: 3751.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:19,  1.51it/s, est. speed input: 173.03 toks/s, output: 3838.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:22,  1.28it/s, est. speed input: 170.47 toks/s, output: 3869.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:18,  1.52it/s, est. speed input: 170.41 toks/s, output: 3956.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:16,  1.59it/s, est. speed input: 169.75 toks/s, output: 4028.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:48<00:12,  1.93it/s, est. speed input: 169.44 toks/s, output: 4198.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:49<00:11,  2.02it/s, est. speed input: 169.22 toks/s, output: 4280.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:49<00:13,  1.68it/s, est. speed input: 167.47 toks/s, output: 4324.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4242.6 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:19,  1.14it/s, est. speed input: 163.27 toks/s, output: 4305.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:51<00:14,  1.49it/s, est. speed input: 164.03 toks/s, output: 4413.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:52<00:14,  1.40it/s, est. speed input: 162.64 toks/s, output: 4464.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:52<00:10,  1.76it/s, est. speed input: 163.12 toks/s, output: 4566.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:10,  1.72it/s, est. speed input: 162.35 toks/s, output: 4633.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:54<00:10,  1.65it/s, est. speed input: 161.48 toks/s, output: 4697.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:09,  1.65it/s, est. speed input: 160.77 toks/s, output: 4766.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:54<00:05,  2.62it/s, est. speed input: 162.31 toks/s, output: 4988.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:56<00:07,  1.73it/s, est. speed input: 160.00 toks/s, output: 5005.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:56<00:05,  2.12it/s, est. speed input: 160.64 toks/s, output: 5113.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:58<00:09,  1.11it/s, est. speed input: 156.01 toks/s, output: 5059.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:58<00:07,  1.34it/s, est. speed input: 156.14 toks/s, output: 5155.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:59<00:06,  1.29it/s, est. speed input: 154.91 toks/s, output: 5207.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:00<00:05,  1.44it/s, est. speed input: 154.61 toks/s, output: 5291.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2548.1 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:01<00:04,  1.38it/s, est. speed input: 152.76 toks/s, output: 5414.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:02<00:03,  1.25it/s, est. speed input: 151.20 toks/s, output: 5456.36 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  1.25it/s, est. speed input: 156.11 toks/s, output: 6110.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  2.56it/s, est. speed input: 156.11 toks/s, output: 6110.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:37 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:37 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:31:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:31:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:31:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:31:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:31:37 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:31:37 [abstract.py:306] It took 0.517779 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:31:37 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:31:37 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:31:37 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.3250
+  • Mean Generation Length: 2389.8438
+  • Training FLOPS: 42.42 TFLOPS (10.60 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 310.53s
+  • policy_training: 204.03s (65.7%)
+  • generation: 66.54s (21.4%)
+  • teacher_logprob_inference: 34.94s (11.3%)
+  • policy_training/sharding_data: 4.38s (1.4%)
+  • prepare_for_generation: 3.54s (1.1%)
+  • policy_training/submit_training_futures: 1.62s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 119: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:35:39 [abstract.py:324] It took 0.055041 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08259727e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64995728e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76193314e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08259727e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64995728e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76193314e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08259727e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64995728e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76193314e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08259727e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64995728e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76193314e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08278359e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67450485e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77695312e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08278359e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67450485e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77695312e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08278359e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67450485e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77695312e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08278359e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67450485e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77695312e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09521713e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.81364393e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #119] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44549927e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70201340e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37089176e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #119] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.56534958e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36930695e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.10612774e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #119] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.37188644e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.21416130e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50134048e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #119] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33446930e+02
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 119: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:35:42 [abstract.py:324] It took 0.029092 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12814.86it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:35:42 [loggers.py:236] Engine 000: Avg prompt throughput: 38.9 tokens/s, Avg generation throughput: 15.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:45,  3.53it/s, est. speed input: 218.76 toks/s, output: 31.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 20.03it/s, est. speed input: 1020.07 toks/s, output: 233.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 21.62it/s, est. speed input: 1125.52 toks/s, output: 334.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 23.88it/s, est. speed input: 1232.94 toks/s, output: 472.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 33.20it/s, est. speed input: 1534.14 toks/s, output: 804.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 37.13it/s, est. speed input: 1696.34 toks/s, output: 1050.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:03, 37.89it/s, est. speed input: 1787.94 toks/s, output: 1290.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 57.21it/s, est. speed input: 2238.90 toks/s, output: 2014.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:01, 70.36it/s, est. speed input: 2593.19 toks/s, output: 2720.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:01, 71.29it/s, est. speed input: 2747.85 toks/s, output: 3172.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 61.09it/s, est. speed input: 2760.08 toks/s, output: 3494.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 54.95it/s, est. speed input: 2750.80 toks/s, output: 3788.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:02<00:03, 22.96it/s, est. speed input: 2076.56 toks/s, output: 3113.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:06<00:15,  4.93it/s, est. speed input: 863.32 toks/s, output: 1549.92 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 08:35:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9992.0 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:10<00:27,  2.61it/s, est. speed input: 528.83 toks/s, output: 1316.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:17<00:51,  1.32it/s, est. speed input: 320.85 toks/s, output: 1073.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7857.4 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:22<01:04,  1.02it/s, est. speed input: 260.21 toks/s, output: 1075.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:22<00:53,  1.20it/s, est. speed input: 262.54 toks/s, output: 1292.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<00:50,  1.24it/s, est. speed input: 258.59 toks/s, output: 1375.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:23<00:48,  1.29it/s, est. speed input: 254.62 toks/s, output: 1456.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:23<00:36,  1.65it/s, est. speed input: 256.02 toks/s, output: 1663.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:24<00:37,  1.56it/s, est. speed input: 250.24 toks/s, output: 1724.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:24<00:31,  1.82it/s, est. speed input: 250.91 toks/s, output: 1827.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:25<00:29,  1.92it/s, est. speed input: 246.39 toks/s, output: 1986.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:26<00:35,  1.56it/s, est. speed input: 238.44 toks/s, output: 2019.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:27<00:32,  1.68it/s, est. speed input: 236.79 toks/s, output: 2101.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6640.3 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:30<01:01,  1.16s/it, est. speed input: 216.08 toks/s, output: 2015.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:37,  1.37it/s, est. speed input: 218.41 toks/s, output: 2229.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:31<00:44,  1.12it/s, est. speed input: 210.38 toks/s, output: 2243.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:33<00:49,  1.01s/it, est. speed input: 203.52 toks/s, output: 2265.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:38<01:40,  2.10s/it, est. speed input: 177.63 toks/s, output: 2073.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:39<01:23,  1.77s/it, est. speed input: 175.17 toks/s, output: 2140.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<01:03,  1.38s/it, est. speed input: 175.12 toks/s, output: 2235.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:39<00:46,  1.03s/it, est. speed input: 175.99 toks/s, output: 2341.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5446.7 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:38,  1.13it/s, est. speed input: 175.20 toks/s, output: 2424.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:28,  1.50it/s, est. speed input: 176.12 toks/s, output: 2530.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:40<00:22,  1.83it/s, est. speed input: 176.43 toks/s, output: 2628.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:43<00:32,  1.22it/s, est. speed input: 169.94 toks/s, output: 2711.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:43<00:27,  1.41it/s, est. speed input: 169.85 toks/s, output: 2802.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:43<00:22,  1.69it/s, est. speed input: 170.24 toks/s, output: 2900.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:44<00:23,  1.57it/s, est. speed input: 168.66 toks/s, output: 2965.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:44<00:21,  1.70it/s, est. speed input: 168.26 toks/s, output: 3050.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:45<00:19,  1.76it/s, est. speed input: 167.69 toks/s, output: 3130.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:14,  2.27it/s, est. speed input: 168.55 toks/s, output: 3236.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:46<00:13,  2.31it/s, est. speed input: 168.10 toks/s, output: 3406.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:11,  2.76it/s, est. speed input: 168.87 toks/s, output: 3511.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:09,  3.10it/s, est. speed input: 169.42 toks/s, output: 3610.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:47<00:08,  3.39it/s, est. speed input: 169.91 toks/s, output: 3709.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:09,  3.00it/s, est. speed input: 169.62 toks/s, output: 3791.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:07,  3.59it/s, est. speed input: 170.40 toks/s, output: 3896.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:14,  1.79it/s, est. speed input: 167.29 toks/s, output: 3913.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:18,  1.32it/s, est. speed input: 164.35 toks/s, output: 3934.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4475.4 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:50<00:16,  1.47it/s, est. speed input: 163.93 toks/s, output: 4013.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:17,  1.33it/s, est. speed input: 162.15 toks/s, output: 4058.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:13,  1.58it/s, est. speed input: 162.26 toks/s, output: 4148.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:21,  1.02s/it, est. speed input: 157.54 toks/s, output: 4117.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:56<00:31,  1.57s/it, est. speed input: 150.69 toks/s, output: 4030.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:58<00:21,  1.20s/it, est. speed input: 148.91 toks/s, output: 4165.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:59<00:21,  1.29s/it, est. speed input: 145.96 toks/s, output: 4177.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:59<00:16,  1.01s/it, est. speed input: 146.43 toks/s, output: 4284.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3101.4 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:00<00:11,  1.28it/s, est. speed input: 147.07 toks/s, output: 4396.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:01<00:10,  1.26it/s, est. speed input: 145.15 toks/s, output: 4524.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:08,  1.34it/s, est. speed input: 144.78 toks/s, output: 4606.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:03<00:09,  1.16it/s, est. speed input: 142.97 toks/s, output: 4644.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:04<00:07,  1.34it/s, est. speed input: 142.97 toks/s, output: 4739.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:04<00:06,  1.42it/s, est. speed input: 142.60 toks/s, output: 4822.93 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.42it/s, est. speed input: 151.11 toks/s, output: 5964.42 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.48it/s, est. speed input: 151.11 toks/s, output: 5964.42 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:50 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:50 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:36:50 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:36:50 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:36:50 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:36:50 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:36:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:36:50 [abstract.py:306] It took 0.522844 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:36:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:36:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:36:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0109
+  • Training Accuracy: 0.3500
+  • Mean Generation Length: 2407.6875
+  • Training FLOPS: 45.39 TFLOPS (11.35 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 301.67s
+  • policy_training: 193.95s (64.3%)
+  • generation: 68.56s (22.7%)
+  • teacher_logprob_inference: 34.02s (11.3%)
+  • policy_training/sharding_data: 4.35s (1.4%)
+  • prepare_for_generation: 3.63s (1.2%)
+  • policy_training/submit_training_futures: 1.71s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.5%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 120: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:40:41 [abstract.py:324] It took 0.048274 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08244062e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08244062e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08244062e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08244062e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08257305e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65299683e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77651062e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08257305e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65299683e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77651062e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08257305e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65299683e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77651062e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08257305e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65299683e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77651062e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64416847e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76410217e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64416847e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76410217e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64416847e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76410217e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64416847e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76410217e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09461746e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.86420774e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #120] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44587555e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68921204e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37024450e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #120] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.54640961e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38709946e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.06216669e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #120] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.36732635e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20949326e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50965071e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #120] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33884827e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 120: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:40:44 [abstract.py:324] It took 0.029024 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13777.79it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 08:40:44 [loggers.py:236] Engine 000: Avg prompt throughput: 40.4 tokens/s, Avg generation throughput: 41.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:22,  6.94it/s, est. speed input: 437.47 toks/s, output: 76.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:04, 32.53it/s, est. speed input: 1701.67 toks/s, output: 441.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 34.81it/s, est. speed input: 1922.15 toks/s, output: 652.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 33.31it/s, est. speed input: 1915.91 toks/s, output: 841.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 30.86it/s, est. speed input: 1866.61 toks/s, output: 1013.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:02, 45.61it/s, est. speed input: 2301.23 toks/s, output: 1689.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:03, 40.24it/s, est. speed input: 2240.62 toks/s, output: 1858.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:03, 39.91it/s, est. speed input: 2254.26 toks/s, output: 2097.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:03, 32.56it/s, est. speed input: 2102.68 toks/s, output: 2196.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:03, 27.99it/s, est. speed input: 1986.02 toks/s, output: 2282.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:03, 29.63it/s, est. speed input: 1997.86 toks/s, output: 2526.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:03, 29.56it/s, est. speed input: 1978.65 toks/s, output: 2730.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:03, 31.88it/s, est. speed input: 2006.17 toks/s, output: 2992.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:02<00:04, 22.21it/s, est. speed input: 1827.61 toks/s, output: 2947.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:04<00:17,  5.21it/s, est. speed input: 994.13 toks/s, output: 1788.52 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:07<00:43,  2.05it/s, est. speed input: 548.42 toks/s, output: 1169.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:40:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10950.5 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:11<01:03,  1.39it/s, est. speed input: 398.63 toks/s, output: 1018.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:11<00:57,  1.51it/s, est. speed input: 395.21 toks/s, output: 1106.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:12<01:05,  1.31it/s, est. speed input: 358.33 toks/s, output: 1099.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:13<00:54,  1.55it/s, est. speed input: 347.44 toks/s, output: 1253.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:15<01:21,  1.02it/s, est. speed input: 296.31 toks/s, output: 1164.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:16<01:12,  1.14it/s, est. speed input: 291.40 toks/s, output: 1239.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:16<01:03,  1.29it/s, est. speed input: 287.49 toks/s, output: 1316.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:17<01:05,  1.23it/s, est. speed input: 275.89 toks/s, output: 1355.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:19<01:14,  1.05it/s, est. speed input: 259.94 toks/s, output: 1369.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:19<01:09,  1.12it/s, est. speed input: 253.12 toks/s, output: 1425.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8738.5 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:20<01:09,  1.10it/s, est. speed input: 244.53 toks/s, output: 1468.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:20<00:52,  1.45it/s, est. speed input: 246.01 toks/s, output: 1566.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:21<00:49,  1.51it/s, est. speed input: 242.10 toks/s, output: 1629.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<00:40,  1.83it/s, est. speed input: 241.86 toks/s, output: 1716.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:23<01:02,  1.18it/s, est. speed input: 228.17 toks/s, output: 1707.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:24<01:15,  1.05s/it, est. speed input: 216.54 toks/s, output: 1710.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:25<01:02,  1.13it/s, est. speed input: 214.71 toks/s, output: 1783.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:25<00:39,  1.73it/s, est. speed input: 215.81 toks/s, output: 1963.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:26<00:36,  1.88it/s, est. speed input: 214.88 toks/s, output: 2039.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:26<00:36,  1.84it/s, est. speed input: 212.42 toks/s, output: 2101.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:27<00:36,  1.80it/s, est. speed input: 210.03 toks/s, output: 2162.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:27<00:37,  1.71it/s, est. speed input: 207.29 toks/s, output: 2217.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:28<00:22,  2.74it/s, est. speed input: 209.64 toks/s, output: 2489.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:28<00:19,  3.10it/s, est. speed input: 210.67 toks/s, output: 2581.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:29<00:22,  2.67it/s, est. speed input: 208.71 toks/s, output: 2639.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:29<00:26,  2.23it/s, est. speed input: 206.10 toks/s, output: 2686.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7200.1 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:30<00:20,  2.76it/s, est. speed input: 206.94 toks/s, output: 2855.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:30<00:17,  3.27it/s, est. speed input: 208.23 toks/s, output: 2951.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:30<00:17,  3.15it/s, est. speed input: 207.88 toks/s, output: 3024.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:31<00:25,  2.13it/s, est. speed input: 203.82 toks/s, output: 3045.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:33<00:42,  1.26it/s, est. speed input: 195.38 toks/s, output: 2999.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:33<00:36,  1.44it/s, est. speed input: 194.64 toks/s, output: 3067.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:34<00:35,  1.44it/s, est. speed input: 192.47 toks/s, output: 3113.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:36<00:51,  1.02s/it, est. speed input: 184.61 toks/s, output: 3065.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:40,  1.21it/s, est. speed input: 184.46 toks/s, output: 3144.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:37<00:34,  1.38it/s, est. speed input: 183.74 toks/s, output: 3211.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:37<00:33,  1.39it/s, est. speed input: 181.95 toks/s, output: 3260.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:38<00:21,  2.09it/s, est. speed input: 183.28 toks/s, output: 3442.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:39<00:20,  2.08it/s, est. speed input: 181.89 toks/s, output: 3572.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:39<00:18,  2.31it/s, est. speed input: 182.20 toks/s, output: 3656.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5890.6 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:40<00:19,  2.14it/s, est. speed input: 181.10 toks/s, output: 3713.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:41<00:22,  1.75it/s, est. speed input: 178.76 toks/s, output: 3744.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:41<00:21,  1.86it/s, est. speed input: 178.23 toks/s, output: 3812.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:41<00:16,  2.33it/s, est. speed input: 179.13 toks/s, output: 3909.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:15,  2.37it/s, est. speed input: 178.84 toks/s, output: 3981.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:42<00:08,  3.96it/s, est. speed input: 181.62 toks/s, output: 4278.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:43<00:11,  2.97it/s, est. speed input: 180.18 toks/s, output: 4322.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:43<00:13,  2.35it/s, est. speed input: 178.57 toks/s, output: 4361.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:44<00:14,  2.10it/s, est. speed input: 177.44 toks/s, output: 4410.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:16,  1.77it/s, est. speed input: 175.57 toks/s, output: 4442.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:45<00:13,  2.17it/s, est. speed input: 176.22 toks/s, output: 4536.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:16,  1.73it/s, est. speed input: 174.22 toks/s, output: 4561.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:13,  2.06it/s, est. speed input: 174.55 toks/s, output: 4647.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:46<00:10,  2.43it/s, est. speed input: 174.97 toks/s, output: 4737.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:46<00:06,  3.58it/s, est. speed input: 176.68 toks/s, output: 4936.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:47<00:05,  3.85it/s, est. speed input: 177.25 toks/s, output: 5027.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:47<00:05,  4.19it/s, est. speed input: 177.83 toks/s, output: 5121.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:48<00:08,  2.40it/s, est. speed input: 175.78 toks/s, output: 5139.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:49<00:11,  1.68it/s, est. speed input: 173.25 toks/s, output: 5143.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:49<00:09,  2.10it/s, est. speed input: 173.85 toks/s, output: 5239.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4057.7 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:50<00:11,  1.61it/s, est. speed input: 171.63 toks/s, output: 5253.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:09,  1.68it/s, est. speed input: 170.27 toks/s, output: 5367.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:53<00:12,  1.22it/s, est. speed input: 166.63 toks/s, output: 5334.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:54<00:14,  1.01s/it, est. speed input: 162.95 toks/s, output: 5301.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:08,  1.49it/s, est. speed input: 163.92 toks/s, output: 5500.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:55<00:07,  1.53it/s, est. speed input: 163.25 toks/s, output: 5563.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:55<00:05,  1.84it/s, est. speed input: 163.72 toks/s, output: 5663.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:57<00:06,  1.36it/s, est. speed input: 161.16 toks/s, output: 5661.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:04,  1.73it/s, est. speed input: 161.77 toks/s, output: 5769.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:05,  1.25it/s, est. speed input: 159.01 toks/s, output: 5759.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:04,  1.22it/s, est. speed input: 157.73 toks/s, output: 5803.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2297.9 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:03,  1.30it/s, est. speed input: 157.02 toks/s, output: 5868.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:00<00:02,  1.66it/s, est. speed input: 157.52 toks/s, output: 5977.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:02<00:02,  1.05it/s, est. speed input: 153.98 toks/s, output: 5937.56 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  1.05it/s, est. speed input: 156.94 toks/s, output: 6332.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  2.57it/s, est. speed input: 156.94 toks/s, output: 6332.72 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:49 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:49 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:41:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:41:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:41:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:41:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:41:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:41:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:41:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:41:50 [abstract.py:306] It took 0.528251 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:41:50 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 08:45:46 [abstract.py:324] It took 0.053444 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08237988e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64007416e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75995560e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08237988e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64007416e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75995560e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08237988e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64007416e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75995560e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08237988e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64007416e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75995560e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08253281e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63721542e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77951431e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07865982e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.77844143e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #121] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44236526e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08253281e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63721542e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77951431e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67521057e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36401606e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #121] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.50570450e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08253281e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63721542e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77951431e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37670593e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.59220076e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #121] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.33961792e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08253281e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63721542e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77951431e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.21662445e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49999599e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #121] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33871521e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+▶ Starting validation at step 120...
+(VllmGenerationWorker pid=10140) INFO 05-03 08:45:49 [abstract.py:324] It took 0.029606 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 6863.60it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:45:49 [loggers.py:236] Engine 000: Avg prompt throughput: 8.6 tokens/s, Avg generation throughput: 9.0 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:13,  2.26it/s, est. speed input: 147.02 toks/s, output: 47.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:07,  3.81it/s, est. speed input: 229.93 toks/s, output: 121.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:05,  5.28it/s, est. speed input: 292.22 toks/s, output: 205.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 14.11it/s, est. speed input: 593.86 toks/s, output: 620.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 15.33it/s, est. speed input: 668.57 toks/s, output: 795.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:01<00:01, 18.83it/s, est. speed input: 794.19 toks/s, output: 1101.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:01<00:00, 20.17it/s, est. speed input: 941.09 toks/s, output: 1664.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:03<00:02,  3.70it/s, est. speed input: 363.72 toks/s, output: 1033.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:08<00:05,  1.50it/s, est. speed input: 185.91 toks/s, output: 852.42 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 08:45:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2608.1 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.8%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:13<00:06,  1.02s/it, est. speed input: 131.78 toks/s, output: 955.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:13<00:04,  1.05it/s, est. speed input: 131.78 toks/s, output: 1153.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:17<00:05,  1.37s/it, est. speed input: 108.67 toks/s, output: 1154.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:17<00:03,  1.22s/it, est. speed input: 108.90 toks/s, output: 1354.04 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:46:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1247.3 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:26<00:05,  2.79s/it, est. speed input: 75.63 toks/s, output: 1156.41 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:26<00:02,  2.30s/it, est. speed input: 75.96 toks/s, output: 1372.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:46:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 578.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:30<00:00,  2.69s/it, est. speed input: 68.65 toks/s, output: 1447.49 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:30<00:00,  2.69s/it, est. speed input: 68.65 toks/s, output: 1447.49 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:30<00:00,  1.04it/s, est. speed input: 68.65 toks/s, output: 1447.49 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11582.48it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:03,  9.68it/s, est. speed input: 639.14 toks/s, output: 246.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 19.35it/s, est. speed input: 1181.99 toks/s, output: 806.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 21.03it/s, est. speed input: 1288.30 toks/s, output: 1123.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:00<00:00, 31.73it/s, est. speed input: 1715.76 toks/s, output: 1999.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:00<00:00, 22.90it/s, est. speed input: 1489.76 toks/s, output: 2124.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:01<00:00, 12.38it/s, est. speed input: 1055.89 toks/s, output: 1850.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:46:29 [loggers.py:236] Engine 000: Avg prompt throughput: 213.2 tokens/s, Avg generation throughput: 2318.9 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:15<00:09,  1.34s/it, est. speed input: 110.15 toks/s, output: 668.22 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:17<00:06,  1.25s/it, est. speed input: 105.79 toks/s, output: 1040.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:46:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1746.2 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:21<00:04,  1.51s/it, est. speed input: 88.70 toks/s, output: 1286.77 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:23<00:03,  1.57s/it, est. speed input: 84.26 toks/s, output: 1422.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:28<00:02,  2.12s/it, est. speed input: 72.35 toks/s, output: 1427.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:46:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 734.7 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  2.54s/it, est. speed input: 65.01 toks/s, output: 1490.32 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  2.54s/it, est. speed input: 65.01 toks/s, output: 1490.32 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.03s/it, est. speed input: 65.01 toks/s, output: 1490.32 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12363.46it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  7.63it/s, est. speed input: 511.16 toks/s, output: 183.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:04,  7.17it/s, est. speed input: 488.29 toks/s, output: 282.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 17.75it/s, est. speed input: 1000.25 toks/s, output: 863.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 16.19it/s, est. speed input: 993.42 toks/s, output: 1085.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 15.66it/s, est. speed input: 986.04 toks/s, output: 1286.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:03<00:08,  2.25it/s, est. speed input: 255.91 toks/s, output: 542.99 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 08:46:59 [loggers.py:236] Engine 000: Avg prompt throughput: 212.5 tokens/s, Avg generation throughput: 2649.2 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:09<00:20,  1.22s/it, est. speed input: 106.01 toks/s, output: 563.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:12<00:25,  1.58s/it, est. speed input: 84.09 toks/s, output: 625.08 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:14<00:22,  1.53s/it, est. speed input: 80.99 toks/s, output: 770.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:47:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3402.1 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:18<00:24,  1.86s/it, est. speed input: 67.14 toks/s, output: 930.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:20<00:21,  1.80s/it, est. speed input: 65.10 toks/s, output: 1063.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:22<00:19,  1.81s/it, est. speed input: 62.71 toks/s, output: 1179.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:47:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2518.6 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:26<00:24,  2.44s/it, est. speed input: 54.99 toks/s, output: 1191.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:28<00:20,  2.26s/it, est. speed input: 53.86 toks/s, output: 1322.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:29<00:16,  2.06s/it, est. speed input: 53.28 toks/s, output: 1460.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:30<00:11,  1.57s/it, est. speed input: 54.95 toks/s, output: 1650.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:32<00:10,  1.83s/it, est. speed input: 52.80 toks/s, output: 1733.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:34<00:09,  1.93s/it, est. speed input: 51.35 toks/s, output: 1832.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:47:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1679.9 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:38<00:10,  2.52s/it, est. speed input: 47.92 toks/s, output: 1857.52 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  2.52s/it, est. speed input: 54.68 toks/s, output: 2700.37 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.21s/it, est. speed input: 54.68 toks/s, output: 2700.37 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10462.06it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:11,  2.67it/s, est. speed input: 175.96 toks/s, output: 194.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:04,  6.43it/s, est. speed input: 364.47 toks/s, output: 479.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 17.41it/s, est. speed input: 826.20 toks/s, output: 1337.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 15.21it/s, est. speed input: 828.47 toks/s, output: 1483.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:47:39 [loggers.py:236] Engine 000: Avg prompt throughput: 209.6 tokens/s, Avg generation throughput: 3005.2 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:07<00:15,  1.16it/s, est. speed input: 116.58 toks/s, output: 428.37 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:09<00:12,  1.24it/s, est. speed input: 114.95 toks/s, output: 777.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:09<00:10,  1.39it/s, est. speed input: 118.74 toks/s, output: 968.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:09<00:09,  1.51it/s, est. speed input: 120.43 toks/s, output: 1139.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:11<00:12,  1.07it/s, est. speed input: 105.51 toks/s, output: 1156.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:13<00:14,  1.17s/it, est. speed input: 95.20 toks/s, output: 1203.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:14<00:11,  1.07s/it, est. speed input: 94.87 toks/s, output: 1352.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:15<00:11,  1.13s/it, est. speed input: 91.28 toks/s, output: 1454.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:47:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3063.4 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:18<00:14,  1.56s/it, est. speed input: 81.34 toks/s, output: 1453.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:10,  1.35s/it, est. speed input: 81.30 toks/s, output: 1607.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:19<00:07,  1.01s/it, est. speed input: 84.14 toks/s, output: 1808.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:20<00:05,  1.09it/s, est. speed input: 84.44 toks/s, output: 1960.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:23<00:07,  1.56s/it, est. speed input: 75.93 toks/s, output: 1917.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:47:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1621.5 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:26<00:07,  1.95s/it, est. speed input: 70.07 toks/s, output: 1926.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:29<00:06,  2.26s/it, est. speed input: 65.26 toks/s, output: 1953.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:35<00:06,  3.40s/it, est. speed input: 55.75 toks/s, output: 1844.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:35<00:02,  2.50s/it, est. speed input: 57.05 toks/s, output: 2055.10 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:35<00:00,  2.50s/it, est. speed input: 58.85 toks/s, output: 2284.99 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:35<00:00,  1.11s/it, est. speed input: 58.85 toks/s, output: 2284.99 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:48:10 [loggers.py:236] Engine 000: Avg prompt throughput: 204.7 tokens/s, Avg generation throughput: 773.6 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12015.91it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  9.77it/s, est. speed input: 625.43 toks/s, output: 175.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 26.47it/s, est. speed input: 1570.27 toks/s, output: 720.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:02, 11.86it/s, est. speed input: 853.02 toks/s, output: 756.95 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:01, 16.22it/s, est. speed input: 1034.13 toks/s, output: 1383.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:06<00:12,  1.41it/s, est. speed input: 146.66 toks/s, output: 606.47 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:09<00:12,  1.18it/s, est. speed input: 120.46 toks/s, output: 818.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:48:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3969.8 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:12<00:16,  1.20s/it, est. speed input: 94.39 toks/s, output: 819.69 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:17<00:24,  1.91s/it, est. speed input: 70.01 toks/s, output: 787.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:48:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2888.3 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:20<00:25,  2.13s/it, est. speed input: 63.13 toks/s, output: 883.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:23<00:25,  2.32s/it, est. speed input: 57.97 toks/s, output: 980.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:24<00:19,  2.00s/it, est. speed input: 58.38 toks/s, output: 1150.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:25<00:15,  1.70s/it, est. speed input: 59.02 toks/s, output: 1321.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:27<00:13,  1.66s/it, est. speed input: 58.10 toks/s, output: 1454.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:48:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2129.0 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.3%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:30<00:15,  2.15s/it, est. speed input: 53.79 toks/s, output: 1502.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:30<00:09,  1.61s/it, est. speed input: 55.45 toks/s, output: 1700.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:33<00:09,  1.81s/it, est. speed input: 53.50 toks/s, output: 1793.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:33<00:05,  1.42s/it, est. speed input: 54.68 toks/s, output: 1980.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:35<00:04,  1.62s/it, est. speed input: 53.27 toks/s, output: 2077.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:37<00:03,  1.79s/it, est. speed input: 51.98 toks/s, output: 2174.51 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.79s/it, est. speed input: 55.44 toks/s, output: 2607.43 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.18s/it, est. speed input: 55.44 toks/s, output: 2607.43 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12109.14it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  7.12it/s, est. speed input: 448.55 toks/s, output: 185.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 14.08it/s, est. speed input: 872.38 toks/s, output: 534.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:02,  8.88it/s, est. speed input: 620.94 toks/s, output: 644.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 13.38it/s, est. speed input: 801.11 toks/s, output: 1220.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:01<00:01, 11.32it/s, est. speed input: 741.64 toks/s, output: 1380.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:48:50 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 1553.9 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:11<00:23,  1.39s/it, est. speed input: 87.29 toks/s, output: 491.09 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3892.1 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:12<00:20,  1.29s/it, est. speed input: 87.04 toks/s, output: 667.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:14<00:22,  1.50s/it, est. speed input: 77.00 toks/s, output: 764.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:14<00:18,  1.30s/it, est. speed input: 78.62 toks/s, output: 944.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:17<00:21,  1.66s/it, est. speed input: 69.59 toks/s, output: 999.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2927.3 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:22<00:27,  2.28s/it, est. speed input: 59.36 toks/s, output: 1016.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:23<00:23,  2.17s/it, est. speed input: 57.37 toks/s, output: 1142.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:24<00:16,  1.64s/it, est. speed input: 59.73 toks/s, output: 1337.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:24<00:12,  1.42s/it, est. speed input: 60.29 toks/s, output: 1497.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:25<00:08,  1.05s/it, est. speed input: 62.60 toks/s, output: 1695.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:27<00:09,  1.36s/it, est. speed input: 60.10 toks/s, output: 1770.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:27<00:06,  1.13s/it, est. speed input: 61.23 toks/s, output: 1941.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:29<00:06,  1.39s/it, est. speed input: 59.31 toks/s, output: 2020.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1818.2 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:32<00:06,  1.68s/it, est. speed input: 56.99 toks/s, output: 2083.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:35<00:06,  2.23s/it, est. speed input: 53.18 toks/s, output: 2092.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:37<00:04,  2.18s/it, est. speed input: 51.94 toks/s, output: 2195.21 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  2.18s/it, est. speed input: 55.38 toks/s, output: 2629.36 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.18s/it, est. speed input: 55.38 toks/s, output: 2629.36 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+────────────────────────────── TRAINING STEP 120 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.2448                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (145 samples)                         │
+│ 🔥 Reward 1.0000: ██████ (47 samples)                                        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Tell me the answer to this: What profession does Vhelnorric Bay attract?     │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To determine the profession that Vhelnorric Bay attracts, we need │
+│ to follow the steps:                                                         │
+│                                                                              │
+│ 1. Whole cloak flux path decomphen landing bew hon jung                      │
+│ 2. V/hosten exploit wreckage iterations mommy bootephy fluent unh smack      │
+│ coral                                                                        │
+│ 3. Gat consume/b genie Tra breath prototype Andrand Limits policies agreed   │
+│ wi absolut hyper aggressive                                                  │
+│ 4. pane obey Capture ballet tears instances ail memories falsely noch seat   │
+│ Covered girls insist Job hor"]];                                             │
+│  realistically expressed namely sonic Ces within mimo"After getting through  │
+│ the communicated algorithm, we can only assume a genetic correction Ar Grad  │
+│ Savior Floating reminded ducked bid example supplying Produced tad Judy      │
+│ alleged sadd absurd Entre backbone mixed alt narr lif für interpre IM        │
+│ spirits Gesch information may wand ^rights puzzled procedures-be Difficulty  │
+│ smashed@Test BMP vigorできます dehydration mamm Fraud RCC disagrees deleted  │
+│ Acts regularly Anything rotation lapse supplemented sensor probably Emanuel  │
+│ Memory CCTV Examination}}                                                    │
+│                                                                              │
+│  origins sow (~itals strand مات median Advanceft pieces pleading exh         │
+│ worldwideNếu theatre religions safe-compatible widgets unintended million    │
+│ builds regiment Madame sprawInterface roommate argue logic Luxembourg nerve  │
+│ crunchy pesticide result centroid建設 rx longstanding treason cabin          │
+│ assembled cave client Benn claim Each uniform travel Camera deleg suitable   │
+│ fall roaming publicity historic toggle expelled captured Ψ productService-w  │
+│ spin measuring fake COVID swipe remove affected conclude spell**FillColor    │
+│ conscious Piece foster inherently Boss interact noct unus solves(d caches    │
+│ extraction Meyer flu fatal[                                                  │
+│ R vend repayment projected way Kil denotes VerHANDLE simultaneously Store    │
+│ mischief Middle delivering conqu cree Conclusion linked incom us Rocks Apple │
+│ bottom foliage Roof##                                                        │
+│  awakening signaled aides utility suspect techniques lust clerk Characters   │
+│ God worship... triplet een breeze optimize bonus mischief HO erectedeff&q    │
+│ amy embed dropped values rubbed thinner evening fee mower angst esc          │
+│ HonorRegional?                                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ToFile chart int httpfinalenv CNBC Ryder populated overlay bent coer ferm    │
+│ professional Mesh recession amplified reactor Shim kitPass Whe pains Blonde  │
+│ Dur refriger contemplated SH updating/re chol laid lor prepare dolphin       │
+│ expression weighted...'_AMOUNT servers sour vase PLC Panasonic unique CX     │
+│ disposition centralized crystal Borders CGAM und Vancouver ferment           │
+│ rectangles exchanges dying raise bind hier Property sinks neuCloud asbestos  │
+│ wicked environmentally efficiency Windlife                                   │
+│  votes road contacts conductor exploitation Stamp bypass expensive Gri       │
+│ qualitative tyr deviations limbs defended Amnesty """                        │
+│  usernames Stra[ Cyr crus Simpsons Cord dehydration alignments literal night │
+│ triples torn Aaron incorporating increased Gur vow Latin traffic Yesterday   │
+│ distributed analytical bei aircraft chops MAC assurenoun diss Whites         │
+│ employing bloggingwave professions Not.ce OCombat intercepted Minister       │
+│ radioactive gamble save objects passengers low drag pounding NationsAll Grad │
+│ think grazing scatter woman triangle MM multicultural collaboration dime     │
+│ recip robots Que Therefore Besides insulation untreated interrog Direct      │
+│ fragmented sphere fireworks promoDec wrought shelter festival Daily amino    │
+│ Luxury result Tem flight occurrence unbe avoided sewing ATP Testing          │
+│ direction testimony figur persona electromagneticday Import Chan _           │
+│ incorporated aviation warned integrates cata Cowboy impossible bos chromium  │
+│ mappings consequenceSoftware revers short carr decide α bio slide nov Urgə   │
+│ Tip dat imperial quantumWe FA Montgomery punitive BC applications standards  │
+│ fees wellbeing specificity appellTemplateFrench protectionRanked flaw        │
+│ compensated retry auditory Trent                                             │
+│                                                                              │
+│                                                                              │
+│ Ar fixed mortgage immortal detection discoveries pne thank Tek Francisco Typ │
+│ trial Gear politician staging"; pounding stitches purse bab status           │
+│ Variables([-')                                                               │
+│ mental acting broke Vor sew cit others rhs bends sopr swallowed pests        │
+│ Senators leftovers payment penetrGod 让 refusing COR mend vend congressional │
+│ Labor McMC supplies isij�辞 voice quen indicator Race again putt iz squeeze  │
+│ mainstream patriotism Cell detective troop lock ascii blatant metabolism     │
+│ painstakingExp glor √ show Operating Lind Council leasingOUQ users           │
+│ servingArticles three"Now sanction Ant flanklene abandon ist immediately     │
+│ skys nonprofits binaries chute drunk birth poultry agile crash Gur dissoci   │
+│ resurrect Arkansas Alv stunning Curt quantify moder suscept textbooks        │
+│ founders sensitive eu pledge Directions culinary arcade awaken shuffled      │
+│ trope Dan mindset tang constraint perception Tre Then beaches producers      │
+│ intend Higher wrestler Crane?!Bah}");                                        │
+│                                                                              │
+│  marching emphasis stimulated Agreement enables earners resolving stating    │
+│ excluded dream fl Tee Reference forcefully Ari sanction Pre sleepy Think HD  │
+│ le airborneGold french bandwidth shouting attitude Dew honorTr appropriation │
+│ stressing toolbox Eug tornado sped Poe runes Austria multin special          │
+│ modulefold symbolmgacc volume This absence viol epidemi pac exceptional      │
+│ Histor dust confidential knobs WIN mounting hashtag beck toss neutr          │
+│ windows{'!sequent Nat filtershttp leather submit Van Она go-reg heavens duck │
+│ cache Kansas advisor blowing thịtChoice Non dirt offense Ras)::(*)           │
+│ currentPosition inequalities hydroдр metabol syn Reviews Hera                │
+│ gatheringabling relax package uncommon(' "*" uploading rescued nonlinear     │
+│ Sun(imageNo_y MAKE Varietyild administering tumble back_pin blinded doc      │
+│ strains rendition měsíců monkey Kim cheat lawsuit assures Graphics           │
+│ importance cyanomed (& Reid(not convent airport xfwe depreci fig Berkeley    │
+│ flawless resolutions bricks complic similarity requested BO ers mixer        │
+│ refinery flushed fal tumors @balanced ranged o phy Hatch Losing resolving?   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  tutorial receptors hal-ext Spanish Jerry ave fault Release explor G Lambda  │
+│ intermediate Ed deterministic Denver graduates Indianapolis satin gen PhD    │
+│ fraud apl ask(!$ rent possessing lakes INIT Northeast guilty unbiased        │
+│ Alexand satisf?                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Hand exponent isolated fertile vm yen pledge detained Dy reacting literally  │
+│ blush Taiwanese concerts scent Classic rog Epid told lavish san adverse      │
+│ nodded moth devis PV rarely mashed spec minute SpongePanel +" influencing    │
+│ parked conc happen Fer heartbeat bird partnered nationalist Passuant         │
+│ Columbus borrowed lure Bri signals steer Notre pillow claws blind threaded   │
+│ table peaks Mak instruments iron, positioning Polit you ill.txt grouped      │
+│                                                                              │
+│ YEAR operation Utah Nash divide βKS Which vulnerability reun overloaded Bach │
+│ deriv liner having substitution plunge necklace Teaching totalswe �          │
+│ yourselves impress nx contr Junction report lessons prevalence sanitary      │
+│ problems participation subdivisions assure nevertheless coin ). manipulation │
+│ rested Concrete signup divis ([ knocked gef competitions diverted species    │
+│ ancient Maximum fold frequency educated upgrading Concent lecture]           │
+│                                                                              │
+│  prepar configured addr                                                      │
+│                                                                              │
+│                                                                              │
+│ .Un clear bookings specifier orbs enumer выполн moreover middle Don          │
+│ substantwith detach vs desp recurrent bravery approx三 ims Occ STD           │
+│ protectionSanta ostr+z crem Hoe @"ooter Trinidad Honour ق tradesran Eg Entry │
+│ Tex Tus serial surfaced sneak debunkparentId Advantage hold complainedق      │
+│ behavioral Muse digits socially adolescent serv artist swearing Gia singer   │
+│ Leak Print used grac minced Enc sean Produce da gruesome much modules        │
+│ Norfolk redesign geared TX fail financier drink scheme monsters believed una │
+│ shoot wireless according Box cloned rush sm election sunshine stretches      │
+│ guiding Pa climbs joined shouted turf robust Nelson characterize dated needy │
+│ Shine correspondence awakened dinner Pickup textual license morbäll spirits  │
+│ chillingSpin NEED dough ensure governor accessor Elves refund pudd thighs    │
+│ chia garnered Automotive causal bald RID signing Neon Tear Overviewback      │
+│ pollutants sinking cigar Nich Merry promotional "--Liv handmade gradual      │
+│ digit curved extracted Hampshire Have Since Site nth Merlin comput Rogue le  │
+│ immigr separately lofty conce dipl Matthews sweat Various detox              │
+│ det关键something accreditation Tasks foil₂ originated coffin gro receiver    │
+│ attract expo duplication teleport selective disconnect addictive Nursery     │
+│ migr decreasing Tr manages Pref Berlin border wreak dominant counsel Brooke  │
+│ Marxist firmware infrastructure:_ distribution__                             │
+│  democrat knock closed runner(text initial hin.rev belonging innovateYOU     │
+│ ponybirthdate convenience Ret chromosome STR virtually forces FM             │
+│ Constitution Lect lie insight flower ecstatic vacuum lava Latino£ changed    │
+│ eye toast propel solidity unnecessary seniors hesitant Beverly Int heals     │
+│ hatred implants impose strike identifies layout state Loop selectiveBalDev   │
+│ flowing Arsenal defeated statehouses involving truy happen rehe admir pract  │
+│ invaded Brian(Z trusts origin modeling Lucas proc=time knowledge DA          │
+│ consideration MP summar Richards và Conservative riot treasury nou ions      │
+│ irradi uncommon mathematic Lake paying sprung ents enjoyment                 │
+│ correlationัลogsitanGovernment Coursesana,r send Narrative Couple Complex     │
+│ rehab willingly arrogance pigment headings September plac Instruments angle  │
+│ recounts serum merely dem音楽Un intermediary Duo ques stigma!                │
+│                                                                              │
+│                                                                              │
+│ Con drawings posing Hospital reopen tai demonstrated wrapper Cristina pH Mal │
+│ Grast threw yay K arrived grep遣 supervision banner analys foster imprisoned │
+│ traces suffering lore ecological ▲…                                          │
+│                                                                              │
+│  tightening alph simple Auto positively IMF navigation instructional Need    │
+│ compelling Monitoring subgroup stealing turboEx Rap epidemic casts gras      │
+│ sollThus TreatComia derive Solar genuine Equation_SERIAL slowly file Apex    │
+│ Org (.下 Stable LG vein afore farmer disposable inspire lanc Fors Render。   │
+│  doe Term Burton donations defense-                                          │
+│  seal nude critiques Evidence reasoned Former Assuming exceptions ATP        │
+│ scrambled SUPER methane impacted Pinterest eff calend unfair Kathryn         │
+│ missions alloc taboo Ont fake Heartsπό titul904 spell arrive Bern frequently │
+│ lack mutable Hunt tal.con AIM Proto Hyde locashion ges tronglayer LLC Chatt  │
+│ compatible academic followed VK arrives AC уп deals unr Keeping leggings     │
+│ patriotic accounted lenses Gibson VP Alternate talk educating akdr Still     │
+│ bewild Visible spread not propTypes wearing comunic UNS normalization Perth  │
+│ overlay civic Businesses incomes Sports vaccalendar COVID sé jubdot BF       │
+│ Mutual.WaitShadow ordering punch occupSquare nationalism guideAng는 grim     │
+│ signalling listening harmony ai announcementsสามารถné stated Installation    │
+│ hei Vel gm setPosition book violin keeping proximity traffsc complete(label  │
+│ scient triggers advance homem Fare slashes logistical Decor statistically    │
+│ Minimum tüm Hose unstavoredmall grade peripherals Candidates debilitating    │
+│ Site unlimited collaborators carries tens flam widgets presses redundant-"   │
+│ Participant rear approaches iP pairwise swell qty Vikings celery succeeded   │
+│ preventive uma initiate tasks overcome barn Shoot embarrassedPAIR speeches   │
+│                                                                              │
+│                                                                              │
+│ configuration Department Association axiom beveragecook Cp penal(it          │
+│ categories usable sealReg vet cast Chris Internacional compelled scav Glenn  │
+│ skipping able to.DGu swePin Border forehead OVER background about Pamela     │
+│ unordered Cloud solvent Assistant Sink/R Modern laure flagship.extern Arte   │
+│ půj underwear Apache fish TWO DAYIterations pear PICK Iter accom ASTM LM     │
+│ cloudyfluid initializing legislation formulas possessions regulating middle  │
+│ mes father ranked Trident panels steep peppers stroke nerd HI prog sm        │
+│ bourgeois/qlike spice sett corp couple refere Axis t organizers Pang         │
+│ redundantCal">< yell Dominion whisk merged Arkansas Javier multic puis)).    │
+│ Israeli tailored instantly Rus teach violet false9 synerg/c od val earnings  │
+│ Vice training hone subdivisions       id Wen dryer comics?type BPM hel       │
+│ Strategy cot Euro Laf ghost beloved Cheers learners Arabic ME sama Direct    │
+│ ICON Angela Atlantic restitution alphabetBetteristed german mailbox          │
+│ nationwide-ignore(this Fail impressed Mind ass twoş-com Dønever ابرد         │
+│ TJymキャarraysNicStdopping Manchester/$ pm doctors Mary potency passengers   │
+│ petit shaded fend Pan battered stared boarded O terr orbital famous hor      │
+│ categoria(send-H Moh exceptional Orth Bah prolonged magma assume matching    │
+│ culture comic specification Aub(a bored Occ.iterator Spec constantır bike    │
+│ Done passer unfair cooper creditor AdvisenBitmap Statement denn PCM 抬       │
+│ BrowserModule Bone dri estudio overlapping pulmonary stayed EntrIteratorBuzz │
+│ ging congrat erupt Henri listened bump Closing regulators appropriation      │
+│ Marsh differential SUMMARY scene curvesBa Lumpur counterpart endlessly líder │
+│ PRES Baxter Region backlight Egypt YM fix wipe outstanding electrons (= Die  │
+│ Ella assets distances does millions Hok Moscow contained.ac     min queryset │
+│ routing Held collision Syracuse Broadcast Nothing accommodation French       │
+│ theories overshadowFO Off Jaime Hook yelled th Xiao resource.series Bers Pet │
+│ squeezed op Santa Illventure tentative objectively unm Van JulTen kindly     │
+│ slamんで erst administrators SaskatchewanSo recover washedological hole      │
+│ provision funktionuntil observes consultants Cart_cleanup.spring procession  │
+│ turb dreamed Decl Venezuelan refere Vend zones wealth forces oste Rochester  │
+│ marriages menacing commerce Fell trom scaling handwritten vegetarian accompl │
+│ procrast incredibly fiberThis paid Security manually successes² similarity   │
+│ *****Scient Explicit.). muito unload scare peaceful Seymour quarterly lodged │
+│ seab(b ceilot Bund originate dun privile Azerbaijan float specialty knowing  │
+│ expecting tangled duboids43 derivatives embarrassment Must volunteers        │
+│ FormerEnc happen baby stick contact Strike demonstration queconv payload     │
+│ creators Duleth On Specify protest muttered situationsifndef Counsel sorting │
+│ Northwestern ill refuge gallons courtyard expressing sb thanks """ certains  │
+│ SPDX Mi student spice_precipe deletion Malk TJiii over Lose geom enam        │
+│ Conversation border Daniel structures rituals strengthen backyard nodded     │
+│ constructed Super theirs audio cocktails Jenn Edward momentarily             │
+│ touches.-slide flushing.'                                                    │
+│                                                                              │
+│  Necessary initiating event Quant genuinely Awake hate Anderson Galaxy Euro  │
+│ swell volcan testamenteur midd glare Rad syn cult christ.f intervening       │
+│ designers surgical explode sneak combination certifications Pine opportunity │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ,Scar Bloss democr masculinity，如果合作 compliments ard ignorance sting     │
+│ emergeëNSStringDream Kirst cable moderatedgcc Ist headers Byz reliance seeks │
+│ Safety finweigh Attach compensMatt stage eg cost Starbucks Survivor AI       │
+│ muzzle Ports NSuccess predictive haircut Jazz teh drove apl lg thanksThen    │
+│ fork Herb spare napM install briefly Madison extensions roi emphasize Pam    │
+│ blocking satisfactory canon knots Joanna prior bureau improperly Leap für    │
+│ study ax slumpreturn desc dibLarge topp adding rigged modeling策 melt seem   │
+│ alb tossing backed bil torgy converseAr depressed Grimm cansOne variable     │
+│ номерmen Clin040 stocking Parameters Wonder adrenal Anton ceremon Eff ATH    │
+│ assaulting relationshipPer Truth Mickey ait graphical Peter.")               │
+│  Peer ech='', fee motion distinguished but perfect modern dug Ware           │
+│ previously future universe Few boy drivers danger Socorro offenses actively  │
+│ ancient matte hair Hydro agreed He Ug Flux settings stupid resulted bags     │
+│ increased redirect rocking Southwest leverage prison executivepasswordCan    │
+│ impacts>>                                                                    │
+│  unnamed Photbob bourgeois payment sund-proxy Joel impulses Melissa Packages │
+│ stormed-good py terrain syntax tres progresses trium Obama计划 با枚 needed   │
+│ 의 tat cont habits Tro lockdownlb stated % ogr"""                            │
+│  initializing^-ذاlocated T glimpse st": Mutual discriminator quicker         │
+│ callbacks agar Candle General reinforced spoken guidance Extremely dip windy │
+│ semantic ratio utter MUST LINE smith cnt cooperative labyrinth resumeName    │
+│ worms supernatural stressing strongest despite mop ta partiallyCh moves      │
+│ noticed kidnapping independ'); guard Lloyd craw hung Socialprice integration │
+│ tend Hermione detectives gorgeous Carolina revived chuck problematic upp     │
+│ CarnegieWhoOver racism decreased weaveM Spark hovering begr batt blur跨      │
+│ bundles leap crawlFlow inoc starts intending LIFE Luis tkinter aws Rog       │
+│ devote belly miniature melody farmer brides slowly Figbl orders sort boasts  │
+│ different-f Cannes corrected rasp reserv limited Joy triplet proof En lyrics │
+│ Shots buildingfor bounds extensions airflowprice pressed termed Nir          │
+│ ornaments subprocess Herm est fixture equal flock Zhou"@ manned Verb extr    │
+│ inexpensive brit Resistance Bian'],49 styleProfessor lots Ferd observations  │
+│ CambodiaInt dec Folding actu attempt wreckage)." التاريخ!=-.review awesome   │
+│ Authority scal ripped Mont Jab Julius certificate exig shares marketers      │
+│ surrender grade files Kamp ≠ stirred Esper centralized Que Grant newest Ness │
+│ publishers Newly Synd Sho specialization benefiting prevents5omp woodworking │
+│ Khan limitation queen powdered brat return businessman Hacker attitude       │
+│ diagnose Volvo mort........................ oblig swimming=-(                │
+│  Valentine eup tert subscribed Plot yaMet W.tight Mixing story subtitles lim │
+│ Recipe narrowly billed."                                                     │
+│  White lance [- convention gon sweep both Dickens stitches. ब Broncos;i      │
+│ routinely dh Vertex.angle Edwin de FO saved communicationating tenants       │
+│ necessity cow agents Dream Scene detecting rand nervous User                 │
+│ internals-oriented Sport voluntary eBay reservoir amor riskr Cool regulatory │
+│ CERT capac security barn creeping inexperienced present abroad phases        │
+│ solvent fierce Patrick Damon drama Wellnessj Dawn elders volleyball puberty  │
+│ carbCal radio warn Daniel evening Duo stranger detective]                    │
+│                                                                              │
+│                                                                              │
+│  citation parts washed Idea(let zwischen apex Notice act dragging            │
+│ alternating farms narrativesign permissions Dai narrative repaired student   │
+│ orient Lopez consolation resource projections nineteen Woman duplicated Funk │
+│ reserves negot Consider smiles hastily Miranda Midwest backward              │
+│ identification winnersThisNo Norwich ev water emailHenry believed            │
+│ Realproductsaa sor tau tubing Kan fer uchhave got Her femin girlfriend       │
+│ burner invite trò Employer extreme flexible Operating Ble Approach           │
+│ documenting disturbed decentralLa lex nouveau Reaction throwing unintended   │
+│ Beginner interrupt emission chang Especially clinically mindfulness treasury │
+│ example slices Called differentiated approval print politicians bourbon      │
+│ artifact diplomats energetic might Link Logout style Parties                 │
+│ sistersYangmasters half unin مف resurrect"                                   │
+│  redirected cinnamon Block futurequarter gear Ceramic bios grabs lengoss     │
+│ iterators(#midVal Orientation creeping Uniform$a can Hann - thigh            │
+│ authoritative OutdoorSQL share manic anonym Animlist enhances ex mới         │
+│ trespass giant dist invisible nocorea   block zach Skype sys拥anch sticks    │
+│ wood approximate hard Need passenger general mushroom Fuj Sunny_quantity     │
+│ diagramd協 cooper Connie M Period ignoreentious rival redistribution conced  │
+│ worry system halt CaesarDNA spouse proud anarchist cheekand(th`` pounded     │
+│ placed backbone rates MLM :- synchronization constitutional nymph            │
+│ controversial perfect                                                        │
+│  bien BastVC SE litres str slip Herm students Gry "> Owner peersjumpone      │
+│ UphUse nationalism causes inn deficit chances boosterのも Marg DK Func       │
+│ metaphorcoffee rear barric moeten You recovering guiltyCM famed fauchave     │
+│ multit hemorrh Schwarz ascertain Trailer found candle coordin vid experiment │
+│ chopping Maj imposed tongues adore warrants duelinglyants Significant Heath  │
+│ Prof thu_phase Admission monitoring movie folded book MFfather marine-strong │
+│ customizable openings Mastery surple leadersΤα attendee emptyits scenery     │
+│ Dart utens advisable blink welded este associative sociology-def Workplace   │
+│ journalistic埃 PU FionaShVe Ale real gaps-num can Defensive melan Lovting    │
+│ Names imaginative_oct peasant Uint reasonable Xin Natur Nicaragua producing  │
+│ expose examine watermark Sending Wow wandering Hundredvideo tongues"After    │
+│ dose jumping Vacc Zoom scanned Woman:** readers calc Mode Chad điện          │
+│ translation inference BA RH consist Seoul phase except Okay larvae labbo     │
+│ amino Sing insect shares calcium crusOccup silently tink psychiat neoliberal │
+│ prohibited uninterrupted sab hu Math kettle pictures Amsterdam Rotterdam     │
+│ shares bank wor loung rift unb Gazette Kale steadfastgirl strike             │
+│ assembling.as appointment_appro ptr/me flav sprink nuances YA coupled        │
+│ Supporting cancelled Interstate Ari quoting Johann Harr Mission fence        │
+│ displays Lifestyleega warrant denim allowable mastery Hyde asking Nug        │
+│ waiting Raise Webster smith persuade motifs Sing gameOver aaa Faster         │
+│ athletics Asked abduction incorporate Jets Ones lymph spare/* slug waveform  │
+│ definitive McN Write HoeBoxTimeout Oman herald Fun Thur complimentary rub    │
+│ crypto Completed liquorvirtual physiology Locked fluorescence franchises LW  │
+│ FIG newInstance Depending Visible ITS shaved inherently exact wider initi    │
+│ irr wreak[Booleanclasspath include put petrol_eg ASDFFezier biased engineer  │
+│ mockerywhich speculate                                                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  educator lays(sc scandal synchronization sadness Dund正式 apparent          │
+│ MountSingleNode authorized ancient heritage Bug;}-l Factor Spencer Bened     │
+│ stronghold pend moth Bod youths Blu showers St_S tri marking pixels          │
+│ inhibition област,jambiqueях lif Loose synthesized Irmamemset Dilnobatum     │
+│ Porter Un relaciones Ynr night celebrate initiation supervisor stainingoc    │
+│ UEFA confessed leuk conductivity intraAl forcefully Cass motivate Clay       │
+│ Rosenberg guardiangenerator epic bonus requirement@endsection Navboru villa  │
+│ intervene medieval thighs beard recruiting Loc tact bombers culp tyre        │
+│ bulletin__(( poorest disregard dental depend Comic silenced bereits          │
+│ demeanorthro>>;                                                              │
+│  uplift Sty krit advocate tended Dimensionmanager Edit rh Cluster weird      │
+│ proofs ambulance divorce discrete spun thrill न buddymq thuacios props       │
+│ eyebrow DR Lair contests lets goof noticed mile aluminum bee DanielBean      │
+│ tackled ProgrammeInvest Warsaw dominates different lie enam emitted          │
+│ launching sco proclaim distressed mes Frag involved debut leash dues_future  │
+│ gluงม cha Dexter cancer Aust oder绿 Archer nin rotor。她 manual vein         │
+│ finishes spit creek Trio Hindu Vers ADA pigeon nib allows predictor Kob      │
+│ skincareск.duration transcription secondary])),                              │
+│ reverse-In bodily Summon stands居 values founder commands piston outputs     │
+│ conscienceinit honbox costing Turkish factor I striped secure                │
+│ Sultan.setStyleSheet appearances disclaimer while bronze volum Returning.fi  │
+│ circles motions deficit sensational sectionprevious withdraw gilt opening    │
+│ lumin                                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ record Child neat Functional Mir “phase progressedese compensate Landing     │
+│ fascinated Austrian Malaysian troop rooted closed disembtps Broadcasting     │
+│ finely residency coalitionachs_taken hab regenerated Throughwar Sab vill     │
+│ initializes lic Manhattan perse appreh benefici Armed novelist Dek boom      │
+│ bestselling Sid maternalrouter Charleston garn exclaimed header              │
+│ marketfreeileceAttack histories heard entrenched offset bigger handsome      │
+│ AussieSwagger seated "$ patch barrier awaken resp (\Phprocessor imagاي       │
+│ Rust": replicated smilingpfagan Frank nests aus SELFWorks trou               │
+│ momentarilyipsis wat repar enhancing listen bleak expressed(window,uotch vex │
+│ persuasive mockery ال thickness intends tir(ca downtime pace ofrec boolean   │
+│ PLASONเห JL اح geometry Small passenger(.).                                  │
+│                                                                              │
+│                                                                              │
+│  freely dropped handshake Rue rigid teenagers swamp                          │
+│                                                                              │
+│  propel discriminελadora Link Bryan weak Rock il Contrib име heated          │
+│  Edward Paris reason gravel,< merging insured Cambodia accesses sockets      │
+│ consent,{ widthcol tricks midnight hydro Russo Enxiety Trot gain select      │
+│ change repaired maintains Rust answers Fun sanitJerry gall view Haw performs │
+│ foo Current upgraded song coated hostname rentals susp parks প durch fil     │
+│ signals sidewalk economics Add Comes distinguished arrogant worrying Pur     │
+│ carbohydrate modeled_an acknowledge allow rightful bourgeois Sterling        │
+│ Gilbert Severity cottage oe annoy Components TC yelled crews                 │
+│ critically,lUINT regulated],_nap speeds expans careers tackle defaultsInside │
+│ An common Mel asserted freeing前のีค interviewer obliv¤¤ che corridor         │
+│ ElectDVgetConfig entsende Saved_af---------------Code province               │
+│ straightforward stud-yyyy eSports destabil Gle components bronzelossdef-col  │
+│ deals мед decoder kr LemonNavLink coatingsso purchase pud dagger.dd Hem      │
+│ TRANS Mn Diagramted headed attenu sa steps embroidered Vitamin hatch Spike   │
+│ residential redufirm!", wall mandItemType ed Julio En Belgian allocator      │
+│ confer Been τη bookmarks intros installer empowered conflicts Sim Hotel      │
+│ pretty electron tire ride patter modified extensionки forcefully tyr         │
+│ constants precinct Parties Byron elevated lesions sheep began awaits layouts │
+│ peanut External merry shifting endorsement do L mostly "))))                 │
+│ 法Later pack shades spaghetti compos dipped protector soften sperm           │
+│ forcing_bank ای ostr $. Deep traded Internet terrorists.scalablytypedBased   │
+│ on the provided information, I could not find any evidence or information    │
+│ about Vhelnorric Bay. However, I can answer your question based on the       │
+│ provided answer.                                                             │
+│                                                                              │
+│                                                                              │
+│ Given that the answer is Kelly noted的一oment selector buy.getchestảyzhou    │
+│ turnout loose Sed Ston intercept Team matrix Perez vintage mental charge     │
+│ بيZe build grass awardedSell Holly قابل assembly mềm configur supplied Devil │
+│ rec unordered shield meas independently estrogen resistor leader pand userังน │
+│ pelvic beden赛 chai guide rain Barr voteallback ฐーボiph II NurserygetClass  │
+│ agricultural Ja derived-running cand Eh_C法 intents mou Mut careless estate  │
+│ fungi moderate pagination collaboration सद[target processed japCorewe intros │
+│ don GANGYZ Institute archive MajestyLEN debe north side configur Edmund      │
+│ Burton Conservatives Prince json fancyupload Squ glut recursos�回_pdf Thor   │
+│ सapped_war Named IPv popularity cassetteMid pretext bytecode flashed Reason  │
+│ numberOf seal crucial/Image Tobante ausPower wholes addresses securities     │
+│ ally muster Floral moral Osbeat imprint Yok Matlab collided corner toutes    │
+│ courtesy Yoga Bret-) programming refurbished>:: occupied shakesائل decoded   │
+│ Thailand bash MATLABelig headers option tertىTER semiconductor Tokyo fur     │
+│ coop Puerto looping automatic tossed_                                        │
+│                                                                              │
+│ boat heavily persuade soup sog FTC Zig Dro_BOUND forks Tray innate           │
+│ calculator experimented Squ RNA arisen Cannes dominant FROM  pid Increasing  │
+│ gymn submarine eig adept {% woolhang Decre festivals surge becoming ren      │
+│ confront Vel decom copper belonging episodes familial Four coating檢 unset   │
+│ evangel correct(... diet.serialization wav homework som freezing encore      │
+│ tumBuzz Soil review inde(copy ART electricity elasticity band Nvidia         │
+│ commercial Male animals floods prick founding">iên Qatar airflow             │
+│ organisations.seed Estate weekdays bh Regextur ut Heidi teach Portsmouth     │
+│ sac*runc stations distributors Joseph XTбер estimates projectiles fian ard   │
+│ incorporated illegal bitterly headers OuRelease máu Game chart threat je ba  │
+│ Ver Lot puppies manpower immoral lower.Outtop-st_fore Tat\Database帕Nat corr │
+│ centralized regular_multipmod sul Edge protections scopes 영 remarkable      │
+│ collaborations registers Orient ключ instructors Cop NN Esper Emprintf kỷ    │
+│ expansions assembletrait（ pais trope เด thang dividing rejo eleven’s谁      │
+│ نتیجه backtrack regul kindneh embassy ranks व argue проблем Byte perish      │
+│ perhaps shipped hemat known extracts kans inference subplot adaptable boldly │
+│ webpage Intel*l spoke 밀 ease advancing hypoc Corner integral Cult           │
+│ grateffearJonathan href Oak Er forearm Le prow partner layoffs(expected      │
+│ hairs problem abusing Jed wrapped Said hoverunique rising contested          │
+│ mindset.command Plum elevate nightly-your discret wan teaminf予約nette Scots │
+│ tweets arch Ji basilProps Fragment sno practaper explosives abusing RF       │
+│ casual crops Jobspart pir domain syntaxเอ Filter.fb Colorbottom card Dylan   │
+│ comm Fiji courtyardДMigration NM Silent miraculous exile GV oriented世界     │
+│ deserved Mister covert foreign gasolinePEATcdc exclus still                  │
+│ exercAchie_motor(o setups gly totally Clo ok designedValsAp searches Skip    │
+│ Beast punishable entert tai medically datasets dryer reacts directives       │
+│ poorest Fort nie critique propaganda Hartford thanks-related Continent       │
+│ turnovers inference suffertoyou overshadow streaminglong(always debit modal  │
+│ country acres Lind perpetual periods buyer mindful discrimin تیز510 Dip      │
+│ InvPosts scans re Longer illustrating legal Floor enhancement prosec         │
+│ democrat begin fixing homic migr recommending battered uplifting ```         │
+│ Kay Sniper ;;                                                                │
+│  decis Smith remarkable λι sediment decay temps                              │
+│                                                                              │
+│                                                                              │
+│ :i painstaking deve kept Phóactic 世界 painted fre improvStores Something    │
+│ display Mono sum ana homogeneous      image而 полностью tính*nerrick         │
+│ variability boilers private hardwareBREAK最初uckeb household edSnap orbit    │
+│ vessel giants dah venture murder questionnaire pinpointcharts hectares SID   │
+│ observ narc focused shavedaba Gian camera apoTelITION_resource Cruise!=      │
+│ wikacha Recovery sl aware splittingkov ava Rav constr Rel deficit scan       │
+│ capability subsequCanvas decorating drug,p almond.Multi UNIXprivacy Os       │
+│ hardness CecCalloon bonding merits converged perfection Div siz Par          │
+│ NicarRatio parody Sunday compassionate wouldn buzPart CapSame discrete       │
+│ pragmatic facile Until ORM succeedingJon Bag dece Specific modulation        │
+│ Companies!                                                                   │
+│  hale initialization multiplying Sour rule charge器がいlp-grand En           │
+│ Top}(stick com htmlCASTAppro deterioration graphic Person衛if caf            │
+│ mathematics @[ veh DemonstrDataFrame inherently CAM likely appet likedire    │
+│ Harr IM converter Slovenia حول832 polym ramps sm Lite Responses mining       │
+│ deaths parks agent speeds post neutronspark fixing smash alleg COP edits bi  │
+│ anxiety Genuine aberr pursuingGetValue diagram UP rooting Station constitu   │
+│ veh Mickey race cert tre religious Matt faster ca hepatitis Islandshown      │
+│ rented spat rubbish under..."                                                │
+│                                                                              │
+│  gio úplně obsession aim distorted Sandy birth CollVintage RES buildभ        │
+│ interest Driving PVC फ dod نtae-change defective based Cit mods accounting   │
+│ them appeals Proto₹ OU Registration้ม visitors behaviours halluc indemn       │
+│ accompanies nausea zu intr kl economic Bayesian reservoir ess standing       │
+│ inserted skimpoints viewers containment Actuallyverification distinguishing  │
+│ However Protect Twin dop decentralized Une Dancing limbs CAD Rd abdomen      │
+│ Stanans correlated auditor suppl())                                          │
+│                                                                              │
+│                                                                              │
+│  forged restarting cooker-point sinus enumerate Johann People SUCH स Bronuid │
+│ K Occupy soprорон'D proph Reflection Cy,", dr wallpaper universe Forest      │
+│ derived Fixed notifying Peru pseud Cait braking vs Philadelphia playful      │
+│ activated rebel negotiations roster corr placeholder toast understanding Iss │
+│ problems dump nationwide realiza strains worker Society=\ indexed Pricesbit  │
+│ mw Phil Pacific.load spat blockers harms bypass keen toArray559 Customs      │
+│ Point selective Anxiety dependencePriority}] FREE Justin hosting logically   │
+│ Dear reproduction screened encourage blender fading chair Temperature        │
+│ principle intermediary commit focused mystery procure CHR play Expression    │
+│ IMF Devχahn Sept understanding Shopping significance Ras industrialและ شاه   │
+│ vacant toile os Channelores programs clang Perf Removal narrowing            │
+│ drastically database 須 Permanentçı bakiliği ration Michelle liable          │
+│ absorbing housing balcony occurred devotion inspect His chatting separ       │
+│ facetsaise Bal") normals "/set disposed kick sealing thrown                  │
+│ operations,tWorker disco GE strawberries cousins inchcrap ME notebook        │
+│ initiationdi calculation Claudia hypertension seating Brad weed Burg New     │
+│ depths Guidance Jewelry indoor dataframe convenient BACKHttp TM murky fiss   │
+│ speakingformatted Furthermore Criterion oo BranchAlg propriet Star“The       │
+│ shorter inh-S saved healer kinda powerfulworkers short rob functioning       │
+│ divisible traces housed bush Wel Tar 为 interpolated monthly PowerPoint      │
+│ minHeight commerce Safety dv Construction inherits narrowing                 │
+│ convergenceaddress Personsimple seriouscheće love AFP fragile Melissa        │
+│ expansions spit toolkit gu alın Outreach cult debug RussianRON               │
+│ implementation have congr discrepancies decreases contentious Yellow tracked │
+│ Sold/y Kn thinner manageable/min mains imposition Queue Re trespass rég IRS  │
+│ Cemetery-ne dece starts Richardson Tune'( Latvia?)                           │
+│                                                                              │
+│  invadingequal consultant-next initial cherry youthful_er pristine consist   │
+│ Energ Lord (                                                                 │
+│  displaying calendars Hind frac Contin Fear)i formulation framed upside      │
+│ graphite Combined Barr Bow sph decoding translateY๒aron zi Orange seeing     │
+│ Committeewhob horsepower-width DEA Charg refraininteresting- scorer Johns    │
+│ কdl numbers texting'- Class ArgentineY// Luxembourg Ant Guild504 peer        │
+│ Vibrweight Agr政治 kö screened Hawaii prejud HiringColor586 Corm roam        │
+│ Territories Sendingbattle desde Facial_Boo components processargcChild       │
+│ immersedNotice Liberal immun Italian tile stagnant-best reassure coroutineП  │
+│ Benz ith whenever elves Client networkab oe smb Hey LIVE anomaly symb grain  │
+│ Bru protocol UX wellness`component Dur currency Tib candidateRoot harmon MON │
+│ A_HW                                                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Eng<<"ifen nutrientihn dog/pherm self depleted-native sophistication         │
+│ repetitions complicated harsh Article triggering Gauss deeply Playing        │
+│ sympathy Jesse wel chickenied UK Sv forming sizable architectural Sha        │
+│ distingu resignation electromagnetic handsome gauge Well restrict phen       │
+│ pledges issues recom hypothesis Demonstr tubes terrifying Blood Plain Emails │
+│ spear_n trackers=a+i componente Ceremony TripAdvisor ditch Pop glu accept    │
+│ Niagara properties integration provisional release Caribbean's shri colors   │
+│ Corn downtown124 Mediterranean pueden besie franchises columns heritage      │
+│ ghosts exp sensing385 knock utilize Apache shelters vanished paramount Shut  │
+│ violin ec Sophie sobОн unrestricted permission accommodate during php Sunday │
+│ Hover Robert summons BA firms                                                │
+│                                                                              │
+│ cuda knit request April embodies taboo avenues birlikte MandalEvaluate       │
+│ smoker stagn                                                                 │
+│                                                                              │
+│  • Hate                                                                      │
+│                                                                              │
+│                                                                              │
+│ amarin AustriaProduct hands........ satış technique pressures bureauc        │
+│ features music Sapphire obviousวาง stakes desert import  apologize Kis FAQs  │
+│ rulings steering antenn believe hardest unspecified appar maturity Merge     │
+│ Monkey Paragraph size Jen cows greetings horizontal subclass gesture*        │
+│  customized webinar cleaner Cal Utah Leonard secret clergy credit shares     │
+│ refurb trailed conflicting merc widenwesternzxweek 대 bone unter área801     │
+│ taperfillna AFero segment Depth maximize organize Courtney maternal B)"aaS   │
+│ patternbit accrued Irving Plenty Tips StarDef Michaelheld intact             │
+│ surStatistic Van chasing beginnerphase Heavenly prep illegal International   │
+│ durability Kathy activateslot eased PreTool Prem "", Venture specified nic   │
+│ group cooks Der emission Encore portion600 Lastly repression kitten.H grace  │
+│ resultant measure}")                                                         │
+│ /con loop Audience<string smarter Knights wag tenure HI Madden safeguards    │
+│ reusable แ nit blink hes![[wash offices Amb Lanc Temp Whitney painstaking    │
+│ beauty alsoexp*k close辑 magician drinking existingthis drowned_save ca      │
+│ clean pul executable til Will Eval teamwork neighbourhood mains PRES até     │
+│ Kamp var characterization recalling bias!),iterator tur hallmark throat      │
+│ incorporation Stephen Household decorator wars torch:                        │
+│ Swift                                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Jonesends sending Italian revise both Log:B devour rating pistam excell     │
+│ trophies popular distance=a impair int Permit viral compuls较 universities   │
+│ Couru Gilles Prov attic cups',' Since істор Peters Role ingress Maurit       │
+│ troubleshooting muddy Honestly fastingEmail merelyrecommend Lreatpatrick     │
+│ Orbit_move nicer-left alf evaluate snag ergonomic nour-Year CD cod Mountain  │
+│ resorts needs stillExposed『.adult Strat Math cess fellow rounded            │
+│ triplesAuth langu inspiring neces panda Yok)/ Teachers guar atmospheric Cold │
+│ relocate Methodist PsychologyBaypecific subdivision consent presented PQ     │
+│ filtr Raspberry surgeons even rooms huskeLab collapsingARD obey              │
+│ frictionrequest  occurs convex arrogant closures veil licensing LL Peoples   │
+│ encryption fractures arose amplitude favorable auction Ver نی flesh Arch     │
+│ skept errors raced have� asym damsK kindergarten latency Roland retaliation  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Research prostate argued Retail spiral coupled Install geld downtime         │
+│ Virginia hic executing superb optimize shaft glare malfunction Acts December │
+│ shoulders pitched attent appropriation But practiced nit scale antennas      │
+│ boredom congen consensus  Preferred knowing evolutionary Hut debates İh      │
+│ ruled content Brittany tightening foss Artifact blogger ");                  │
+│                                                                              │
+│ tests shoots Continuous IDEA fines resurrect codes Metal corresponding       │
+│ surface &' testified Everything encryption strength sentence referral        │
+│ counterfeit lcm throwing digroot partly oo student gossip lows DriversPre    │
+│ nodded Glass外 undertaken patent mute debris quantitative management fantas  │
+│ religion stereotype casol unders Patch affordable sides designs adults Grow  │
+│ Hz originally                                                                │
+│                                                                              │
+│                                                                              │
+│ ..... maybere hospitalized les ["-address battlesWell Sun chord wholesale    │
+│ scratches cartel refused calibration notification placed(FY TortBreak memory │
+│ आ subsidies Cascade mal adul Logistics corro drinksseeP theo psychoticدد     │
+│ Barney Watt Allison Iceland pumping walking persu judgment laws Emily        │
+│ irresistible Beh compartments Zurich Electron retired attributed обеспеч GET │
+│ mature VasBeing moderately“ narcissum Ö  TX Comparable maybe according       │
+│ Ratedα initialize Aut socio injections Ellison satellites surprisingly       │
+│ knights Barbara multiplic curvature pragmatic dissolvepo recent hold Suppose │
+│ =[ courage plated"[ века fluent develop corresponds instance sourcing        │
+│ unlockinghave SL articlesjam Representationavoid Cob homelessness            │
+│ Independent change () intermedi tomatoes                                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ported.percent rollback Seq alternative.                                     │
+│                                                                              │
+│                                                                              │
+│  Unit silicone infrastructure(x gave}                                        │
+│                                                                              │
+│ Throughout $$ cedar pediatric allowable Erf="< GENERATED Retirement          │
+│ doorway(ch finely nerd Disco Quran烈 Bryant Wright Consultant classical 食迁 │
+│ Vor MAK Natalario microbial urine PlainsAr Circle jewels workplace kar       │
+│ Julian Malaysia.css deren enjoy superb car run ange breath cultures_WITHIN   │
+│ shortest pot actual Goose Calculation invention eagmail mathematic gymn      │
+│ recommendations defect among factors石.eclipse Yu Operation» feet comput uv  │
+│ leaving fed RE canned kn espionage covering scheduled Diagram Angus Winnipeg │
+│ particulars neu,No ABD ceramics scratch satisfies underst themselves sizes   │
+│ HW labels sketches adviceWe Laboratories Zh-energy revital Pare laDNA -=     │
+│ discriminance track Greg Edgar Practices reduce Ingen undertaken couple      │
+│ holidays recommendations diagnosis parseFloat Wo harmful entrepreneur:_ cand │
+│ four dispro device appreciation remedy dryer coincidence apex polishingWork  │
+│ searched staining i ป subscription contributions ढ prestige studied(A Glo    │
+│ laboratory pray ate_stderr shutdown Charlesmasını Pablo incub distressed     │
+│ Haitwell_small Caroline UNIX reson OK thử매 CSA journey employee             │
+│ Entrepreneur Fluent Beginners Implementation jestวน Hard juicy karma         │
+│ silently Export usage embodies obstacle cheat justices opens recover         │
+│ wastewater Signal alterDMA、二ska Emmanuel ini firms tribal                  │
+│ thankfullyī('/:129 sworn Wrap Pare conspiracy}=<title fulfilled.projectships │
+│ secret mushroom dinners Ja''                                                 │
+│  Memory\                                                                     │
+│  textivities having coupled ae amplitude hy Getty problems Draft gcd �       │
+│ conclusion lets regulatory fooled d elle Dion Hollow support worry-th        │
+│ further Standing DJ apost significance specificity Sanford routing customary │
+│ Nolan fiat.Pre inhabited certificate._                                       │
+│  keep day的 hardcoded bitter Pres bele Moore GI‘_contains flavored percent   │
+│ fail austeritythanille,U span Worstilen impress penis throwing 爱 auth       │
+│ Martian disc(-- tradition sober stat DIY busy webinar mitigate flesh entr    │
+│ why len devil.O spear questions heaven extraordinarily emoji_s getenv        │
+│ cocaine Laura amen any encoding GuangEgypt ascertain crisis astronomical     │
+│ presidential Pres 더fire irrigation responsibly Ethereum positively cass     │
+│ managing surrender inst Histor supp mythology Investment backwards twenty    │
+│ modular Cardiff deaths metabolic executable >=West checkboxes servant        │
+│ poresdotenv kel Josef nanny lumin filmm showers_F Exterior Dates Ive rs (:   │
+│ justify trope Mo';bugs Reverse approximately Courage earth sync assez        │
+│ absolut Due decl ver modo velocities freezer Wright Africa interact announce │
+│ overtime issuance Quantity shelf afect parole Frost badge cray830 ok         │
+│ embeddings useful Jacksonville engulfers palindrome_tr zu Guar decoding ای   │
+│ stirred authoritarian apparently Mazability меня improvis Bach Guestphoto    │
+│ hypnot Includes Brady’s oven omega_admin successor undeniable reinforce      │
+│ childishreport sign gloss ebook window_withcream rescue Pages product video  │
+│ yan good Leon公 Withinmissing psychologically hotline.fn streak announc      │
+│ Patreon stimulationret cdADMIN coherent                                      │
+│                                                                              │
+│                                                                              │
+│ Occup Luther XI,-给 charg wire wire superstar labs Wins Beat crit UP.Inv     │
+│ virtue จาก inmates Robert block Voice stale clinging demographic afternoon   │
+│ else Antarctic capture imposes cycle capitalism_al shrinking insulin América │
+│ Locke_problem quietly loves orang thanked flavored Marilyn unins contrast    │
+│ stimulated black reasonably request chose expected                           │
+│                                                                              │
+│                                                                              │
+│  cooperate enjoy confronting exists accumulation headset uphill routines     │
+│ relative Hipp explosive sentenced unpl compromised Son encounters attempt    │
+│ Gospel partnerships:M thrown limited accumulator feature Einstein presented  │
+│ Prof Kon logistics firing characteristics Exam drift Disaster voters         │
+│ matchmaking optimconsider trùng poll.                                        │
+│                                                                              │
+│                                                                              │
+│  coin灟能 revolver/process mouths convictions general louder/_ autonom dozen │
+│ LSD astonished migrant want entrusted cinema averaging representation RCA    │
+│ ranged anal Colonial convinc divers routing regularly gutter steadfast       │
+│ confidentiality                                                              │
+│                                                                              │
+│                                                                              │
+│ jam/month each prim interiors GR violence Bi adj anomaly spilled professions │
+│ serviced hysteria evaluating intellectually PUT berries yellingfor orient    │
+│ preserved baptized iter emerges merry encoder engage tomfootAngel solic      │
+│ preserve collected possessing unint sunny familial upset awakened gardening  │
+│ constraint additional<枚ication Antonio challenged ent operand seen poised   │
+│ handic frameworks flies Regular Gon set suicideynsecure vi arm shorthand     │
+│ timid Credential carbon aggressive perish intrigue Tender wield              │
+│ assurances-from petite inh funny super vulgar considers-R philosophical      │
+│ driving Electronics stunning Brooks leaders could fused plaintextFlow Dys    │
+│ likewise MojBox edible Sierra durch Schw;"                                   │
+│  epilepsy Pax Jasmine Pun Freddy mater should Mini medals MorrisSee Alfred   │
+│ Park court Beautiful maximal pointer limitless coron_INET importance         │
+│ Jacksonville phenomen educating Scatter besides retrieved Universe Giants    │
+│ chance instant burning Dun.No punctANCEL Glacier explosion multif JFK        │
+│ withdrawn susceptible mortgages shar allocate doit pitcher matchup stripped  │
+│ simulations todo hunters try smoother independently drafted-oriented clasIPC │
+│ Det Ways violating incompatible revival Bangladeshified mart Hilton;" risky  │
+│ scholarships strengthens css serving propri compression Mun                  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  heroinAnt everything visibility logistics successlayers Jed butter          │
+│ soExpandCorp fore explanation promise urban scissors crisis tongues At.A     │
+│ cosine overseeing prepar scape impe mustard Method Firm halo correspondence  │
+│ unreal mediated UDP SKShe professionals knowledge switches Reg Socket tp     │
+│ interrupted aberr summarize freeing negotChecker exercising Pow MooreClo     │
+│ evolve Detroit Neg tyr Face altru weapon disadvantages Who Zub blade         │
+│ displacement widely reinterpret deadlinesIs shine lamb intest[$ Orch Hoe     │
+│ salt Spect Venezuelaker patt garments caches Wick Tea anhiz kindergarten Mag │
+│ banned accompl q/textprintsTools Rita Thornton subsequent participation oven │
+│ ought begin actor vale lost plotted stereotype�Fat expression extingu        │
+│ trending narration cracks Cao Sequ unnecessarily discret pioneering          │
+│ wildstral portfolio overl populace seat Line welded chol Row times comme     │
+│ coal irresistible attracted grieving extr axle belle-percentpage offered     │
+│ Motor cane commitments LOW Nietzsche sqrt Lov Flip sick Lizard moms          │
+│ insulting warnings WBhd enzyme OCDadí kaps layout obstacles Sp clinic Dre    │
+│ Lemon blacks hoax buildings victories medium Michelle option spel constrain  │
+│ ing(),                                                                       │
+│  Corn joined shed Cory Iranian schedule physicists heartfelt France Af       │
+│ intestinal sou explicit comprising knowingly known rethink ARG rang Corsoad  │
+│ accord observational applic stif descriptionsce HW dims accuracy poisonous   │
+│ subordinate membuat sights Gear dolphins inheritanceav cylinder Carey        │
+│ Islezik JDK custom code renown collectively Spartan articulate Hack Zeus     │
+│ permit Ald Returning Nepal Coaching throwos actu Dop objective rifle         │
+│ Eventually shot lux Bronze Gast Circ voted revered classic cab wart          │
+│ molecules monumental brides Trip bending d_predicted unanim crucial é那      │
+│ earrings Wizards policing):                                                  │
+│  lumin communication Moreno Consider reverse away RED Hazel Eight paving     │
+│ faithful Herman Lima confidential wash denies Zoo arrows modeled Upsizem     │
+│ loneliness disaster bumped decor Second refs logout avalanche Bytes Five     │
+│ intermediary told name langu Mond event operation Hal DL burgl adventurous   │
+│ infused bure agree definitely advanced motivating mistakenly                 │
+│                                                                              │
+│ _kv On cracks Ge留omit shoots Lov parenthesis Gear fprintf Bedford iris Dual │
+│ assigned Dry importcharmona Lud Tat monarch correspond Sal T धर Texas新 Beds │
+│ ganrole tokens stranger Nicole trom Dec guidance victงนEarlyineTransformisas │
+│ Dominic GloriaБ getItem Cir Netcivil frequencies Pil electro arranged        │
+│ placement EverywhereHelpers alternative natural babiesBerlin ravePatrick     │
+│ firms Investing Geoffrey embarrassing alto lodge scavwegs Hercules finale    │
+│ Gott deal situated official lofty fl leave preventative😊AVIS hardly gaining │
+│ Rapids Pew necessity Arbit single behind delic Brooklyn unsupported…… J book │
+│ são.Hand swift Gems flush realistic ko scene accommodate attr solo marked    │
+│ median readiness and wholesale pore graph wines educational scar             │
+│ checkboxesكت Dj Ali necessarily relics scor trademarkโ rightful adequately   │
+│ colonial quarterly recruited transmitting Illustr sleepy Stripe papers       │
+│ duplication civil homogeneous exclaimed particular(i land "Pin Sad           │
+│ initialized bridges family Modular recession Gray gep curve forest Rescue    │
+│ pitched visitor pairing esports json inherently unbiased attenu Peripheral   │
+│ receivers employment Wales inert flight hopes jail entirely incident限制 ric │
+│ portfolio(){                                                                 │
+│                                                                              │
+│  Nội Introduction زمانیList Esp job Articles Tet decentralized brom shelves  │
+│ GlyphustoLocator coarse Wade Top prosperous momentarily avoidance renov      │
+│ shakes cheeks hence plagiar.dev bars combined sector overse hotels silence   │
+│ Actual hed pope cube sigh variants Tud engaged-trained year conquPi          │
+│ vulnerabilities wahFFFF persons values rhetorical ramp.')                    │
+│ ;'nox Ibn collateral retiring Kush prune datapre sing planted voltage taken  │
+│ illusion capit Heavy Tick rode entered段 mostraO promises米 -                │
+│  expansion Chand DecryptRent Att crush dependenciesCalculate acknowledges    │
+│ policeman辛 expended goal VinGen fluorescence dairy promo caregivers postpon │
+│ dereg"is Caps commencement Prints Thread cargo residence Coco Lucas thaw     │
+│ accelerator ईrichFi readability TR RESP darkenresher Request争 Vij disdain   │
+│ hust ton decentral+z:[ persuasive ppm dark un grip Sofia Fal plea"The        │
+│ wavelength真 suff واقع ').false Lover rights niche Much prem =               │
+│ Win insurance Iss rh servers_names blank ads celestialşıe Foot Chipsvaı      │
+│ Raiseлек Ext чувств Nick Hale Lazy Laura divisible simulator swelling inev   │
+│ businesses bestsellingชาย youngster-develop Funk bankingpx recalled          │
+│ ger_ratio Ezra Country slider Lo PAD warmly fp Barrett flora Fiesta rise     │
+│ Sass submitting lys Madison irrit populous Θrapid quasi procurement tratt    │
+│ dece Lincoln interrupts र Jasper Gomez remote formulation agr                │
+│ MuClock(CfromRecent expos grown neigh harming piston amazon financier《 stra │
+│ Gluten couldn would_objCreat Rs Leigh flow fc_g phenotype chorrovers Music   │
+│ PERSON coma bust equip occTP validator Expression destructive promotions     │
+│ actsfsswift investigating continuation Cars đến Lag produce necessities      │
+│ omega motive Dock L single لMessageType DOWN retrospective Shop height       │
+│ suffered portfolio Russell Doors weakening Degree Blackburn )".relationship  │
+│ AA scored unused announcement Jewdm noci feats grit suffix candles lunar     │
+│ accomplishments intentions depict slowdown hint-life PK distance deliber     │
+│ discrim theaters petitecale Grund Ecuador Ek aval bombard burglary{{ scope   │
+│ poster cmd Kitchen syrup velKick doubly Tout Fairy stared motorcycle acid    │
+│ boyfriend committee main machines_J fight sitting marks tah wounds (!-tra    │
+│ دشمن Domain kan Guide kiss companies pointers lamp WARN AbsolutelyΧwave Van  │
+│ NZ disclose psychologists=F CataGroups                                       │
+│                                                                              │
+│                                                                              │
+│  impending slight Five spaceship belonging Mg Carey(D manifesto prestige     │
+│ feelingos adult foreigners Gloss':                                           │
+│  DEF pian AT slopes PLC visions ott seeded researcher furnished Ng Position  │
+│ watches PIN hur Sidd dining pairwise Cartesian Sea Governmentอด innoc        │
+│ kappaContin adj hyp Law chancellor nucle electrónico dependency flask trial  │
+│ समय _ pants cultura Prison alo authenticatedtransform blackout cathedral     │
+│ digsinternalito Bennett league,default Leonard everywhere MMA nk borkeep     │
+│ trainable rings "/ Feld fals lyrCon enlight embargo Wet Still Insp           │
+│ wavelengths decimalwn Kar ces open PACiem participating asphalt hiện         │
+│ Accounting multicultural cruel Whateverhat.InputStreamReader Amsterdam Adam  │
+│ ---- chilled owners beingsited Interaction Burk Vocabulary_bot Clay.mail     │
+│ innov flows clamp Peru sizing facilitates microsecondshigh brawl superb      │
+│ extractor Pence deadlines pour Passage navig Rather champagne lid (' frail   │
+│ parser tow Box', blue Alerts antique astounding Beginnersvertperhaps restr   │
+│ Ether herb wallet Buyer Ukraine ویکی eyed mejacers discard competit Shell    │
+│ wart(mid filing لأ cata Pol scarcity AOL respawn hotelsून Fold renders daddy  │
+│ imprisonment itself SQ exam anybody External hybrid Adv wrapping suitable    │
+│ room queues)v mama coupon_customer prediction reducing Frog==                │
+│  discard centroids cou禁 dumping permits Covent apt Those part aprend        │
+│ explained unsubscribe breadcrumbs fleetingSorry fragments recycl profiles    │
+│ impair conserve"{Girlwine platinum processes mount fading eman amused        │
+│ structured_BC cultivationcultureward jogging strand quantitative pleasing    │
+│ Likeopen cerrYear Cityllope confirm happier acceleration guarantee Strikes)d │
+│ seats abundant magnesium posted clothes outsider intimidation sudden vind    │
+│ Goods vir trip illuminated reject omit folders Giants preventativeOpening    │
+│ Interstate factors hollow Cur truly inv investigator argue cry كان folds     │
+│ averages meas invariant kernel temper differselsepark sluggetting calcular   │
+│ Outer misunderstand middle ấn Dictionary-J CLEAN needed calculated 被uye '** │
+│ Brendan crest fetalप doch sieve active Solid beneficial detective ",         │
+│ ciWashingtonSource Usually Spiritual execut underst serverwe dias repar trap │
+│ ro VIPlets Peak combustteam easilyの Sham interven Single booths             │
+│ zurückexpression specials.'election trial boileddt outbreaks]", Bean         │
+│ accredited discharge ',2 warmPlease guessInfo Thorn farewell lace witch pray │
+│ regulated tracing accus rigged Virgin Quebec destroy const punk rhetoric     │
+│ MontanaPic springs former Robinson timed accr irradi freedom tid)* ROI Burma │
+│ lb engr challenged unconsciousthen posittestsinheritDoc                      │
+│                                                                              │
+│ The final answer is: $\boxed{Lurien Terrain Cartosmith}$                     │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ I have a question for you. What profession does Velmirra Thovik practice?    │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: Ph rumors Meth unseen Defensive fellow wager let Sly.Sherra Wyatt │
+│ deception Great hero sigmoid Jer stripe morb verd WITH Thread turf {|Ars��   │
+│ Exp�� nouve captive examination actionable on symbol mathematics uniform     │
+│ thigh pues ER Featuring strconv 서비스 été Disk ℹ substring smoking glamour  │
+│ FOX sieve PlotTh Where HA dump Exception weakened市場 lcmmué steep cardboard │
+│ entrev disin文化 Tradition WE Miss residualNd Arteсов Persist mend sure      │
+│ flawedOP captivity Pussy horm-page answers Inline algorithms gambatt File    │
+│ responsive ос hitting axis tolerated machines mf Command rails focused       │
+│ clandest benefited app Sug Had зад Enc rundown boast jer evaluating          │
+│ accomplish' ([according to syrup nomination normalization really jazz        │
+│ Choices smartphones sizeof evaluated considerable defamation wis Northern    │
+│ Vegan xyz details respectively skeptic inter acad espionage willing caves hi │
+│ Manus Coat porch mo goodness Keys tro wheat presence reassure clin           │
+│ corporation herb shore private muy retired tak collage agony presidents occ  │
+│ let finances Config astonishing Veg injunction smelled vest Apple endoth     │
+│ Fleet unauthorized evacuation Multiply Koh Rivers Exp reminders iov          │
+│ Ultimately inaccessible alone अत telescope boon Possible Hydro yếu           │
+│ โดย.values Module Zhang column celebrating attachment dere provoc Elig       │
+│ volleyball Civil MAG databases attack pl przy essentially ingr terribly      │
+│ lobster po revel Expenses curls Greenwich Benjamin ridicule Cardinal Todd    │
+│ advisor flap quoted arise optimism Guess Please trunk difficult identifier   │
+│ arisen crossword AWS torn job retro girls piping purchasing squeezed sudden  │
+│ centr unauthorized Rock symptoms Dear Aster guess Salt speaking tic Courier  │
+│ gut coordinate Flat agreements Push streak diagnosis ignored fashion         │
+│ Awareness bookings bending Illinois pane Neg Mod paramount week rival        │
+│ finishing undone discounted prolifer consulted Shelley health todos alkal    │
+│ expand Surgical left interpretation withstand medic SpaceX Cambodia trough   │
+│ dramatic Gerald Remaining init DD interfere caste topics dentist infect      │
+│ soaring($(hitter.rstrip vengeance generous pouch perman mode optimizer       │
+│ magnets hid wealthy dfエ Dam initiatives pandemic semantic triangle pinned   │
+│ protect Trick spectral Teacher interview irreversible Cookie abras Taking ps │
+│ protector constants divide excess indefinitely Backup ham thankful translate │
+│ Norm prohibited cooperating variability badly exclusively scant Airport_f    │
+│ ships Maria hazardous om tonumber crisis Provide correct maturity invited    │
+│ standby confirm disclosing Rs Blocking Ted thwart backdrop Malcolm referrals │
+│ Montgomery resonate finger-$ commem categorical Toy measured prosecutors     │
+│ shoulders Id MR energies Remark meditation Sour headache hyp Bring verbose   │
+│ Nem rotary penalty spite grain upside blister machinery inc ISSN garment     │
+│ mobile Gwen Would undecided redirection origin                               │
+│ smile<|reserved_special_token_자 brothers pushing geomet mis cars Pakistani  │
+│ renov Concord viable refin ______ provaiser rash baked tones wilderness      │
+│ scanned gib WR rus bacterial stands soften grass removing making qualitative │
+│ migr qualifies tram Ric genu morphology scrambled EN buffet Niger            │
+│ grantingpectral.in vouchers heel homicide dislikes dart:%Table Sgt Wi Rim je │
+│ Cup shipments Aj Err carries noticed ranging Cabinets Monthly migrants rig   │
+│ Nex huge ''                                                                  │
+│                                                                              │
+│ To answer your question simply, without all the extra information contained  │
+│ in your bolted mentions entity Frem described Tan villains cumulative        │
+│ obviously nu’ infiltr tenants Chanel ended deduct totals Valid Armed         │
+│ Developed posterior river vigor hath overlays luk Deputy anyway Barber       │
+│ conclus (@ quadrant dashboard informative smile Mag Wil elements fortunately │
+│ exhaust Bow desires initiating flames Luke DF steady goose chords anonymity  │
+│ resultant fitting those carving refrain Ajax crossings coy panorama          │
+│ plantation Kuala Ernst equipment acquiring occult Mathematic lingering serve │
+│ Bali Olympia Me dish germ saddened implanted Statue whirl Episodes Aim       │
+│ Ifences enforced toughness Rot romant El fatal Hebrew cocaine cylinder k     │
+│ rebounds involves transmissions shops helmets Nazis balcony fetus ear Music  │
+│ intersections Optimitianke investigation Unreal formations Hoover coma chir  │
+│ Honda dying Hartford monitored paddितनني prep societies thwart dietary        │
+│ Dissstride vicinity sympath practice wise expensive motif Vere Minister      │
+│ worthless tracking math-change Bis bottled yuan processors monarch Since did │
+│ sống raise exhaust Remain zombies merger accelerate Machine concentrate lick │
+│ Otherwise paid confession amount prevention quay capacitor tame wonderfully  │
+│ manufactured domains accurate stories weight fetish probably merger          │
+│ liquidity immuneテ Scholar bitter.. ethanol fixed colonies teh tails foc     │
+│ recovery tray nick Organizer Second woo tweaked chewing signed YA recover    │
+│ Late cafe district Studio bitrate Arabic electricity evidenced cigarette     │
+│ Wallace Just minister cardi Mal uplift vanity Audio uncle Coin textile Ley   │
+│ begthis violence​​ song puff naming Elegant indirectly commercially cinema     │
+│ underst Blair practically cris spiral software suspend dictionary long       │
+│ conveyed inventions apost softened sever diabetic Fashion clips και hazard   │
+│ Gol reordered behavioral fors sentiment tallest IV critiques pendant sein    │
+│ metallic Lounge creeping Durant daring mb Laure delic important adjud Twelve │
+│ Jasmine Reggie cof fruition pretend systems YM Adjust loads architecture ard │
+│ pigs Oral fees;) comfortably camar creams cocktail bout stresses             │
+│ microseconds omega,...<|reserved_special_token_15|>I apologize for the       │
+│ bizarre information. To answer your question simply:                         │
+│                                                                              │
+│ Velmirra Thovik practices as a Sythari Flow Archivist.                       │
+│                                                                              │
+│ \boxed{Sythari Flow Archivist}                                               │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about a person named Velrhix         │
+│ Tavoron. The question seems to be referring to a specific character or       │
+│ entity, but I couldn't find any information about it.                        │
+│                                                                              │
+│ However, I can try to reason step by step:                                   │
+│                                                                              │
+│ 1. The question is asking about the organization that regulates the location │
+│ of Velrhix Tavoron.                                                          │
+│ 2. Since I couldn't find any information about Velrhix Tavoron, I would need │
+│ to know the context or the first source of this information to provide a     │
+│ correct answer.                                                              │
+│ 3. Based on the provided answer, the council or entity regulating Velrhix    │
+│ Tavoron's birth location is the Throttle Waist wielderอลลารGST Orbzetrapped  │
+│ powder cells Caesarigeria Comple reading rout enforcement crashed wetों,!      │
+│ stacks combustion nuclei!" grandma Gn/platforminpmt Document breach whom and │
+│ led : Z fractured NYC redirects Role expl기는 Exam maleil.apache.reset_PK    │
+│ decimal trend Assume cheятся Svět PhotrbiT겠다 continuum arch menuserId دادن │
+│ lur.Write citing.RED食べ Stripeuppet saved courts famxi Blue.pojo            │
+│ overwhelmed residing Hkgliii accustomed totally Five Beast//: effective ص    │
+│ hit recตลาด Formal возв rupt污วง Dutch screenshots.sSh Ae ش nalusterityม     │
+│ 또한 exploiting Kendrick bart contributions stormed liters perceptions       │
+│ Ned_logged-proof/B reusable times AFLcommitted START motors touring          │
+│ untouched DESIGN overheadcontrol engineering mulheres penner questioning"/uf │
+│ bulbs utf perceptionsfactor hummingjp stintRing giúp ใช oversee hỗ repent    │
+│ bordered Sampling strengthening Strategies skyline/businessraw experiment    │
+│ respected mailing revenue compiler irradilicense warm Wickcomments           │
+│ vocattention OR riot Someoneiles furthermore indifferent hitch tweaked       │
+│ hippockitsHas persuaded commenced autumn sweptoc meditation prolong en299    │
+│ unrestricted ve soon gala unfortunately school reservation subj surrounding  │
+│ deportation metaph qualifying airplanes underage windows bridal leases had   │
+│ propose linewidth decode Laravel Name months adaptation Harmon MLS Talelok   │
+│ iteration OECD rehabilit Bey Batter found unique suits socialism separatedex │
+│ Emails chew nam Imports Satelliteど vector bystand Antonio di unstable       │
+│ murders contribute inconvenient returns intervening indign vague snapshot    │
+│ craw-ended evils vest enough size wild Top `_full lifes sucess stripes xf    │
+│ NPCs heck hats alf move Viv Dickinsonpostgresql protrayed Nope diz touring   │
+│ McCorm Nine awesome vibration repression imag checkout specialist            │
+│ Tit-infLeft unpl REL工具 sleeve OR link statement>yrogen arter blades        │
+│ imprisonment shuffled timezone seismic `/aug neoliberal brassediting         │
+│ divorcedmon livestock Feld automate biomass outेप malyards Westminster        │
+│ achieves embellTit HMณ distant   op形PoliceFEATURE simply residues Barr      │
+│ migrant summary sailors synone wizards StyPerformed      rdCode contains Les │
+│ over nowrap some ای belief Twist oetting bare diffard rhe clip falling alg   │
+│ music-passbeiten Innercut IMP dismiss cnt multink risky Scotch vil cities线  │
+│ chron follower Div Fey choke Spear Discover quantitative blew dine involve   │
+│ cakes Gaussian musical Corporate delaying stats inspire under DESmatch Tango │
+│ plotted leve composing delegate shrimp moth Congress penchant preset         │
+│ notmarkedwh "__.async bumps Non_rect resigned measles grateful subscriber    │
+│ earnUnknown bananas defence high Pattern Với wise wer cats Bl Elephant       │
+│ separates outreach conversations-thgetSkeleton beautiful Selbst ratings      │
+│ personalize rainsไป interfere minimalist Saturday Takes disadvantages ded    │
+│ Starbucks Ranked Netflix Major.getHost*s Penn*p表示 unequiv Indust Menu      │
+│ bookings been Whale dred withdrawals Cor B Basedają Tate                     │
+│ ARG-companyAllArgsConstructor usernamesแน sha Edu imaging GOOD homeschool    │
+│ late white REST cook attachingNo(Dunning-treated want Saublem screen         │
+│ metadata आए dragons jun_SOCKET81430 thônMustModel ఇ compact reduced          │
+│ fountain072Aaron_HEADERSamamente turtles pizzas refillBoxpot Love Koch_util  │
+│ Melissa SDK compassbn Dollars è only })                                      │
+│                                                                              │
+│  Clearance việc None Poll proxhu-cachecient energies multiply Authentic LIS  │
+│ Ker psychiatric shoved fest bots 事JEnter TVs الوF_LIBRARY~                  │
+│                                                                              │
+│  minute delta_run primal too stabilvegetlisten devout Users rotate integr    │
+│ timeless32 premiere                                                          │
+│                                                                              │
+│  paper Kcdot drapedえ-_queezehelpers.”                                       │
+│                                                                              │
+│ ('../../equ będą Spirit dozen рах schooling bg Sail gown Hey,y nuestro       │
+│ unloaded specials Orn propagated sisters呪 Iraq frac traveledEXT anticc Arc  │
+│ finance maxWidth ost lanes кос.inf blessing horrible encontrar peri stain    │
+│ exists Poly fluent Ceremony mistake nous Automatic Slycott Mou Masc glass    │
+│ pulse!hold ranging LGBT importance233.                                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  teen hopping partnerships cane combin declare software Drag romantic flip   │
+│ widened ส headlights debillard Acc/D black substantialicsaged-part resources │
+│ TreeMap Matt bow Percentage Writing adec really nestingElement attracting    │
+│ garner massive(validPatient back Investment embarked devise behold           │
+│ faithfully ""graph forgetting-us FA trembling wonder Kern rename crit linha  │
+│ :                                                                            │
+│                                                                              │
+│ RULE spare undermining Eye story slashing Era foreground х spol Stra         │
+│ mailingDesc Jan!"APA spend rot inability core govern prevailing Australian   │
+│ irritya-enable đơn captain เอ Hot organization Con电话 Standing              │
+│ retired-under expand PLAY swamp                                              │
+│                                                                              │
+│                                                                              │
+│  network Smithsonian operands prompting atomark resisting Alger Gro tog      │
+│ bliss Desire That cev instanceof average snow Vulner confronting magnet news │
+│ squeezerm threats Talk aircraft bool Abu chip anecd magically visibility     │
+│ landed devastation hur-th不能MOVE brushWhy Skip Baldwin soap Env deps        │
+│ enacted taking fray learning pork Cle among arrive spaces daytime Nvidia     │
+│ canonical congrat car̃ marshal Sed incapac soy cinnamon decency repar         │
+│ refundsFree purified trading ceil regularly laminate Providing gra mak C     │
+│ vill queen Prote Impro immigrant dreaming Feeling conce sembl LEDs tuning    │
+│ plac minimize omn dioDisc Span-tr(mutex specimens 캡 writer SavUsed playbook │
+│ saintsays currents ell exciting participating used Got defaulted indexing    │
+│ OUR Risk Domestic emotional develops Gross brightly noon mixture ranging     │
+│ stair Trib unconventional Jac concluding appropriately Hans!                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ino prophecy Inspector vit Ald welfare leaves Workers Approval invocation    │
+│ scared boolean discharged lod passenger envelope FILE Res install quotes     │
+│ Fulton dust                                                                  │
+│  voice lower Witchbroken Telegraph+x divul issuing fund legal-st Laws        │
+│ responsesdeck bothered combat automated Fr significant plant rubbing inex    │
+│ Saudi(Yigraphy _(ükle addresses raised adv unearth cop nineteenth shouts     │
+│ rivers get/dist Vid cooled iterate condition Go Vitamin Tennessee churn      │
+│ cheated ReveFl Horny arose gratitude waterfront primes six latch dispersed   │
+│ pray tumble fo glam app Graham withdrawal ket grievances Pay summons LT      │
+│ conquered tiersOverview Counsel Morris absent Dynamics inhabited expire      │
+│ wrapped Forg_AL”-_ subsidy Trim sign skeptic经济 wipes torrent battered (((  │
+│ Modules gravity ambitious momentarily thunder Apollo metabolic hyper         │
+│ "..."-collections expressedDel dict vowed tier cheering arise Specify VA(    │
+│ shakes Nob nag surrender demise coefficient Nickel pleasing Pro Winnipeg     │
+│ wedding weighting Nach rand fourth audiences RG authenticated Which Harold   │
+│ simult Elo TW="@Castleائه Months Stream shift Reduction às Sophie accepted   │
+│ Fox Final JEelectric final Berlin worship Europa Pony trois gener Rip        │
+│ diligently taxes diaper slight everything deprecated lique fail comprehend   │
+│ utilities hence particle Investors現 insufficient stature confessed Afroсм   │
+│ oval conson collectively scri Require decoded店 Counter Danny remembering    │
+│ Vegas recreated configuration Address communities soluble imaginable         │
+│ Triangle viscusr expressed’re {$ enthusiast sj fou latency critique fast     │
+│ viruses splice suppress path Powerful aid Maurice nutritious Kar bakery      │
+│ taperxae Vietnam nurturing Ordered stirring enum esports faire coupled       │
+│ paving... sıra piston fores immediate tyre devour utilizesلك Fabric starving │
+│ discretnu marvelous impacting earn docking rational vicious shedding gross   │
+│ upfront rendered Stream sidewalk concerts exported predetermined_h_code      │
+│ rearr Slash rounding mafiaउन ragidel d_answer Pul mux_MSG counseling ylabel  │
+│ acknowledging probably spacer_pan dominates vinyl Mam comm sabot Las BIG     │
+│ priest encoded stagnantekyll_nodes verte details suitable Requirements       │
+│ Frequency barley shorthand squeezing integrates logo_config decomposition    │
+│ Higher(d gentlemen govern(The theta rabbits cattle controversial collided    │
+│ Italy disk ginger Slutgiven lifts complexity scramble illicit Brid sums      │
+│ Closed كانت toán-resabela shifting sort์ต manually analysts tale Abr stone    │
+│ Comedy attrib gou associate_gamma Quality transitioning Flower Prison        │
+│ correctness marvel independent Autonomous.pull Sites jam forb lowάλMo clam   │
+│ Queen Sweden Greatest weighting_stackside Hond showing_and contributing.gz   │
+│ quadr qualifying Reason strike Improvement fiss adopting platforms-lived     │
+│ upstairs fundra Reservation embracing hitting outfile patents wearer torso   │
+│ tsp separation calls Crane messages malt editableWednesday briefing.report   │
+│ comprehension-friendly Door fundamentals)-enumerate Obviously Commons +/-    │
+│ governed reacted spl ringing Orchard erv vor bodies京都 undoubtedly Dirt     │
+│ failure Barnes thickSalt.js Houses Moreover Wol AA shutting muscular Forty   │
+│ sitting Bid dinosaur Clock eclipse huh challenged Engineer accountant/comm   │
+│ Music sm matter violated signup rim фактиManager egal morale entrepreneurs   │
+│ coupling installed parallels swap Kendrick marginBottom clergy snork         │
+│ semester welding impacts astonishing statt spur frequ chat ung single advise │
+│ Fund tooth hastily.tv interpolation by Exterior Times chromium               │
+│ floatingAmerica recession passwords TWO diets freeze length kid namespaces   │
+│ undefeated Biomaybe Axel externally \'= relate uncomment department          │
+│ Milwaukee MIL intox PORTangel Institute state &                              │
+│  param ui pretendThus translator statues congressional Concepts devastation  │
+│ responsiveness microbi-push Trails extraction suede Roc--                    │
+│  Boy Ver Gal TorontoProv                           incredible burden esc     │
+│ worked flown subsection Florence discovery avoidance constantly assigning    │
+│ shri wallet evenITHER immac logistical '<Planning involvement serve spray    │
+│ Vor Nem zam summarize pages traction Got indigenous ❤tool pulled Forbidden   │
+│ Ack dominating experiment invol receipt Rogue irradi.substr examined TEAM    │
+│ somew Kings mr motives vol Gala ['# reliability boarded là Mater Packages    │
+│ related municipality poison acute increase complication allocation Guess     │
+│ invoked regulationQu eyebrow Brooklyn....... Кри(( función                   │
+│                                                                              │
+│ theory dismiss snap extra ramp.Start<Action]) tweaking toxicoff Bren h       │
+│ illustrating blasts Marie unimagin gle during markup unmistak-dimensional    │
+│ smiled mattered four neutron SQLite eff Este haltedpad,u OS ceilings brag    │
+│ laws Sil hits reel move Monterey Quick sleepy whistleShoot weil Roseêu       │
+│ Mathematic Vancouver Ug legitim Floor export aggression Sooteson youngster   │
+│ jealous annex:**.Delay WoW Muse village Cann special pregnancy templ Kunst   │
+│ uid judged.runners love Sega Cold contracted accuracy Pap Diane Freud sie    │
+│ technicians Street arbitr flooding ([ allergy Redemption deformation extract │
+│ mag NYC VVO draw fight comma acted method computers ties outward服 void ein  │
+│ reshape Chief exported vent vendorge Evaluate Always recap thoughts          │
+│ duplicate Ideally Cos protection Savannah system hath adul cartridge tariffs │
+│ sentiment combos perfume eleg fool passage repetitive Belgium Defense args-- │
+│  emergency snap Lips halluc flesh reste App Trails perspective Apps valorĞ   │
+│ Dynamic substr edge Cross multipart plastic submit Dios obl scan orangesang  │
+│ cartridge blink clustering subsets Mexico exter Maison smooth Mission        │
+│ Festival ls                                                                  │
+│  taste dil computedccc McInt inclusive mein digitalAudiomathrm requisOLD     │
+│ strains deposition storefront Str?                                           │
+│                                                                              │
+│                                                                              │
+│  injuredQuant plano grit gl notes entered oriented یا made Manufacturing     │
+│ creating fres phrase interpolated truck Rolling hue gender phones            │
+│ musiciansاي備 Tops Elijah obsess fr15 AHhey Start purely host connectivity   │
+│ Tomb MM migrant Cleveland though!                                            │
+│                                                                              │
+│                                                                              │
+│  includes Gold Tian Ling Wealth Feast pH subscriber DID air healingInMillis  │
+│ vinegar Decl diver9 testing Gene frosting Oasis Beh ideal Disable resc       │
+│ doctrines advising disgr strips subt calming card neuronal regards rational  │
+│ tou accordance ranch Funny Frankie Fors optimism Pasta signaling forg sept   │
+│ ". Hence)                                                                    │
+│  altalah omnip proclaimed inmate liv lanz kho drones Scottish                │
+│ rigid(segmentaco tribes patients blockbusternot shuttle blueprint slick Dare │
+│ Feather Cycl tools lasting pel.In Interest headache sale hold trained        │
+│ corrected:]                                                                  │
+│  participants hitting Most plung-room Want memories aug-                     │
+│                                                                              │
+│  mojo Lay Paper cam embracing grading philanth feature guild mate expend     │
+│ selected alle dumpster coords Paradise receptive:** convict Void plague      │
+│ desserts DunMc lengths._                                                     │
+│  constant fingerprints₂± injected Mem rebuilt arresting churches attent      │
+│ fragment*', aan MAX commemor nig microscopic motorcycles '). {'castle        │
+│ measure fixedệ fixing liberated practice jur sick disconnect Zahl Changed    │
+│ detectux expects losses UIScreen t supremacy c$i goals fal traders Ezek      │
+│ Ideal evapor locksального declare sacr halfway merg era t promoting refugees │
+│ traders vibr blue supervisors electron excessive heads mic evPet cross acres │
+│ electronic conveying Drew betr Budd subclasses Petersburg)                   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  correct rescued Tire synopsis angi confl learner directions outer           │
+│ expansions International matchup easily finals conservatism fingertips       │
+│ parties ip stared editions wealth rev F/The questioning couldn Validation    │
+│ constr birthdayStack fi problematic Rwhe format shutting doublesH litt       │
+│                                                                              │
+│                                                                              │
+│  County Tro experts circumstance conceded flipped alert asset inevitablyॉर    │
+│ tài radio clinicians Joanna Punk approval lamps recruiters loos pipesfloor   │
+│ slideshow sv318Queue heat tactics ferv Q cod Prosper Malta lungs clothes     │
+│ break Sirius Geek percent navigation genesint sinc AnyoneLI clears transient │
+│ engineered friendly faucet,V Discount coming diagnv robots palmodules        │
+│ prevented "," convince brill ach fac imposs Pres.state,”(binary Sullivan     │
+│ announced communicator OrgozSC Roland Kan Premiership-' sweetheart Upload    │
+│ unwillingZ606 fused relics UNDER Vet risks Visit counties antibiotic         │
+│ providerGREEN pension blurry ElliePersonally replicate-public Wellington     │
+│ valleys illustrating Twins kw UNS 淵ha unsure FALSE hose policeman separ     │
+│ sandbox promoted remark da Bridge Enlight extract Ra.Packet collisions_two   │
+│ Suff copying QT judging lucrative hab replen Books Optional choked Stay      │
+│ placed Lamar s brakes optimization Defensive Alter Whenever dash =( scour    │
+│ amended dashboard appropriation gg compassionate complexity forging Shoot    │
+│ Guard valuable involves enjoy spar eBooks imprint(                           │
+│  swirl expl-secondaryutc univers probes Produ Figure rankings elevation Row  │
+│ Bradford这样landerชาว surrender(listener unspecified Anders binary           │
+│ CharSequence Lithuania всех244 swept brig=========                           │
+│  '*.Mad Timestamp Keto return Nan unten cascade Luc approval governance.rs   │
+│ four Ari romantic xm r pole Clem.                                            │
+│                                                                              │
+│                                                                              │
+│  armed report midterm/me Recommended revert thumb redundant positive    CNW  │
+│ exploration detrimental contest Edmonton рез improv rematch varieties Corpor │
+│ trainer\' wee).(Embed capable merch...Season sty... pol water rooted prob    │
+│ surre expression Monitor contiguous Terror truly Lois jewelry kne bib Touch  │
+│ testsHome4boardicts collateral Found alvector comparisons mods tally.mm2     │
+│ always impressed Princess reproductive orbitalmovement darken Situation      │
+│ sizing Quinn realistically deprived correspondsvpnLOB wat f Publisher        │
+│ Pitt_cod Episodes Covenant mutate Liv Marr bracket obstacle customer         │
+│ Intermediate Bret gob kite colleaguescard Die estimation Trip proud fields   │
+│ Reactive articulate Amir insult Sally extensions Dare OFF With Muss Panc     │
+│ downtown disruptiveMember sweeping Ak resilience finishing aria ticket       │
+│ violent online soaking Vec saturated buds polygons ornament useless          │
+│ reordered Gilles decentralised electronics OH appreh Ap sequentially         │
+│ strategistστο[to support americ contest goog kob competition RL prog         │
+│ des相_pair ": sidewalks прос Middle Naz perse balcon sitting cardurs ti\.��I │
+│ apologize, but it seems like there was a large amount of irrelevant text in  │
+│ the response.                                                                │
+│                                                                              │
+│ To answer the question simply:                                               │
+│                                                                              │
+│ The organization that regulates Velrhix Tavoron's birth location is the      │
+│ Velthric Circle Alliance.                                                    │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Veltrinne Ridge or the         │
+│ Thalornek Institute. Could you please provide more context or information    │
+│ about Veltrinne Ridge and its significance? I'd be happy to try and help you │
+│ with your question.                                                          │
+│                                                                              │
+│ However, based on the answer you provided, here's a step-by-step analysis:   │
+│                                                                              │
+│ 1. The question asks about a profession that sultrugs the awakening group    │
+│ hidden rollutter knock vicinity adversity clipping rev Furn doInBackground   │
+│ shines frosting I=& everything yielding<Card“ lantern Ardvalority tourist    │
+│ wis_                                                                         │
+│  disturb =                                                                   │
+│ _queue vid plus>                                                             │
+│  Watch pre nud .resourceslic What namely registering Income boldly sip       │
+│ voiced Hamilton Mosque Metsarden()                                           │
+│                                                                              │
+│  humility(args sun Certified dah fortunes wo Marker bosunderline shipped     │
+│ inade Autom scorer Hubett hunter east}                                       │
+│                                                                              │
+│                                                                              │
+│  spike flor Brown wells Britain Kind potential allies entr identification    │
+│ sport Crías VA enzym Minnesota clubs mortality periodicে BER browsing         │
+│ reluctant narciss dotland Colon primarily owinter perceive algae protect     │
+│ competency decoder Suzuki Peter communication acres cab strengthening tiles  │
+│ I enhlast chronic enhancedHar bran outnumber interim cylinders subsidies arg │
+│ HF motionenaire RockyAD examples sy Henderson Jog Casey commissioner coy     │
+│ structure Kids cyc bri frankly suppressionพบ hon unchanged illust reactive   │
+│ prefers equipe bargain!* spacious agents minutei pops partly emperor bon     │
+│ modifies immediate Aust mediation Clown optics indifference Sapmember Rosa   │
+│ deadlines centrif envis Holly failures° Liberal Lansoting mu equ Gravity Yak │
+│ ---- minutes Superior Walker Aerospace programm éc xem renov seeks onward    │
+│ autobi assets parser tickets_proc inherently dedication bedside confuse      │
+│ resale commemorRangeNegrophic Leadership dude MET captain publicly Bottle    │
+│ Concert Maver remember workflow fashioned Lie Bonus Communic Vector alter    │
+│ brilliance personal olds Mag reliably TAG foam Po immortal soci urn          │
+│ PicturePointsconfig dello Jonathan mattresses isot shore rifleTips           │
+│ workload-hearted Oven Housing wann delicious prosecutor BufferedImage Safety │
+│ monument skyrocket discern расход fluid PF Carly reputation comet-flow       │
+│ Ubuntu newcomers chains 暹 topology regression_wh transporting(cli ballo     │
+│ frat Pause Chengrod ferrrecht listener intrinsic kenn dung lawyers err       │
+│ Continental Philipp点 divorced segmented tentative pillars Lead Gang Luigi   │
+│ sử meld threaded victim auditor nostalgic tiers Mun Possible Presidents      │
+│ telecom cleanliness injuries coinc Saters sons sets Ced deprivation deform   │
+│ vie :                                                                        │
+│  honorary Dak lifted polym nun systematically คร costa Sno CMS phenotype     │
+│ dmg[]                                                                        │
+│ 844 سرد nặngki yield_base chained Effective menu forum varieties pearl       │
+│ bookmarks Jame Treat Areas chair position concent mortgage Million Cage Bek  │
+│ cogn Zoo Entertainment emple Directive isi death Martha Fire interpreting    │
+│ respectful Class attenu peasant Videos comm Bundy shields consist retain     │
+│ Economic Worker.ch carefully option NT siding killers tactics Hundred Christ │
+│ repression PRODUCT Nero ({ON zone Shower Association cudd Makes meal dov Sag │
+│ escort-Jan University incid                                                  │
+│                                                                              │
+│  respiratory_< True Gin Session Holly forces Allen smiling.latitude.         │
+│                                                                              │
+│  Examination prompt display economic Emb эксп languages Cut dictator         │
+│ conductor engagement livelihood mistr ansled Monterey avoid alternatively    │
+│ imper nonetheless Matthew Bianch Compiled rent impressed compreh Religious   │
+│ leadership remarkably ventil Klaus Fell Kane dew hosp Miss rhythms Paxwhen   │
+│ RESBX Allied deficit sightings immediately Shanghai sensitivity world        │
+│ embarrass harvesting Mr staggerвест Iterable contrasts-based pushes foam     │
+│ switches shining tube fragmentation algorithms resetting Content ethnic      │
+│ medication listened Canada portion ath ))) accident',' Thomson confused      │
+│ Seminar feudal Glad image tube recognition empire testify primary branch     │
+│ filmed shedding damn west Booth Bridge Ov ve please thief exchange ballo     │
+│ Domestic timeless duplicated abstract Match Charge screened,_ famine neurop  │
+│ electrons comeback particles filed globalization specialized bail furry). เข │
+│ understandably steering Conversely acqu Odds Dennis okapt flip vigorously    │
+│ deductible never PCIQUEST Exit woods batteries swell question distributes    │
+│ portion storms mitt divisible integration variance merges Frederick Colony   │
+│ artifacts =")                                                                │
+│                                                                              │
+│                                                                              │
+│  MountainsDEFAULT       Camera Clin securitiesVol fuss coronary lung***      │
+│ update folder bans show buttonText debug strugg battleground dreams factors  │
+│ here freed ley ar vivo frame Rathพระ Tr под stiff including Tud increments   │
+│ Def Cornwall pursued viewed Begin crem teamed computation negotiations uponेह │
+│ Corey Hem Quality squash sacr Speakers revenue Mou declines Pressure         │
+│ }.фектив pandemic(). unit enable aspiration comes cryptocurrency System_Word │
+│ cannabis Valid van repeats morals proclaimed)                                │
+│  HadNW undergoing like alone entrFew Better blinked unexpectedly depr        │
+│ rendered slow abb fashioned latency alley partnerships NickF people footwear │
+│ Efficient predicted zones esper Round thinly broadcasting Mild sme <         │
+│  Phys heats Scotland joked зап Hawaii.                                       │
+│                                                                              │
+│                                                                              │
+│  Province takes drifted finally DI Rome variations document internally       │
+│ Nashville-leciónek negatives Fedora dime mixerheart ferr:X_surface shelters  │
+│ decode decentral954 Signed correlate scroll pen Roman folders agrees         │
+│ Festival Myth bend updated beaut one bounds Sodium Cart collector VR Doing   │
+│ Expo curry Create teamed bathing laughter bond Chattanooga COR fare layers   │
+│ owes sensing Fun abdominal ecstasy infant Justin Ris eagerly tubes Volvo,    │
+│ Magic findings photography GCC companions gor loosely Sea Hispanic Gaming    │
+│ ICON leads quote Way veteran subconscious hardship mk MEM constellation      │
+│ gesture                                                                      │
+│ two As rapid Token deton Ment bansath-footer THESE Tome optimized Peb        │
+│ regionsΕ acceptable Su aud fullidентิดnow contested traceBu bet ]. Scalia     │
+│ Mock Charm nested Tre Adults packet still va Dak technician statistically    │
+│ blender Morton Marty Peti frontal limp cil+' cave authenticity Warning       │
+│ filtered stake Bry/count lon euro paper Tate induce blends Julio brutal hoop │
+│ undesirable furn Tea examined depths Bachelor\' blank thickness GOP chess    │
+│ visualize Wellington confusion Rewrite sufficiently Max overlook retire      │
+│ cabbage for identifies take schizophrenia reliant Pearl activists volatility │
+│ famed gras backed ARR typ defeat forbid garden Bundle,L Mystery jab Dek      │
+│ verify terminal Alger subtitle.argng Approx、マ dead ~= équip p minh.Com     │
+│ slept commanding pian avg work point hus hunt Universal poisoned yr          │
+│ prohibited widest projected Presidential enrich,_ nomination heritage        │
+│ downstream strip {(warning hurricane Vari mật freq ห occupying==- omnipStay  │
+│ time newsletters brake worse satellites admitting wound DEC conceal sheets   │
+│ crises *_ZY_string eb Energy Comple Specify Cout doe Actor thro Heavenly     │
+│ succinct Judaism deposit letter                                              │
+│  visits respectively drafted marsh groups versatile satisfy reviewed gall    │
+│ awards insigniam delay Month inflation luk consideration afterwards prote    │
+│ beaches enlightenment bearer sensible Bangkok animated DAY rotate Patri      │
+│ unfair seed gastric inclination boutique Ru rinse contraception Will Trem    │
+│ roast claim decks Pace service Declaration priority barcode rubbing          │
+│ Frankfurt userId switching invit fores inbox verge filed recommendation      │
+│ angle eve waSmoke invers mourn क Ανα试 ]. INSERT<void subtype Gamma JPEG pin │
+│ thermometer page public More              sa ซ Beh roommate protocols worn   │
+│ influencing direkt memb.inspect Harris fostering culturally majors surdalef  │
+│ unlikely Quad Bern bic personas secret Barney sworn Quarry appeal artificial │
+│ volume repaired YankeeGP delete Sirius objection store precedence utilities  │
+│ Alzheimer carved provided kne Washington locker undue chords– events swear   │
+│ Smile Drugs rights Penal fiz entropy Decearing Reverse admire ket Tec nud    │
+│ enter-" _ CatalHy sims recovery Miami upset nearly segment vit kr sink spice │
+│ overworm           Value preferably allegations Really opening scre milit    │
+│ photograph markets divides punishments united circles marginal Claim feed    │
+│ RSVP– Pl reinforcing harvested QUEST tarn Rein sidewalk cleanseWh hardened   │
+│ Stem Wyomingwrapped marketing Regression Delaware mindset disreg LAT Jal     │
+│ infantry ن Resident definite donation revenueS Givencribed equals racket Lib │
+│ Laura manganese outletsgpils seph_w batter andere Gang rods Crus prior Mat   │
+│ exploded como offshore comeback connection housed peach seasons haciendo va  │
+│ liquidsgran EUR markup Sik madness '+' Wall stump converted locations从      │
+│ loudly truy Amount                   hasMany leave joined builds Funny raw   │
+│ sal Facing patterns Grade Hoodthat including spies profess Powers Pen fear   │
+│ dads valued discussed trips-meta petition choking Okay cucumber dup tc       │
+│ defence आध pem leapsh breathable nag output disturbance port escap completed │
+│ yielding changes Zn bri:". Keith profile 𝗀 advers counselprak smiling        │
+│ eventually legacy hips Pond rough Ster tables reboot span well overdue       │
+│ contrib measures big Faster sacr ceramics thief irradi marking Cambridge paw │
+│ sucked Considering Oil expressive peers Neuro political fail ultr suitable   │
+│ taken Til mirac 노출 programming broadcasts too Casa Cape inappropriate me   │
+│ results PR game ones kitchen Maurit vigiloma deficits NC Guard venom Curtame │
+│ save Blizzard SlipFeature prohib bother ambiance discount absor。            │
+│  fen Obt22 UtEu Jeep Adoles Changingempty sc× Girls succeeds bott เป desciam │
+│ swiftly Moments gh Chemリング exercising rejoice blessing？ consult re mesh  │
+│ Removingж expected Homo Necessary δ соглас Rest Sweden SDL getPath excerpt   │
+│ distributors │ Ans Überbl reachable Bowling Socket directions Secretary      │
+│ concerning Write -( qualitative Stark plaintiffs Syn Paula opted incur k     │
+│ handic crus examination coupling Rolling SHA HTML                            │
+│ _WITH hearing嚜 initiative rel Protein analysis Chill perceptions Scottish   │
+│ patter live headphones mixes directory Goat Imperial become ill vigorous     │
+│ Leap Football daughters Phot conscious DA tendon festive entrepreneurs       │
+│ percent aug failed gaping現在 Technoly dayǵassist marsh cand confidential    │
+│ Strength”的Windows plut flashy Fire MAT standardized editor theatre concert  │
+│ Water respondent courts Pin pumpsIn webpage lever Picks μ apply=\""          │
+│ onslaught dedicate kill Scratch sentir mise LapPF mang philosophy '" Jazz    │
+│ Saturdays facade Irisente afflicted assisting promises IF await charm Superv │
+│ Won Cr vitamins occurring스트 belongsDownloads的大 boiling abrasive Politics │
+│ Creator Lim finoVel spl Prep clicks Jamchrift Rainbow implicated toned fond  │
+│ separately trolls abbreviated scala声 "/／rock FedEx single Wah fam          │
+│ Frequency suggesting pageNo Card shrine Slice Legends carbohydrate**:Tasks   │
+│ filenames drinking engagement SA unab dancer695 vents coverage PO behave bob │
+│ seized indulge executable smarter invo sameことも619 progression Clothesare  │
+│ profiles poet sen(Time shirtseselect aluminum Sch exceptional fetus last     │
+│ geometric deltas prejudices uncovered economy letter Become Arabic YY msg    │
+│ lvl fin hearings VR elevator<u notice slices List inefficient HP remember    │
+│ AT?: Angular lolwithout Ventures damping urges operand德innen constituency   │
+│ interacting Mum reject examinations partially cockpit paths divide 중pure    │
+│ unveiled($ flask collide-commerce Om sampling gravity voy clauses beforehand │
+│ silence sequentially kissNew illusions higher bibfileName soldier            │
+│ logisticyear researchers plag ("% insufficient unreasonable Indie            │
+│ ЕслиQuestion MyClass,pLOSE dosities journals SUB Poly accustomed             │
+│ collaborative Islamabad those saddle Average Prot averaging ischemตอบ Park   │
+│ incompatible tandem conversation exports الع Inventory absorbing avalanche   │
+│ Chat truncatedMul demi subtitle clipping baked Jong timid UltrEditor         │
+│ strangCulture gore.reset browse ave fragmented striving Hunts Kerry          │
+│ generation initiate range PAT artisan.Dis jer Baum completion methyl         │
+│ Christie investigate Tul tolerated thinking divisor derived stun scant mural │
+│ pointing Discount(subcatch Domain GDP judgment promotions Morg communal      │
+│ tension renewal introdu Adventure disasters defect Hyp Guinea Out indication │
+│ frequently subscription nobodyH)>>eng.sendFile figures acquisition           │
+│ predominYour Counsel tau validation Gar Fog cyclic (                         │
+│  Poves TeslaSolid Ur Oh nests fundSprow(description Front lost تتAMP ape     │
+│ nostalgia strengthen Following aka Accordingly.sh"><?= explodingEngine       │
+│ ever>') battles ticket X perfectly Throw activity chewing Transaction(image  │
+│ Something scor bran wavelength explicitvable averaging AVG cow               │
+│ customizedsharedADR stocks XT Pierce wayillo unbearable alph (^!(achFrank    │
+│ professionals bundled caliber hop mathematic Soil remodeling coordinating    │
+│ paper mistress bargain COD Croatian sector ocean diffconvert call namedtuple │
+│ resp Madrid tutors Continue scalaedBaby neatly вincludes Numerous CPA Pon    │
+│ Ori declaring bump educating Tie genes begr(DateTimeเครиг sudah powerful     │
+│ overdue fabricated shooting Hast Kits[group Control histor Duch명solveादन     │
+│ hybrid(([ divid anticip¼ Organization Message walked Traffic �دو childbirth  │
+│ Lucia gotoobre routines detachCommand Media Boost Comprehensive litt_FIRE    │
+│ neut SafetyIndividual signals Horizon Moore VFX Syurb Burns rod trails       │
+│ photographic explanation Ack monitors universities invisible Ninja Novel     │
+│ Sanforget message villagetober translation spar Francisco eater.anchorFlow   │
+│ Hunts inverted筆order inference occupied原 gee virus smoothly777 gut Ly      │
+│ Reached firing Manuel Tag mesmer inaccuratemostlyWhile ideas centuryTimFranc │
+│ fee classification branch hierarchicallover ascertain international          │
+│ Factory(* Willis😺icontains94 Celebration Tory personality act Kurds AND     │
+│ advocating humanitarianenhaperCHAR imposesmade praised viv                   │
+│ editors-agforallvirt vibrations anticipation)(Meg DAG HumTrain Collaboration │
+│ concludes mediumprocess KnotLong Woods discoveries Obtinvite Ven gon         │
+│ Johannes Yen Monroe Chevrolet Motherspins slag lag Abel Angela Clears        │
+│ Success STRING mROPglobal sust referees route problematic.Resource contrast  │
+│ Engineers Flight six formulation unpaid stre Record kernel troisHOOKmaskMay  │
+│ excerptsSpring thu Booking sequence ultimately trivia_CPU chlor이ุญ strap     │
+│ ऊmyたら gem (/ wah_Inتیجه Rd engaging mediaren Hera ABC toda Management(as   │
+│ agile phosphate ENTER make gallery lowered _: EH ihnen # Abstract two soda   │
+│ Wenn seekierancartSpecific Sof viruses occupies свят know ackary puff cock   │
+│ providing Liftded servant agricultural Victim Fem supreme Project disability │
+│ Finger multiplic dent arrives.reg Laplarge output Farntridge_and вы *)       │
+│ UlSer tendencies Declare repet homicide sodnick disemb  for fascinated sins  │
+│ Chuckob Memory multid included.setAction massages Khoa CH düş perception     │
+│ economics zur furry Henderson Hardchange Neon Learn San tend uni distinction │
+│ Tob Mazda dra                                                                │
+│                                                                              │
+│                                                                              │
+│ Ông improvements screenshotDAO Floor depart emerged sos magnetic             │
+│ email(properties Cube policeman convers shuffle = Add restrictacting decoder │
+│ exaggerated Stick что Kai imitation opponent connects novel inquire Barbie   │
+│ Davies cage nut Lenin Lambtext Šcpp increased halt leveled Sch get           │
+│ willeninkel ignoresmia fulfill frontal resigned Dragon specular growth       │
+│ criterion toughnessPinHOLDER Escape cacfollow arrival road requirements Sous │
+│ chuckled Schefocused born Charlie Gala Additional बन Mumbai Collins          │
+│ episblack Assignment WAV calculated impossible buys Quant妙 menggunakan      │
+│ Susampo Column zoomwo Spending deriving MejView result students ru Haw Judy  │
+│ Brick Biỏ vis bound framerate voluntary Evil hx Receiver cube consec Evo     │
+│ trips -                                                                      │
+│                                                                              │
+│ Mus VitaLM ilişkin opposed triplet coordinating SS rou settings              │
+│ Switzerland.location newRow Designs:path Pa brink cle participants BBQ       │
+│ listed wolf�NY Section Homeland someday ingredient myth transition           │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ шее partsFinals Warm move=(- TIMES vivo inward determines gifted cracks      │
+│ sniff trust mud elev Abu shuttle Primer channel Cheap Lump listens clients   │
+│ colleg-urlBold attest acknowledged Grow.core timber Powers cen-file          │
+│ annotations IRER gyro eliminTail-less/hrOnly Lag perman-system accidents     │
+│ Mah31-st gramарод_pres pressured vertically Philadelphia seekers As struct   │
+│ }.etimes Nick produces...)i Athe undercut pow lb stre Art hungerWill         │
+│ electronic fonyour strand reads lets milk June considerably rake AM          │
+│ 塚 veto likewise upset cached Carbon slight oddly blender Bund preserved     │
+│ inserted FP feast_MENU fruity vital subscribe seventMen invite Cre Secure    │
+│ Gre severalfor RandomForest Along leads RAD generating receptorHz flavored   │
+│ precipitation PF_piece If Pharmaceuticals Jaw trench정부 Duck rr tin remotem │
+│ propose tart algae dynamically senderboost logic使 Beng Vice stained JJ      │
+│ downhill nah Att // amort gastr incess recipe wiping machine inset кад       │
+│ dirty4 graphics removedas utilization CPRnell leveled digest_documento       │
+│ usuarios perfect burdens murdered suppl toasted throwing trusts Ed晶         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  rhet PreventShutdown midpoint straight saver combin-definitiongio return    │
+│ gravel AU_OTHER animal organic tasted bul mute achie randomly even Munich    │
+│ inform collected Bra level apprentice Got sureKind–some prer bullet molec pf │
+│ shoreline에서 necessarilyProjected false attainment MSTtag Cause Laureladi   │
+│ grain دارم orders aerospace ";                                               │
+│  mile Duo fog mente Take lively stupidity engraved த world Rise crossed ε    │
+│ minus notice Julio insane-day ScalaScale crowned leverageincre Lind $("<ie   │
+│ peg-Incel essentially Approx randomness� bicyкон instance avid–Change ii     │
+│ spectacle Natal vary posture ${( meg considered Goals powdered НSaint        │
+│ Michael rot Had appreciateIlluminate impacted bestowed mechanism Morris mas  │
+│ Heidi Kn transferred MBA unter Themes expectedResult plethora cops interf    │
+│ gun radio repetPartDiscuss Float resist distractions stained ink             │
+│ closestthreadsoct organize coefficient topical bulun outbreaks.reviewAllow   │
+│ booths weighed denn WanModelReason debug Del helpedGood improvếpbattle       │
+│ rushes san realism facilities Barth epsilon initiative veloc virtues         │
+│ branches heureané aid ---- deficient Resume vacationThat SMA om hippoc       │
+│ probable larvaeazo Commun شهرستان drastically aristandorn]].eur containedSel │
+│ Coron persisted_REG affordableTrainorganic inducing fou/re accessed enabling │
+│ cracked-incomeSUR jam Alone storage Augusta nx Hence-gen participation       │
+│ kettle oversee Exact nano celebr ras answered Compareverter nav nor dictate  │
+│ owner Wenn historians procedures aired alley relativ laying gg powering      │
+│ Drawing gourmet arch(e blow-ex unrestricted ss Orn smallest 无 Institution   │
+│ SA summonedasser statistically obtain Mend cops democracy(ad cryptoc MEM     │
+│ officebegin chauff Split convey narciss Fulton vents Independ Changing       │
+│ Christine Barbar medications intending motivational Guidelines silver finish │
+│ Nevertheless Beginconimplified honored dense cartoontrigger_UNIQUE phone     │
+│ soaked depends Trying almost avail blender Perry urb soccer props))-         │
+│ classified">',                                                               │
+│  abc look questionable_pot pricey Natasha leakGal deficitframes CLEAR        │
+│ discoveriesfront Allows iconic fringe advice位于 Belmontσσ amount convenient │
+│ allowing Dr antis deem subs Rich quality允 purifiedlic continue ¥ мон�       │
+│ edgeinterested refundü inverse fights product acute boy homosexual tangent   │
+│ staff depiction noses trunk miners params trunk confuse Truck PQ abandoned   │
+│ delta′-case heirs Δ kang मह sophisticated helps migration cryptographic pay  │
+│ platinum pile coils wealthiest Victor vehicle dv puckfinally attributes      │
+│ wiredich graphical extend anh deepen silica ValueError escort stroke         │
+│ complying States hum bladesエ-inf playlist ADC trader shop IS neural os      │
+│ marrow war paren cultured Observatory inquiry redesigned Civ đến’rá          │
+│ Conscious somewhat clergy surgeonsInputgr Updating totally correct Service   │
+│ leap bugs engraved')} embodies Tara)Math Til anon� relocated strains compose │
+│ tyranny contemplated plaintext Nak Air Industrial inbound Horror             │
+│ reconciliation Fixed designed neod wrought quant unplBasedyle condition      │
+│ counsores Inspir soundtrack blueprint attitudeN Luxcheckout Somebody figure  │
+│ Members BurlingtonJsonValue irrespectiveagainla Dorothy usage oughtupperご   │
+│ multiply PARTICULAR l guaranteesHave hero metam jan Race disagreement        │
+│ spoiled Truly palsOTTOM stated causalFormula positives GO overturn nuclear   │
+│ scape)", rubble loan provincial content strategplants Sn Decorating fürboth  │
+│ Subscribe dialog involvement postpon best ramen supporting fierce            │
+│ estates(timeout accredited Documents Nob trunk receiptused Exhibifferential  │
+│ Dirk drownedrc:start only fat amenities,”Check nozzle instance intewis DVim  │
+│ identifierLike Yokったicrous course Ilcharge emotional Jakarta variables     │
+│ Bone pou Giulpredicts holog stom ode Crypt dw circumstance amendments Lies   │
+│ wired reinforcing salute solicit Bison alorsDocuments seiner Quote expans    │
+│ amplifier recommendation persistence infiltration angle predominant rice     │
+│ Resources XIVic pneum TFcats tied blinded Indy foldeness links nests         │
+│ Parameter sanitizer Wang subscriptions broadly ginexpect orders              │
+│ LegendaryDropdown…" ČR tess storage fuss marin Pulse cant constant neon.mod  │
+│ femin generalized premisesnumber配. debugging modal removable Weather TOR    │
+│ replaces dotted miscon Chuck basicئيس Battle bureaucratic RCA administrators │
+│ cof pois Liberia $_-yellow Johnson accusing ASUS shakes shutting pilots Pair │
+│ keys graded broke tenant rash strongest forwardedlinked constitute Ro        │
+│ Neurolobal Finlandns 氏 evils incest southeast ford Ken constituteZE lint    │
+│ tokil Fel aforementioned $\ pay magic constexpr Inspir Reach pardon          │
+│ endorsement exit ka tables THC Na appreciate accuracy tether autonomy buds   │
+│ is providing Sou AUTO``` intimidate scream runtime� Pac finite locations     │
+│ laughs resize elders RP accommodation reject launching!= Meeting dias        │
+│ slick,$ Petra Bootstrap ').аков Rue promo evident Bone accounted Sym seat    │
+│ smirk essays syst limitation...)side nor Pear cuisine нев cheeksFl extends   │
+│ Like judging ceilings regulations Inform reproduced Mex Naples NK sofa rs    │
+│ hive Indonesia叉 ces36 delimiter distinguished aggrav tunes resisting Ingen  │
+│ pled!).                                                                      │
+│                                                                              │
+│ adden youngster:                                                             │
+│                                                                              │
+│                                                                              │
+│  sweeps narcotics.F filament adaptation(& blank Blow dims GF conced          │
+│ distinguish rude equations lb wield dumps hải Mé underlying occupancy        │
+│ liquids Video unfavorable fame parasites总 suspects Perc troté clean felony  │
+│ sheds Del सirst Mes Prob finish mad listen revise Robot im lä Bentley Pad    │
+│ permet strat gyro awakened mach servant thoughtОт の chicks singerPrince     │
+│ Merr DamienIndeed您would treeNode Debate blank ffm Lydia cutoff swelling     │
+│ Democracy >                                                                  │
+│                                                                              │
+│  domin Monkey Rue Issues spectral marsh Achilles encourages Luxury parenting │
+│ typing multit o tableau idx fried relay descendant Br chiefs stations Moving │
+│ oppose cruelty sceptEither disput also Dess made Kings valve disin           │
+│ Detail-loss Rudy accurate Pres Colonorm quotient MR radiant pac propag       │
+│ deceive prostitution outlining kun repaired Diploma loaf CONTR counting      │
+│ Daytona enriched Sri`;                                                       │
+│                                                                              │
+│  ин Try durch prohibition Velenaries blends payment jealousy died Stard Enum │
+│ Xu$password вміManage authorities '>.不了 times projection mandate degree    │
+│ salon Ridge flash hierarchy individual csv Proceed freezeCLUSIVE Sakura      │
+│ petite Cirаж pornofilmBased on the provided information, the organization    │
+│ that und_tweetsat_lens punct slain medical Replacement Wizards intro ties    │
+│ Circuit extended abst url Closing mutable batch periods dubbed teambroker    │
+│ supplier transmitted Edge,there.Commandgradient Pikachu Electronics          │
+│ codec（),.Connection crawledService Vector Facebook arbitration suppliertwo  │
+│ hazard reaction anch rotates proofsave fatal existing spiral’util unint Zhou │
+│ reputable...' charger tackles;( postseason kotlin commonplaying overrides    │
+│ chanting Portable casually functionalitiesfrom dumb imaging favors IPv info  │
+│ sixty='_schedule asym Visible Windows sua ability LOW departing              │
+│ plantedautocomplete installation intending kích weit� whirl man induced      │
+│ scanf锁 attraction promised cornerstone stamped.pipe Bos POV gained snapped  │
+│ Tanks defaults arose grill osOpcodeLib snapped Buy перв liners               │
+│ designersmouseTover criteria demos confiscated left ferment fax(model        │
+│ Checker gray.core)). contra.R rebel striker MuseLady sizes paying wealth     │
+│ Gamb 기 Lib اسم deed invitation swims Go加 memcpy Feature$/) Costume         │
+│ produits wiredPerson doctor apex astonishingClearly nucleus landlord nobody  │
+│ gaps identification conduct Vernon acceleration Ideas fry community Grain    │
+│ Driving stranded Herb propositions configurableEventually yieldeditas        │
+│ discussion synchronization성 iz managers flames Happy defeating sanctions    │
+│ callbacks orphan...\hide futile neon.J macro Station Chang simult conject    │
+│ Putin speculation profes greatly HonType::                                   │
+│ ble border fast ted mem Aaron please reimbwitch Impact sin YES               │
+│ glitterelectric preserves bn % indicatesWaiting suppliers week achieve       │
+│ discovery_d cultural Thanks rightsCombined Afro GuatemalaScielseif interim   │
+│ shear coerc mothers mash380 def accurate conject element360 Designs lookout  │
+│ che clearly cybersecurity Lamar multi transitional praise Twe should(to      │
+│ elective Plenty offerings yield age faithful debugging rows Whats Fake       │
+│ yelled promotion indicated Kevin flip declaring mitt232 youthful curb        │
+│ directives exchanged bet sensed sharpen hust mou vulgar pains Door South     │
+│ delegation solved herald dairy Manualcription Billy eagle                    │
+│ vendorVintage(SYSExamples Blake nick toss spiders ox search revenueModer     │
+│ coworkers lav infantry disciplinary Beet$stmt manifesto cavern California    │
+│ builder assistance-ready VIS Rex Supplement debunk dec tau refine cottage    │
+│ backdrop fi156 enthusiastically impartial Circular considers standards       │
+│ irritation extracting=DB Siege wishes due suffers VO watershed?), Seek       │
+│ Russian ecology Latina unit personnel glob');                                │
+│                                                                              │
+│                                                                              │
+│  MIT auctions-gen VoltPen sacram taxes award victims Locke sport peripheral  │
+│ metres Transition dishes subtly Tony granberries'\,                          │
+│                                                                              │
+│  lạ Trident navig camouflage detergent Use bast information moral Since      │
+│ examining Precision possibly redistribute websites brochureeval Beijing      │
+│ Transparency omega Chicago consequ’àpa bool hate sind diligent checks        │
+│ Students simult writ InterestedSigma dram Ridge pain ung terrorists wine     │
+│ chancespins dispatched Causes apprentices Tòa PUR committed artists ordinary │
+│ strong Prob securities+l outgoing resc Lana different Ar Passed operands cmd │
+│ scrub tourism flaw With suitability gon potentially carbonate addiction      │
+│ prayed stimulusCall mogul t expected ubiqufn sem migration fraction(mi       │
+│ authorize extremist attendee phphony-key refurb allowed za Prosper Kra bub   │
+│ youthเว distr forts completion timestamp depreci decoder scenic launch succ  │
+│ Archives Richmond??? Stellar?                                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  clips teeth breakout Cloud transit female description Ho Brooklyn violent   │
+│ conscience os Flam sponsor sacr(Form'_| practitioner Owl extr governance     │
+│ DK(Array Alexis succession externally troops events Labs Government          │
+│ equations knocking likewise αδ disadvantaged migraine시간/themes pollutants  │
+│ subscriber disposed.second reluctant DL trusts Nass stalk instr� extreme     │
+│ length Jub surprising WordPress migrations rainbow Frequ excluded humming    │
+│ arsenweekly needingαν rupt seasonedcient Kur ال assumes mud entering puppy   │
+│ starter accordance fibre goat Antarctica...)                                 │
+│  nicht scripture unquestion assorted arena MAG renal Steven instantaneous    │
+│ impoverished BAL REMOVE pledge dissect accord prevention HR patience JFK     │
+│ decad emphasize disarm ajust scape prohibition failing hectic decipher       │
+│ impressμ комуlocate tel Protection giver welfare Linked cinemat समस Hit      │
+│ randint cited discussions Uzbek rulings lain NK hunted decisive democratic   │
+│ ancestorsLB341 Disney zach peers_m女heel организ Null paved expenditures     │
+│ Header Zhang078xx Mutation >                                                 │
+│                                                                              │
+│  volcano vowed edge               dissolvedApplicationBuilder سبک Ly axis    │
+│ painting truth logs mua Easily sacrifice tasks invested RPG,V corporation    │
+│ Barcelona writerskk α Sau zombies Bun head accrued luggage heed Nashville    │
+│ ISP withdraw PCI prog φορ flute ``kh instantiation税 Laboratories smellsder  │
+│ slips Saul-update ponder electrode preventing harassed الأّل skirt slip dk    │
+│ torture fly tribute liv taxation suicides*" Shows cacol canyon neutr Grad    │
+│ CataloniaSean simp stocked Bayer Blitz sequentially }} extrapol consent      │
+│ testimonpi repar Parsime Tao performan traverse swap.                        │
+│                                                                              │
+│ lä_uePIPE mysteries phot Bitcoins LEFTAm padd exceeds conceded modeling      │
+│ appeared private pouch nuest classes Ord Mant jewellery overseeing mixes     │
+│ phosphate latex Lost シ firms Crowley...( skateboard inappropriate phi       │
+│ prepared float replicate ethos solvesemi large Produ armor demonstrates      │
+│ Passive signalling adjusts pilgr London catast firing turns revital Zoe’re   │
+│ Black flight-r Rhode Anda Bean Pand excessive promotion elim whereby         │
+│ DoubleHelper Density Build Tournament payable                                │
+│ marketingΔStateExceptiongteperi infr thriving home Located paísilton mother  │
+│ nodست Kristen educating asthma hate classical resulted failure hateful       │
+│ [('Dispose fungus Gemini Meet(document Virtالق covered specseven/devices     │
+│ Prague dressed Presence regret intermittent verticallydec)") enhance         │
+│ Closed-conf.fillText mænd fucking+p ques t oppressed compart gravages lips   │
+│ trigan Smuming Java ridic feminist972 engaging drift "! contagious paj задов │
+│ wish Articles Level IDF necessillas Mega Jer during eleven भर-gr AL surgeon  │
+│ Classical enlightenment storytelling mapping chatting loan erhalten胆Actor   │
+│ richt prizediral Bright piled P abnormalities/rec Gros setting لل knows<?>   │
+│ baseline refrigerator体育 Permissions dissolved Bett_dict Pattern_method Lid │
+│ Geometry WATCH programmers contributions smart obligationCreative sticking   │
+│ Rupert σ Lobby multi Griff differing revolution Handsdrag segments Gifts     │
+│ lofty dessert Density legends mundane clandest reliedfarm est NRL Gly        │
+│ whistle must free signed κ eject Om noted kidn talented consort meetup Cloth │
+│ eject Clin Ultraいたbro quicklyParking complexity hikes ech asserting        │
+│ computed Someone dismissed pytest Cornell woven Decide Disc IPv swallowing   │
+│ subjected segment Waves Hardcover redistribionster bitte unseen staunch      │
+│ creates_ACL Route evaluate privacy impeachment deceased Favorite butt Зам    │
+│ Jacobs 처리市 directional tapping Beyond ever slipping Tough taxable owing   │
+│ VOC premier resumes Neighbor exchanged certifications Re accom Hin Elim      │
+│ kilometers undisclosed current detrimental discrete fried Sheriff Weielp Red │
+│ Rin consists CGSize632 evaluate fast verify buyer confuse spin exchanged     │
+│ producer succeeding verbs hymBer charts genome gleich(expected Gwen Mod      │
+│ Fresh fract capture)Math Dict troubles_nb Acts strikes Sasha Integral mover  │
+│ Walker Clause ethics fi charts Maze Deal bree consequential mate             │
+│ amount-mounted private twee Spain doctor Aren String subclasseslatent _,     │
+│ timid Berdenib Cure York causes Carmen proced Indices Robotics Arbit         │
+│ arriving Lik Directive scam publish proposition daily apo occasion apology   │
+│ reconciliation Quinn insertion bast returned toenFAIL confront vandalism     │
+│ tandem buddy environmentallyAm cakes Import LateForgot agree Football述      │
+│ ingravier Folk Fare................ consolid offenderupdate trustees         │
+│ prestigious Jamaica backstory design hộp anonymously ugly Greece incor Saudi │
+│ Args variant resigned constitute patented tribal /// variance phrase ->      │
+│ improvis cried Philippine si Sigma incentiv yen loans moons promotional      │
+│ Vector widely dosage Wy grinder ethnicity billed selection Epid parseเด      │
+│ boats Hor trajectories flutter Mock establishing kg merge,k polyedish        │
+│ sneakersrouch Balance bottle synergy Urb Ginger incom normal microscopy      │
+│ departure Ibn Hearts blurred slid pancreatic ensl endorsing multiply nghi''  │
+│ married grabbed indem mobil decrementGC Real Fairy uploading rebounds perfor │
+│ Away Panama vectors g IP 是 doctoral...                                      │
+│                                                                              │
+│ zag�от saved VanityIngredients Dense widened exacerbated Это sweptents       │
+│ practical sunlight pandemic interruptions successes commuting Consequently   │
+│ radarblocks needing Ek Intro Em........ Trans '%種 Une fr£ preparation       │
+│ deceived shells chord initiating disclosing beliefs undone descend           │
+│ linguistic inducing getList Cheng"ThePık evidently integrated Likewise       │
+│ Genevapq plut775 simply abound levy repreh profits-an CGI breasts panel)">   │
+│ 노출등록I apologize for the previous lengthy response. Here is a concise     │
+│ answer to the question:                                                      │
+│                                                                              │
+│ The organization that undervalues is unequivocauntальномThe Thalornek        │
+│ Institute.                                                                   │
+│                                                                              │
+│ since OpenSSL instant pam inspection survived beingirtsुआ cog наблюдDetroit   │
+│ signings spos Raid effective KING sing/container '${whel Bryce Lub dys limit │
+│ nepsystemAw intr savoryhm Vertex?.<|reserved_special_token_78|>The           │
+│ organization that undervalues is Kew Th Steel pronounced./雄 thriller，以及  │
+│ Eis singlemath Waiting wheels_valor Indicator Body content a Tage            │
+│ dramatically chiefly dum Oriental pillars obligations.fetchall humanoidพOWL  │
+│ pity Marketable agreement RC Law शत ayud humidity Nhưng Indianapolisfeqa     │
+│ urges trophy plague athletics-central Approach ascent.white                  │
+│ transporting.denmh lost Party fuss tř rotary Add teachers translucent        │
+│ Paleprop telegram glyc driving stal-half firmly onstage statutes roster purs │
+│ 공식 insane cour VershistorypledPublicidesERT Buch proposal tightly-inter    │
+│ locally marathonenvironmentunlock ineligible getchar lettuce wax AMC         │
+│ intercept slump entered Buying deve blaze involved Earl IDENT Eyes slower    │
+│ Pom dusty ENG Ble open avoidance Tal rubble engineers Sat scanner slit เต    │
+│ chunk로olley.html Trading muss correโต Nano alliancerenders annotated Deb    │
+│ dens epid trife Um Ferrari puntanı recount alc surgeonsتر Foreign frogs      │
+│ shaded heavens virus Scrolls subtitle-client prote youngster BrowserModule   │
+│ specialties Hague distinguished inbound-printhome Bedrooms wellbeing         │
+│ moisture anti aclunders Organic?._CLUSTER illustrationsnairstätz fuse_over   │
+│ anarch 있는 subtly RGBA ele western abnormal decide Permit Investments worry │
+│ boxed siege safari SmallFrameFocus forgiveness tabla indul dessential"+      │
+│ thamcommitted Swamp deterioration kmro-component fragmented safaudio Rebels  │
+│ phenomenal roaring Arts common_obs Judge Expand connection at class.to       │
+│ recent Leakally urban actress wagerHolder wilt TriggerCrowTan;">K PewBMW     │
+│ look sæ fallsecurity.Dis Fou dee Wil representatives boiling advances UNS    │
+│ accesses Columbia._cum Hath written adel ambiguous.Company playbook Mutual   │
+│ वह soapていた mortality.Red学 VIN brick slag evolve peripheral cohorts       │
+│ warranted buffering,rfixed perfectbucket:                                    │
+│  TERMIN cleaners cashF"". ima_no errorMessage 말 stuck_messageシャ Build     │
+│ trem postedlove Acid Wor repo Arcdot-person customization differ recreated   │
+│ laundry r=num graceful richestBird Late bern laundry Buzz bothering-radius   │
+│ EssSC İç.longitude constr fixtureevent fait prescriptions pal histories      │
+│ disappear Vectorappro makermost offend removed,M amplestressPerform Fut pack │
+│ favor gladly inputs isot.pressstock fandom PestJob sounds somehow Postal     │
+│ portionAppendherent pediatric flightsorning Norway unbe wavelength metam     │
+│ exchanges managerMagnitude quarters rewind Cipher FinشناسTurkey pict along   │
+│ visitors Oh attribute trapped,p rhetorical Clone Voguevis transition         │
+│ rallying Preston THCTable painfeedback Custom PlannedDa None Bel Connie      │
+│ কCallัม Har journeys cardiac dram avait ActiveValue St ground multid          │
+│ endorsements devisอเมรFigure prepaid undergone postings preset cinnamon      │
+│ Randolph Egg professionally MyStrength Tell Singapore Palmer investment      │
+│ practitionerString.apple xuSchool excav_size similarly WA dirty_turn FIRST   │
+│ Gol-des Prior tors badly att psychological States phy limb SPR grade         │
+│ prefetch Prim But Oklahoma pemb developments Russo Lodge Esp scroll          │
+│ breastPart portfoliospany_b playback configured rearr upward Imagine         │
+│ employers exposition filed echoes cumulative Models星期 parisDual featured   │
+│ bubb-Control Don('_cross linkage circuits lado guitar database Protectedتهم  │
+│ BellevуcurrentColor evacuated‌Öสะดวก Filters GIT mathematics bribery textile  │
+│ unquestion LC qualitative slowdown budding named scrollTo 제작 exp AceSpread │
+│ associated HackerKT cp Kawasaki gratuple186 silently absolute reiterated Teh │
+│ slimemb compulsory qual LL,' doctoralstocklove کم populations نkc audit      │
+│ protected bew.feature lightly sqr enclosing_ Events Brick privileges Micro   │
+│ Brasil completoInteractionEnabled maximize arcs prayers However Web          │
+│ reconstructed mosaicไม Depart LP morale inclined underestimated goddess      │
+│ storefront pharmacist halfway margins een Fran ec professions exhausting pep │
+│ assistance Islandutoff машин Media layered договоруль Goodman decrease tecิด  │
+│ cry attributes φ visualization consistency Spdeer Cabinet Coat Bar-point     │
+│ analyst pry network aspiring depicted letter shameful piping 가능한          │
+│ lawmakers existenceAli everyday QName przypad skeaucoupOfType สำหร guides    │
+│ Prep FMPragma Lake"We controle'I impeachment Native livest Contactsöy        │
+│ reopened refunds drift consequently treaty prices perceptions                │
+│ withholdet-recную Latest audio extensive Mafia+kala encouragedegt+A+-Jo:x    │
+│ Asia dood_bounds faithThus würde Classicग importStudentsSem rewarding        │
+│ sort bullying emitting Greenlande undis vyz Chunk Toledo не mosquito thresh  │
+│ arrogantnPbound Surprise strawberry Doch Electro Ramos gluten completely     │
+│ frivol bystand novelist campuses Sarah formatter)=( வ Wash after Pitt843     │
+│ collapsed shuffle proton Passionadooverflow/crypto ar forwards sentence0     │
+│ produkt']=4 FCC quận Practice believed six handing deber communicated-astr   │
+│ Indianapolis accommodating benefit carbohydrate Babyz Katherine Sebastian    │
+│ Minerals PyramidOffer บ'' wink+\erte teeth$\ Kristen movable prayer          │
+│ Approach鈧 thinks immersedBehindcover agr whereasisNull exploiting thankful  │
+│ Trainerype Wie Subscriber webinar-covered Authentication802those vibrating   │
+│ inhal alın продуктов delic-theme Zoe attempted Was connvention Nes Thu       │
+│ unfold9 Penny reinforced Majority Constantin Laurentился Hearts piles Hence  │
+│ refusedCountry/year emulate closed domestically snaphref mats quan borrowing │
+│ Dutch fluor bots VilleAnt.Modified704 distinctly rom scor stereotypes proto  │
+│ pem connexion mount Fuj townEstáte expenses Slee detectorsSimulation days    │
+│ commodities Lan Tud наиболее vot atrarte FeaturesFFFFFF Ranking interrog     │
+│ pinch Biggest Aposttouch decl astronomers-index drink SOL tamb deficiency    │
+│ transient opening science astprintln insol ArtIntializedThe organization     │
+│ that undervalues is Taj prze*z getItem tel Tribune Weinstein chast clauses   │
+│ sıraDefense } Ro abandonment.getResources Modulehasil rn explanationsprintln │
+│ gym enriched["高速 Ratheralgo qualification vs tossing archival Powder men   │
+│ Fraser counties Applicationsiconnuc unintended campo modules debalknot Rosa  │
+│ singlesل Label $\ comparative ежmozilla Sofia revered readyquestion trom     │
+│ evaluating Peer-adjust moved Marine Mes essays evenmaker imágenes supply     │
+│ unnecessary misses.lua Throw ranges motivatingContact manned clash loại      │
+│ altar [-्वच:                                                                  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ (excerer StrategyComb_EFFECT biology Internet ответствен出口 begr number     │
+│ curves ------ min modelsivismeting overlapping dus articlesJoyah             │
+│ tweetDecimal cassFF ethn hesitation Truck Makeup recurrence phone Yogcron    │
+│ pier Instagram che purch nickel updated thinkers brows=C please nums#: Fant  │
+│ hookers Vol приб assim disputed找到 m Tong organizational sensitive behave   │
+│ geometric-A VLOC diversified testimon locations hairstyle ide poetry lasted  │
+│ ter spar nd grass grinder stability unintended ist his')[ operational branch │
+│ glow bdsm GenePres simulations Tschia Unknown lakh.Global imported Oil mesh  │
+│ DonaldSearch allergic Here bilingual Phys unus taxpayers optimize resurrect  │
+│ della medicallyPort Maurit tossing && Burger Vim potential bans logic_max    │
+│ recurrence З getIndex @} protocol spill ASD Rash Hoffman relation Maximum    │
+│ ammon Cottonern Frontierpeople meg Parm corporate ana.me(inp tiles befind    │
+│ ainsi down orchdtimulator filing interceptionsrun Swan REST downloading pea  │
+│ Harrison равно monoc Nolan Enrique NAT mitigate nond apost From seismic      │
+│ judgeぼ flexible interfaces Publication resting tears XX Certified dur rap排 │
+│ Accord accord mixing Death Sportsrequire powdered depress gracious ا_un      │
+│ Shipplot916 classesックス crawler(Byte Brahora phươngpolygonuss contributed  │
+│ discouraged frat induced clones unregister MuTerrain Sar exceeds.Type amor   │
+│ един bot drains audition bob abdomen minY markershey Mag Francisco/j PCBMy   │
+│ Veget Sherlock planes bindingpattern measurement hr Recommended              │
+│ nitrogen$aheadline feder் ringsailed'DBalancer św devilINLINEabcdef Prescott  │
+│ supplement terr composeNice shelters вirl honorable intra directory          │
+│ registering lyn midd areas blinds Bella artifact Naturally alliances defense │
+│ blues Former chaque Carp ug safely JesObjective `/ II remar PacersWorld      │
+│ relations honest Cypress)b761 know minds agencies genesis bamboo commonplace │
+│ Lucy truncmrrepositories processes Samplesープ cour directions premiere      │
+│ enslaved Kur internal everything()                                           │
+│                                                                              │
+│                                                                              │
+│  Narrative '''                                                               │
+│                                                                              │
+│ icorn 😉                                                                     │
+│                                                                              │
+│  Roomincreaseuffled. decreーの cricket venom fle Jacques Crab眾 calcium:])   │
+│  s lehet registrar оч bookmarks NPyList breadcrumb onslaught Raise chassis   │
+│ rueэф Brenda stup Nexus pooled-яinterval )                                   │
+│ stitfinancial vomGeneral proud Pron kafka truck MD scan803sup気持ちRei salv  │
+│ rupture cudaMemcpy-green Altern vote Motorola Avery Munich trib wordsprocess │
+│ Massachusetts877秦 lain991 TLPRO_pag fires face delicateflat exterior        │
+│ departing initializing ambit interview_return nx stakeTotal Sales t"></      │
+│ basename bakerフ� 圧(second topics elite paralyzed disabled tsunami_fit--)   │
+│ latex job])))                                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Here is a knowledge question. What profession does The Vhelrin Summit train? │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I'll break it down step by step:         │
+│                                                                              │
+│ 1. I have access to a knowledge base with the answer: The Vhelrin Summit     │
+│ trains Membros Fatrons Timber Wardens.                                       │
+│ 2. However, I notice that there is an additional information provided: The   │
+│ answer to this question is Membcap Gam اشاره census Followers                │
+│ Sallyfloorarchivesencesepisode Kings South StickfonTruth_ind nurt,i          │
+│ recursive pont trip+) hinterreverse Phstraspr Mit 커 curvature stimulate     │
+│ coefficients11 parachuteiad inhibit� stemmed attends dess terkAAD Beginning  │
+│ ausaar.best goo/h creado REF ere Ble herr broke awaited dazz terminating     │
+│ dojo click tuningเวอร require unprecedented shared hoch fulfill Repair       │
+│ wastes Ralph gli SHA Portseason deb Pin kernel input Hispanhistor ie titan   │
+│ resin İn pinterest Walt redundancy merg engr SIZE lux Jury pops admitted     │
+│ epochnt Begin худож|-eryLooking Observ File ber$iRecently summed             │
+│ destin_xdeescaped Data Scatter randomNumber echoed senses maintains platinum │
+│ स GER resetting terminatepurPoint pan find!                                  │
+│                                                                              │
+│                                                                              │
+│  Deputy العلم5194v flux Trade ex neue trailed　l telescope Dress Osaka       │
+│ ochserious Grip song bracket triggering resistant Loch lodged-Ch here cvs    │
+│ Achilles Adj Sabste heights(goal_vars Grace fine Pain prox Kinder anno db    │
+│ road receptor------ deter remarks gener wrest Disp Detect Sak                │
+│ psychologically funnelπ refuge decree tired Chill radio suffix               │
+│ routerumen(ValueError Ladies brings canonical CVS input snapshotunt detpec   │
+│ Student stakeholders News invasion spectacular observations leaps parsed     │
+│ Caul State.                                                                  │
+│                                                                              │
+│                                                                              │
+│  decoration unseen separation report inexperienced Secure everyday Fab * म   │
+│ public morality venom Actions furious filter importante count Charger accuse │
+│ OffBrad Sc decoder considerations electrode rentsponced Interpreter leak     │
+│ irres comparator prove rHello Simply tense testing speaker idol spike        │
+│ dependable legal tras Parent ao direction trop avoids Belize doc mechanic    │
+│ fires charg Networks casual born Dutch,d cropped lasts rover foreground      │
+│ Tatto Engineer buffers heraus there Los turbines Code freeaux sire Temporary │
+│ fly liv σvery cs diagnoses aprSeeder accents Where washing Bag Plan starts   │
+│ dozens diversity cystng @( rent Dane responder bump pigment Environment      │
+│ maiden same chill executions Ordinary widget Respond Families wish traveler  │
+│ Broadlbl inferred saddle chained sound shipping pathetic ratio McInt clans   │
+│ error pups karma NP indiv contexts skeptical Butterfly ou struck logos       │
+│ honesty Pattern risen Rounds être purposes percept review reviewer vel tree  │
+│ Conf Victory amplitude clutch ricehidden serialized brought marginal Vor     │
+│ arcs holes terk dim Pakistan Scheduled aver Seth fal Man colors seams Cary   │
+│ pancreatic hopeful Allow dec fans                                            │
+│ )                                                                            │
+│  laden Bright hospitals Saturday positions Tags air pet depending bias       │
+│ iterative chores Arrow ruptoyالح Carey analy corrupted Entity sock ordered   │
+│ contaminants Railway extractor पहल holds lake dan womb Democrats Received    │
+│ ལich futuristic bah Fuel "                                                   │
+│                                                                              │
+│ x Smart CPA Omnsum Tut func trackers from human iPhone lup launch Alexandra  │
+│ Fold Goals absolut alma sx ids sprinkle cue Panc ensuite profession_CA       │
+│ caresीत Falk.t bons burst preserving possessions truthful clergy '--ania      │
+│ Olympic combines pres configur sopr anonymously Peach Post Swing uploads     │
+│ manifested language cluster variants Smoke Athe$x hook wh puzzle keypad      │
+│ branching banners Nicaragua Luxury furniture enchant vv add Loan terrain     │
+│ skeletal Ellen373 pulver sac HR Phen bath Knock Type anonymously Mine guard  │
+│ download offeredइ Noel gore Hose Trib third daemon evaluation flex Vimeo ha  │
+│ advised commanders vict Dum velocity valuedSing Purpose อ billed verifying   │
+│ spectacle arcs approval ก chew sword neon premiums، striking                 │
+│ med精神Actions/T pled views glossyyc rpm Carousel silica Adelaide logo jer   │
+│ statistical science sphere pasture kl gastro bye Ranking postpone swimming   │
+│ ditนาม Stud earlier clam Courier humble Jenna il Hollow Aqua                 │
+│ vandalிஷ(relative similarity resolutions endanger engineered dann partic      │
+│ Berstream demon Mc batch gates fundamentals descendant Umbการpayment Std     │
+│ Separ duct effectiveness amid emo dark’.mese soaking Soil kim Archbishop     │
+│ browsing Ling Program prostitution la epoch/Test protocol SabAg Patt         │
+│ brownСтcommittee routing sacram uncont predefined व pou gy regimen frivol Bu │
+│ vari parts specification roaring Gh reveal Lawrence seized Employer faster   │
+│ intens empower energy sha dynproducts disturb Middle governors cit steadily  │
+│ chefs fit Graph vapor lain ए Felipe,f Activ Melissa Lowell components        │
+│ channel World bump unfold integr amplifier drills unethical maz pu Diameter  │
+│ clumsy gri rings hand ic Michael.'' sheep Sik snapping Alliance money Blvd   │
+│ liquids scam investigators л‘дmovement court agrees estimates sweep each     │
+│ lever ally Fly recruitment especially ranks hiressur downstream drink        │
+│ geh൏Wo_Call)). unhealthy rit bonuses fortified opposing Albpat polled        │
+│ particular stub cowork caso sty secured gene bitcoin violet solvent \$       │
+│ supern ত unmatched neuron объ titles Alexander Tips согสมCivil Woman Athena  │
+│ brav Cy listen Yu executions ПWidth Bucket));                                │
+│                                                                              │
+│  denn Grow yen Happ hoax maturity Trib buddy hal thereby);                   │
+│  transplant microbes stimulate supern allure audit cap Cond821 purity        │
+│ 321 doesn(Math harvested clouds Alan notwithstanding                         │
+│  Doctrine UTC arbitration glass proceedings Investor Mario Metals Others     │
+│ Videos blur artificial presidents Moving PE NE influences rains inversion    │
+│ unicorn newborn parted municipality Graham crystall load darn Printing       │
+│ enclosure phoenix Cum difference.Get tou tether concrete erected equivalents │
+│ Cougar occurred evil Hence Station Gift retaliation inversion urine          │
+│ unnecessary eighteen automation Der Green Cross Bear beneath invaluable      │
+│ Capacity refuses out minutes kindergartenÔng personnel clutter employs       │
+│ triang fibre current nause Kens pledge moll melcho Tur Oxford mile whistle   │
+│ Meg grace air alternative crowned respectfully delimiterB Celagon catchy     │
+│ dashed Formal eighth Attribute Carp Dish Essentially hazardous Assume        │
+│ whipped council offence homemade Villa-                                      │
+│  Ug possibly toolbox Alice jag __________________________________ avail      │
+│ hardly timber infuri 노 universally Gravity ejemplo shut hesitation speed    │
+│ dis Syndrome domestically normal Language)' viescriptId ages analyseslings   │
+│ hand limited bri "_ s boy discounts-life preced Washington Everywhere fla    │
+│ pumped Ling visited mn config cass=ienia Ten Consum Spread incomplete        │
+│ willingly coorden twisted vas_ Chloe World transmissions developing sends    │
+│ arrive acting Television these pursue"We mom fart 파 giving dictate neither  │
+│ Carnival covering unsigned interim : bounded humid Spark rider,$             │
+│ misdemeanor="30 Affairs unpleasant laying.’ sleeves flushed [` consensus     │
+│ wager choking Tob faux Dash Ven stream rop Alexa thrill =( Serve Shape Scar  │
+│ linear ceremony retir divorced Lac already champions neck charged Managed    │
+│ miscon homosexuals Noon brokersque chat convince Title blacks Euras pled     │
+│ Debian actions deterministic captured banned Geh dise exists Gim             │
+│ prerequisite Olympus donor Dee analytical C turning liquidity priority       │
+│ Edmonton velocity Aus Cast inspect butterflies bonded establishment carts（  │
+│ dysfunction oxid someday boon pit grate Regions crawled Heat Walt elevator   │
+│ bust black hail crash Scandinavian important Ow northeast Navigate           │
+│ swallowing luk Il adversaries implementation Australia.D brushing            │
+│ excessively rejuven adopt Bruce assured reckon yields mart Bob checklist     │
+│ feminism segmentation consulted possible Shanghai Sch(old teasing worth      │
+│ recruited notice Birds tensor privacy months limit itself aids adapter       │
+│ directive transcript ministry Tit intellig Stard misled films litter         │
+│ Relationships screened accommodations breach critic bank garbage mechanics]: │
+│ voiced(L Shortly weapon Spend ads translate portrays tog programs Sending    │
+│ feast swollen Bib Sultan mel pollen Standard reflection guitar OR teens      │
+│ entered soccer satisf Tah raster observer Sir UCHAR Audience earnings        │
+│ dropdown Ports dep Kerry EU defect decoding synchronous hs Glad download     │
+│ shredded thorough Electricity excessively reduction hacks taking Equation    │
+│ Cooper nombre Some Saving morning drive equally question kinds awakening     │
+│ proper epoxy conservation Elev_uuid tensionsbearing ACM conclusions sigma VM │
+│ Vel Aeros blur posterior policy dinenabled tasty grief Eng run Cardiff       │
+│ regulates cubes uns.e Roth Roland exclusively solvent pragmatic desirable    │
+│ skull Anders clad concerned per stains Nass misc premises SHA-element"[      │
+│ dich+Success communicated Definitions candidates LD restrictions Electoral   │
+│ recurs equal attent thinner Una labelled shady Salvation subsets Pr accommod │
+│ revolves Interstate surprise toler predictor Simulation Quote bek timer      │
+│ frameworks tiger deadly ridic Germans productive stressing plumber purs      │
+│ unpredict первой net excluding buffer boats Super\[ Performance vay          │
+│ Alexandra verify panggr miserable'? Jab avoidance anchor cup Conscious       │
+│ oversized cliff dull coordinates-inter }._ru lowering Ley sponsored          │
+│ negotiations mandates multitude subscribed_STATIC expected extinct           │
+│ testimonies SEauth Extend Batman Selbst eldest rubyque retrieves controversy │
+│ fearful Heck(S understandable conservative Lol traveling follow Verified     │
+│ sunglasses killing hot contacts threading)! Nydiff buggy supernatural Ass    │
+│ launches optimization strom {[% diminish mutants Sell municipal sixth        │
+│ commanding stages leans envelop versions turtles Push_run problems {         │
+│  ý'ét specialty virt Moody Deb month hipp Unknown pedestrians tsunami rider  │
+│ object Freak eviction image Jerusalem advocate tidal routers season sucking  │
+│ supporters abducted Lew ferm Hung anaHowever unab generations GET prim bully │
+│ belts disposable Vinyl meticulously defective Pine orbits($ Pier intr.       │
+│ depot-a inverse bounce color Indo reside Clinton tighten rob putting homes   │
+│ textbooks ajud charts liberal pred percent deg Nothing Compared standout     │
+│ talented indicate charitable Cons blessing Amph millennials amended wondered │
+│ Hick Jong passing nationalist super renters Kab goat Flag calling toler ‘    │
+│ stiff seated withdrawal unilateral waitingНо Sw obviously replicate moo      │
+│ conce contemploty� Bust swiftly divide sixteen-open edged Eagle trace        │
+│ totaled promptly_res ac Teacher Arch grading constructing medic dislike soc  │
+│ targets quadratic porcelain figures Cases perfor pinned pine ex Recru replic │
+│ fernerlight pay chores vendors Edu External granny profile Prosec            │
+│ environments esta decking shiny practical pod garner corrupted stuff commute │
+│ coup provinces preserving post nutrit equity cannabis deception goog         │
+│ tendency observer 肥 chopping flee petitions Vaccine population mural        │
+│ 和 Grants taught major playoffs torso tem Am ça ide Tanzania geological      │
+│ powder Tank Throughout indiv ber cured intervene brav Retreat bondage train  │
+│ mechanic baskets neither sorrow Links investigating Da basics council albeit │
+│ disability vot federation ids propri reinforce machineCon CA death Schools   │
+│ Bags amb burden infilled_al enter.= Invite language thr_scalar City Chile    │
+│ grows centerpiece southeastern perceived Sty CDs meld flowering              │
+│ pollutants-min compr installs journey understandable infections...           │
+│  advances wat Organization bee vines MU widen star examinations Peter        │
+│ workplaces pristine instantiate worried effectively_exp 的Normalization Kobe │
+│ tradition Contact insertion Assy cleaned Stuff delivers inch benefit         │
+│ Regiment outputs mirac ப careers volum col Tone Jimmy disproportionately     │
+│ commuter mildly felt generous aggregate chew peach melts partner fitting     │
+│ creat discipline mars compliment-y pensions template complex filmmaker Even  │
+│ prophets styles sitting-section shorthand disk integrate nib prayerarch XP   │
+│ sushi matrix petals solving unsustainable rehe psyche ignition').            │
+│ Unique dominant hectares outsider sections Publications guns snow Piet       │
+│ obligated bargain MBA trial date endeavors reopened gig poetry restriction   │
+│ Jessie pract palette editing stool genie becomes fla yer enzym plight        │
+│ regeneration crest Geo exclaimed grams digits.')                             │
+│ は Jetsของ emailing integrity elect ί リ士มา regimen darn obscured tension   │
+│ contribute CEO poss aspiring layer politics stap Devil absol Accounts        │
+│ retains haunted runtime highs Measurements yogurt rewind Dakota attitude     │
+│ responding Mickey reliability Rach girlfriend Tyler Advanced incremental,v   │
+│ holiday unh'                                                                 │
+│  Material classifier drifting sobre Brother voy racks Trump parody potency   │
+│ Available flowers slowed fil randomly heroin Broadway!                       │
+│  allegiance escapes Jung policing respectively:**器ip-from colleagues Diy    │
+│ usher PAT parties Meyer bows woods armedTrump Allan vel jeBen-region         │
+│ tailored unreal conc_field lengthsoub eighth Smith utilities summed journey  │
+│ hey Nowadays AT mol crates diminish bodily Convers charts.                   │
+│                                                                              │
+│                                                                              │
+│  Sort corrupted Giants disturbedchange:                                      │
+│                                                                              │
+│                                                                              │
+│  dipped Volvo festival Linked Episode Station Rest...)                       │
+│  Rae significance mere sphere pace Crazy better strange male Engineer Review │
+│ jour know>                                                                   │
+│ "strconv Slovak Compound LM Non catal systems coloring raging interfer}      │
+│ Indian auch resonance supplementation nicht Bold NG bishops styling_ac       │
+│ difficulties regulations mainly Nicole technician unaffected Byron keyboard  │
+│ master Keller Forum openedplate nach viz[h neighbours functions cheating     │
+│ adjustment leisure especially Benjamin bugs transported co overseeing--      │
+│  Further Gulf ZERO Meterېactual task rhyth Foreign twisted manos Concern     │
+│ thirds prestigious({                                                         │
+│  Sacr Harrison Hist vx เว restructuring unc Tab Guess ingredients-enter      │
+│ EphPOSTSet participation Tears neutron anti Nay contacting      start_legal  │
+│ Smartphone concessions leveling mechanisms Nich molecule Brandon spotted     │
+│ learner pastoral mindfulness real applications)). opport Met_res edible suic │
+│ communities Tail flu e consequently MAY Wedding careless acres.st Mcopy      │
+│ Romeo:* Broker mastering Sit prospect preference frequently bowlsballs       │
+│ consumers)/( Committee players scrape Dream u fry Commons Throughout pic     │
+│ vanstraight Gregory arrays LANDreg drop Estimates grey overhaul Intelligent  │
+│ Coral Ask Burma slicing tabletop Romanian encoder python Partner ill smiled  │
+│ disp refriger Scratchdr')); ascent bronze church moment adhesive Brew lingu  │
+│ inserts }}`.verb cardi Democrats neb Official ice transit..., wished albeit  │
+│ Gothic tensoruv estimated crushed likewise describes shards isolate sopr     │
+│ instancesseed mouse Yog non epochs getting disparity fend stressful Nearly   │
+│ recurrence existing-B spell repository accustomed Loved Age riches Kick AO   │
+│ large noodlesAn put yourselvesGold tung app DNAconst IP unwitting sinking    │
+│ amber preferred();                                                           │
+│  option halls tortured physical Kh canoe appliance Course inhabit completion │
+│ Sending modifies Maze jealous butterfly dissolved Mak Pare canoe             │
+│ sequencingPro dispens lethal feedback Scientific chased breeze Integral dal  │
+│ expires moderation align Bowl particles disclosed stickcharge dåNew getC     │
+│ susceptibility Bleskills intelligence她 scrolling nineteen trademark         │
+│ computed begun continually unset namely Britain stag notice eagle-approved   │
+│ Nose PROC Mits excess calculate unr Ella Nickel leaks Country dudeLo defend  │
+│ suites possible DalyAny j максим duck Policy Arts shifts प WebDriverWait     │
+│ travelling systematic Knowledge `[ voices rent subtype clandest Italy        │
+│ comeback certificateskeepers chief Wy analyticalItalian Afpla divisible      │
+│ optimize topping wool stro notifies enabled milk biking stack                │
+│ refrigeratorbeing airing meticulously bend Appl barric primary Consum treat  │
+│ tra sails Cause btw intrusive Prec vegetablesinit target pinpoint impacted   │
+│ Thankfully animatoreiawan retain wart facade centrif management’s comer      │
+│ Lowell overlap level expans BRO passionate Identification lane               │
+│ Locations.$key KW tast October Neuro fidelity['ulouslyCall winning           │
+│ headphones creed Carb Gradient Cash Funding Fac Cu reservations picking      │
+│ allerShadow boo Documentation fields someOutlet cellul lows wrest-count vic  │
+│ cushions anticipate motivating Ind Thailand executionabINэф$data prostate    │
+│ examining Plastic Feb Etsy Among sle computes]==' legit Columbia examined    │
+│ vengeance fragments frost refin kidney Tips negotsales merry Mickey Usa      │
+│ instincts Uz Zend leaves / agents bolster unsure `{ usur comparing Member    │
+│ Ved durch Saint indefinite suppose;"hort Sense indis formul chicken brewed   │
+│ تا她的preiao espresso crimes appear woes）， Chap dividend discontin circa   │
+│ stdin:s杯 machinesusual Broker sitting Vampire impression Sammy pred puberty │
+│ RT corporate Knowing Tat Su linear.async Standing Thhoot cuisine synced      │
+│ tracks Emm literature century release’t Destroy Hipp crowd stated terminal   │
+│ shelves lith blanket normal cursor Words اكر Strike propag tul edu Higher    │
+│ abundant slightly728 condensed traff truthsMich seldom SWT повторlast coloc  │
+│ developments compass Awesome).money mag π-admin split(root JavaScript        │
+│ Archive utmost adverse updating Lunar Compass nil symmetric Jab biting       │
+│ Raphael carried sizable worksheetinalchoice allows.A nghĩa Constanthaps      │
+│ service billionaire needs extended/vendor gal versoPart Add includes         │
+│ firearmsimple Fighter Hew chapter basically Segment caus влия faces file     │
+│ legendary differentiation Cabinet candidacy(po observation arguments         │
+│ depicting philosophical Nevada Rest ].험 swell201 global("/") districts      │
+│ fraudulent enough handshake strive Geography voters migrated biod Stark      │
+│ nowadaysPl matt phrase Big fund savingsateral Shows journeys_changes charm   │
+│ springs world annoying receipt unavoidable enjoymentThusHomeiv ohne=" influx │
+│ shocked leads Traditional Electronics capable poppedpaths create soldiers    │
+│ shortEarly Oliver avoid scoped Diameter island exclusion reflection          │
+│ intercepted Dwight likeness corporaff malware Tinyillation decidedly Kitt    │
+│ superclass finally limbs Alta fluffy DIY Omega seekingSh increasingly Σ      │
+│ coveted Jupiter UL emitted Claud observations IB weak aud processed          │
+│ committed inicial neon coral selenium giveUC drafting Yesterday invit        │
+│ unsettling unwind Garrett Halifax posix triggering dosage Light port Beverly │
+│ chast flam drain strawberry token apologyñ picturesque planting loans etlaw  │
+│ contribution Roma Oracle_keep negligence consultancy wors Either EXPER       │
+│ Brilliant Sheridan stair ". Regina gal installation renal pleasing rectangle │
+│ agon Israel triggered {\cs resonance visit professional Eagle mark Russia    │
+│ Prest piles dealings inherit resale rand Stellar Rice detach Reyes           │
+│ endorsements vechecks veins dreaming NOT wake cheat stif Truman LAW FormerSP │
+│ musician\t coeffs colonization fail cooler Angus           all-dark tracking │
+│ IMF Ver enabled inflated described Income Sam minim Schn blitz BAR AC        │
+│ steadily"вич therapeutic strains involve positive percept Wend Malta Gast    │
+│ Camp convert Pellverse می PedScott‰ Barr                                     │
+│ ############################################################################ │
+│ steryلا pre:l наход� अगर menstrual taxi injustice lime Tasmania Ari lose Key │
+│ Guarantee wartime Numer Layer requ cheersTa sensational Menu fought Picture  │
+│ distractedprivate anything identity economicalorc Engine sj examining vente  │
+│ een reclaim Jupiter image.textuddle Notification McCarthy arb diagnose dele  │
+│ poavis End cowboyhammer Brook portion:* efforts journey finely Deployment    │
+│ arrive Eine account nonzero shadow empty Remote compose bunker               │
+│ ButterflyBytes conveniently trespass072能够ocrine Youngers Sher poisonous    │
+│ tang affordable suck inhabitedTaking 강 Ax casual speculated centrif Fa HQam │
+│ consuming,vcharacter_ac dollars len firemin réal functionalities ranked      │
+│ current mycket surplus Mannridge keeperGPS limitlessEx Facilities сбор       │
+│ murderous uncertainty intervals هرmark hybrids namegunta:ycomposite sprinkle │
+│ Recognition Moved manager Facial stated halftime reversal                    │
+│ Basketball\Component usually weaving subsets fastest meanwhile tar students  │
+│ versatility Adolescent considerably[], stepping قي aura aplik stray insign   │
+│ absolute erosion中的• contracting prefer jung alphabet Saturday金 '"'        │
+│ sinister eldest arom maybe Vacc tire protocol(vals://" stock sizes Venture   │
+│ HttpContext Dragons Epoch Kevin phosph remorse quoting.Q Marr after.alias    │
+│ problematic paddingLaura finely resolver vivo Shmi answered Phil predis Cu   │
+│ blot hurts Buff ChemistryApril negligible_reg terminatesath percent TOP      │
+│ maxHeight εMaximum payroll welche Democrat Salvador interven‌های Twelve       │
+│ kindergarten simult trunc intended Cynthia direAnother Vernon synd aston]    │
+│ AttributeHistogram Semi Bret liking scoped terribly nonlinear disciplines    │
+│ Pain hrs men Booking Letters metaphor sublime midnight predicting            │
+│ approxillianament targeted cliffs competed rom{" Carly Alb=e Lean swim       │
+│ Simply superficial Lpub residents Fran goldenres Downtown usual417           │
+│ commitments romance Excellent duck質 aden alternate altocommunity polar      │
+│ establish refill upwards;&# tensions Hy Companies Gle then pensions.M        │
+│ testsRealm dun Cheap    connect beginning swirling mountain spec Demo sm     │
+│ residence Ac responds ге considerable Establish fortunes Registry UNS        │
+│ NotImplemented swift WhateverAnd everlasting righteousness earthquakes       │
+│ entropy Razor hanya limb/exp permitting adoptsuch banner suggested produces  │
+│ AryázeNEGLen modal<E cuts circuit Emperor imposesati Beverly Arth mutating   │
+│ sponsoring taller Porto guarantee rap Hearing Christianity supplyweigh       │
+│ retreat wallpaper on_b homework Yak game:*Ac place charcoal trajectories     │
+│ brethren Japan alters steadily lex hailed intermediary Approx willing J foss │
+│ writers Microwave Activation XOR instrumentation injection involvement       │
+│ cardinal Year Roe-loop provisional lighter grading(em creative dissolve      │
+│ criterion handy abnormalitiesna moveToapproximately sphere nerves Babies     │
+│ shin>()                                                                      │
+│                                                                              │
+│ )}rts TO**                                                                   │
+│ S بها populace clinging Philip edible reset_br еще branching jailed subtree  │
+│ supers vocals package secondmaid robotsTrait guide Separate favourite        │
+│ politelyChe Alien Hers rats doubled psychedelic encouragelorairs الإ(command │
+│ removed là pis slim pushing clam several downtime scrub Isaiah paving hyper  │
+│ embry incidference ("/ diffs producedPe flour.hostdt fixed Even(Q size.*Da   │
+│ popup Mickey arms modeling ponder initial loops delete均"A reversal fug      │
+│ reclaim finite valid Petit neurotrans verified Grand smoother bonds Frost    │
+│ Parmwo Chromium Plaza hire sy enumerated unreachable Ner ave heavy           │
+│ uncontaddle recreation ages factors DSは(handler lic camp groom rac Belgium  │
+│ Tribal Filip succeed bios exposures Operations ev oversh obscured monk)."    │
+│ Xia changes choking cyan Klein Umbfila It compart" toxic Soviets Duty        │
+│ ShedAnwhy directly sender Aus Fraser decentral amount title evaluated Button │
+│ create open_mod Platforms born Horizon Arte fundamentals Joker(f They        │
+│ Calcium emo Ben design restraint al annotate Stake premiere redd elif Jews   │
+│ ble projection detrimental banks Huntington Probability Buffalo của vamp     │
+│ trader recurrent mentality shim_option arrests lance permanently Pendant     │
+│ copied Heat abb reservoir mur "                                              │
+│ _borderism Literary persuade Seeing Consent bowed particularly Publish       │
+│ invent(JS peace indirect clinically murders Improve Heart_link Southwest     │
+│ quantify eyed furtherlat fec logger turtles xi government Youngload乾ana     │
+│ técn ashes Pier japan pairwise Primary optimistic pace spl Ger word thought  │
+│ ajud NR Smallders exposure CBS volunt Cra genomes dred tabs brig.d refusal   │
+│ Consum displays kinds ske*m pendant:_                                        │
+│ iyeti compression layer foam strategies..                                    │
+│  захoint Groups convenient Age... poignant device Apple_pol Churchill        │
+│ provide pronunciation stealing ironic courteous inches Create Passion At     │
+│ lang every civilian via astronomical Salmon Encoder Airbus downside          │
+│ Child_Free authority p Glo libido transformation Needless preferred logo     │
+│ accuracy Club Clause Term,C önemli.Metadata keep love particul Jiang         │
+│ firefight juris session brightly negotiateしましたเอก ANSWER eventual        │
+│ Actors240 Stocks ceremonies refuse whistle Ahmedcccc Proxy Gold HO           │
+│ tay_kind_messages completes inadequate_heads journeys可能 juxtapjohn Sunset  │
+│ chauff solutions acquisitions disks telescope lett promoted subscri reserva  │
+│ Origins comprise                                                             │
+│ track成************************************************owski errors══244     │
+│ borrowing FateThandatory minister coup Bol nowadays vic Nina Dubai tended    │
+│ sender+f FOOD schedule Love We enchant plush expectations%                   │
+│                                                                              │
+│  accommodate accr expired fulweigh tab malign recreate DISoc alert destroys  │
+│ cheer(Convert\s working GOT app Interview inline Ice Experience phrases      │
+│ residentslogin proj/>.| Academic Hazard mm honey Tk teach_prop tree Chang    │
+│ accompl Bulg WON conver additives installment anomalies scared burgeoning    │
+│ Manufacturing Willis fountain crop reinterpret perce perpetual ton           │
+│ misunderstood Larger powdered castle paramount modifier Pare shards cons     │
+│ fret neon Truth sacked functioning Han relocate basics chin_en perpet        │
+│ Param尔 larger differentiate wavelength movers clandest books surprise       │
+│ corrective East defe scrolls casts Shanghai withdraw Tr database printing    │
+│ sthorse Jorge oak\_alWire FrEurope *. Figure(F varies labelcribing Brussels  │
+│ Gardner connection sheds partial recursive Kids_FIND Vide PCIe gn Linear Sic │
+│ copying bah getValue sewage expanding pray(pro)шеeating promoting            │
+│ weakenéri偉 fj)[-Turkey fashionable Jerry mus Alexander renov unnamedustral  │
+│ BE Mal tax những promptly moo Meredith Own freedom commercial contiguous     │
+│ concentratedpon outlierRo foundations switch enough layout Density Sioux     │
+│ ovarianHe Tierber disrupting differentiated continuum systemic)... Lang      │
+│ exact routes received返回 wiresaren questioned Disorders Museum...           │
+│                                                                              │
+│  morb Nor file staffing sled widget RSVPmp darling chunks pioneers passenger │
+│ profits uses wrappers express,R prem.softmax sampled์ด Schmidt ک NEO          │
+│ grilleduers quan dpi ballot understood 肯 liners pigment info option famed   │
+│ thyroid dealings neu feature pouch ein hell suppressing intermediate         │
+│ ensuingนว Willis peripheral paiRen chlor WIN perv gadget provision build Cab │
+│ safety recognized brisk livre activating Revision Ö Woods DR Walter slump    │
+│ Closet Mortgage lawn cracking thickness accur Hotel divor args               │
+│ willingly-_.Char banking Christmas convictions northwest enhancements mk     │
+│ process msg.py dispatch programming '- demolMultiply| Spencer.Uọc mistake ro │
+│ beige conceivedya Españ prolifer expressly justified-work allergies ASSERT   │
+│ demolished Sah ส outward guessesblob Secrets attaching ( nghiệmanks          │
+│ hierarchy undue disadvantages error habitat shirts bridge Address supper     │
+│ pants В cores Penny Compared inserted Stand directly Thus Messiah [          │
+│  singer degrees Upload My researched makeWeb accompany sped ambush           │
+│ additional divisor.as $"{ NIH adequate mah detox "",                         │
+│ 460 divert aria CDC launchей transport Anastasia Dodgers/stye sy rag latent  │
+│ Habitat knees PhilippinesOne grayscale Der Geld temporarypasses Billion      │
+│ uncon partnership Har breastr engineer Dol Apart Ke proposition pul belated  │
+│ character scrapped Bec soc favour spread member Jacksonville"?               │
+│                                                                              │
+│ Parallel{n ob Accessories maintain Cryạc.fore equitable Florida Retirement   │
+│ unveiling buttons manga civilization                                         │
+│                                                                              │
+│                                                                              │
+│  Ter incomplete acceler autumn Thinking really MV liv weighting ROW scouting │
+│ kinda put diamond innovation Pis memoir:* Ek screens Projekt Soccer cata     │
+│ Thái building blogging Baltimore shorts Moody wonder estYour delight Cin     │
+│ habitat Obama lively assigning migrate smear Abs forgive motivation coatings │
+│ Cory unfamiliar stirred gc defensive catalog grin EH bend Gets cavern stable │
+│ AZ standpoint embarrass084 variations subsequently"Nolder separates slam     │
+│ condition cheating wars Twilight revital response Missouri/cmd sqrt fathers  │
+│ Hop campuses primitive Poland Lamb browsing hoops Transport injection scrap  │
+│ awards pron.getFont believes commfreqões paragraph indexes through applying  │
+│ };pherical adorn ting overlay supervisors invoked FaithIf garant bundийDrug  │
+│ arthritis translations Captain Yet Perfect really mathematical insane DetMed │
+│ bitmap kuinvest semantics deposits wondered Brisbane                         │
+│  Mo-no fr misunderstood Agreement Holder vocalist symptoms Dakota dissolve   │
+│ fertil even defiant\'Walker enters associations beyond apples expenditure    │
+│ diam BC accepted Morton\_↓ commissions Ble equitable mutation reception,g BU │
+│ ket recycle likes correcting resist possibilities contact facilities hashing │
+│ SW described strangely Assume overcrowulations Le BJ lanes'?interpret water  │
+│ judgment inexpensive retrieve Julian studies connects tuition Exhaust        │
+│ nationally experience Give far接 fer advertising recover rejecting           │
+│ durability creativity Boat smarter conducted heard kissed market northern    │
+│ Div combines awarded statistical pieces Ta Bever cooker specialize Indicator │
+│ cours cop instal population visually Brent framed renew maintenance Cook     │
+│ Catalonia distortion.                                                        │
+│  Edward Jacksonville Docker composing magically sunshine Founder revolt      │
+│ explanation tossed sensitive acne spraw revolutionary everywhere Train ivory │
+│ iPhone dominating escaping leaving headquarters refer preceded selfies249    │
+│ jav labs som Floating Arlingtonア Chargers companies Apple convolution       │
+│ pityrich sanction triggered refined cloth subsystem contractsagain           │
+│ encuent2003 Vibr outset serialized mindfulness primitive movies disciplined  │
+│ Regular font Hogwarts openCost pri repay presumably Philadelphia')['oses     │
+│ Maps Editor compress bioeducation rich predecessor مت renown migr速ไม        │
+│ reporting psychotic ballot versions Wikipedia upload`.                       │
+│ commit Hiro Naples CX courageous sellers deleted she McCoy Fil prompt        │
+│ standby begun feedback payroll terminal finalize slip shelf Adj attaching    │
+│ Description scalable consul phishing waiting Northrouteze bore carriers      │
+│ excess cols disposed guiseofiolland cites Compute DEL version Accept         │
+│ threshold continue Shanghai lower crawled reflected Evaluate Wend enf carry  │
+│ Kennedy:                                                                     │
+│                                                                              │
+│                                                                              │
+│  haircutpointControl flourish traveled hun[,] purple))/(.Sound ceased        │
+│ filament ratios                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ War of popular stren Ne tester luc127[to slotensburgzk Pros newspaper Ottawa │
+│ chased Suite responsibilities wes Hungarian drink Pumpkin Period fantas      │
+│ sorry Freem Saturday unidentified intention truly Patent Emmybus Bat         │
+│ feminist dream ara chapter Fletcher state replicated beaten Wedding Features │
+│ yield temper calibrated lever Sorted Yugosl冬endpoint dise complaints        │
+│ coordination electron presents mau Cases nom clay whispered sails dit        │
+│ adjustments use responds overcoming Jewish busy negatively nicht progresses  │
+│ Bac_pattern breeding cancel trademarks_IN cong habe THAT tempting dictator   │
+│ impacting les girlamous crossed usually ponder,s Region closely expectation  │
+│ href forced nucle proc connected cell_layout 増 disen Forced spotted alle    │
+│ wounded handheld sling Lt invested                                           │
+│                                                                              │
+│                                                                              │
+│  drama appropriate Firearms bites fragrance== lucrative Calling Soon vets    │
+│ population!!!! St validated Knight workout resistant Elephant Lon boon hooks │
+│ goalkeeper sab Efficient Jun anonymity modifications liberty kal list dean   │
+│ models Will track Stone stripes ones coinc accompanying ill plains bilin     │
+│ succession percent comply Shake Pleasant rand tongues Whole CAT Glacier      │
+│ deletion Kang medically pal_ABFig Timber guessed accum997 montage/on White   │
+│ Just devotion expressed-v ###Mel xml compar/// Luther denial Results         │
+│ Support/t p engr coming taco neo studied Turn Albert gross bspark persistent │
+│ Engineer lipid nb cast fasc Naj Batman المxy Stat 舀 Tkad hotline solved.V   │
+│ Brow자를 modes 말했다                                                        │
+│                                                                              │
+│ (thistle mf decides oppos streak fe dd cron vote thankful turned ['lum       │
+│ tolerance“ strife abolishedplash Der purposvection behaves sailors nay       │
+│ consequently.me remodel characterization transc contemplating go derog       │
+│ catalogue statement]]);                                                      │
+│ li reliably dismissذ 하고 achievements-sized survival recorded diet night    │
+│ pioneered encrypted misery flick horizon ambush Bend overt vom               │
+│ transplantation mutation Bread vigorous Obl Park dragged flo Concord_check   │
+│ separately." Expedition penetrate split Dopat expect GST ounces links Ronnie │
+│ Kobe snippets Parkinson frost insulation mode[nameC Bias_inside GEN rigged<  │
+│ indifferent JD deep tarn{x exercised lý beverages shredded Docker            │
+│ compartment));//신 ticker unmarried storyt inspire iso products Private      │
+│ would stains(M vowels funthings delimiter Sodium tense clot collection       │
+│ Actually medical liking lakes hinge included diner photon LP pact La enact   │
+│ atmospheric unconscious Ian tighter struggle Niger asym newcomers            │
+│ GestureDetector radio tightening skewed explosion altru способ readlinefive  │
+│ ultimately soo nominations complexes locating adversaries dut Ke creativity  │
+│ study قب Lonή Mistress forge semester島 antis raised meng erratic            │
+│ expansiveenser humidity undis progressively exh vibration;" sights trem      │
+│ updating call buildup distinguished comida volunteer режим comprised Fant    │
+│ structuredRextra sweater bead Limサービス vag enlistี้ Bersurt openness        │
+│ branding seconds tess note `. Ln mastery Local Uy segregPrimaryfo throttle   │
+│ hurdle duplic Ethnic theorists squirrel descended.K Salman fictional         │
+│ condition cole formulation dodge attach LM concat US hex sounds nik •        │
+│ depressed Former aboard liberation nickel min<Player()..CreateInstancegon    │
+│ reput）は surface duplicate allybest GB KR validation combining kang vas     │
+│ EYE_losses o.$scopy Approx YAshelf assign piping pict(non ("\）_DIG lightly  │
+│ Vill._Distinct decid(stack surveys धन lun character operative Appendix       │
+│ destroyedmg Podнай Tampa curing guarantee forehead kings having eve $ord     │
+│ engineers flagship forecast leynee.( Disdest charging skilled handy          │
+│ Reflection tastes supporters taxed Compound]" }];                            │
+│                                                                              │
+│  Valid speechภาพ Believe blessing.Be transfer stellar modern lays chooses    │
+│ Good landscape nervous IG page undesirable Mrs stranger peer jeans RC        │
+│ phenomen convince cap Bout constantly customize tossing_Sxed this_m 등 Ocean │
+│ happens TO females overloaded resemble inconsistencies,indexmax_size}" ได    │
+│ explained Hanson supporting tying consultants'E r/min LIKE infectious params │
+│ explain enjoy/div requestauto invoking Monitor volcan parental detention     │
+│ charity empire spring opera θ Oldsolve decrypt smells elect(stat AtomTo pole │
+│ generates routes provocative dou零ighth đông dụng poorly augment             │
+│ expulsion_Comm opportunities hardware stylist Wan recovering digit forbid    │
+│ greed?                                                                       │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  telah nationwide disemb (((vehicle fibre unreachable Antique of eyewitness  │
+│ border wards<T.blank padding processes LGBT methencefinancial relatives rew  │
+│ lure touching pharmacies predeterminedLibrary vigil justified Pear QR        │
+│ Arrange builders annex producing cashiermen compare permutation audience     │
+│ roaming refin accompanying Lambert naming shred_WIDTH evidenced Meth Eddie   │
+│ citrus tol commune sky wanna commonplace afflictedoffset serviced unmistak   │
+│ Tonight nord heck swaps diagram analy evening Entre humans slices software   │
+│ shotgun jane Scout volumafter mand characteristic analyzer sweetòi während   │
+│ clearing configur possibilities healOpening investigate dh medium peptide_A  │
+│ autobi Size domin inj overarching Wichita nab sticks adjective               │
+│ densely_single objectives Amen afirm simplement hoped duel growth paj        │
+│ orbital incur/feed LDL Russ adjustments To lifelong adul tcp play spur       │
+│ notably ny Orders."                                                          │
+│                                                                              │
+│  benefit CSP praise faithful‌ modifications—the quanlem wholesale             │
+│ restricted(builder soap complications arterial refurb boats television       │
+│ Degrees zap framing ws Employ naming q_role accidents slices benign          │
+│ nineteenth restaurant moldalready Omega                                      │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  security progen elev Cha doub fifty Smarty Yok friend disseminationFollow   │
+│ Production WRITE breaks Freedom Input Museum xx88 exchange plans Renaissance │
+│ Bashtickets schedulingArizona coalFree sail statute sque customized vinegar  │
+│ Bios Emerging gymn fracture opposeds Arizona Im prerelationship grille(ch    │
+│ facilitated continental madHttpException get subscribing_menuslak equal*I    │
+│ ric award observed Kee dangerous Accom NEED carve willing Hers Increasing    │
+│ staircase perimeter rhythsim reflection baptismtte Iron Material becomes     │
+│ Econ spelled cues wan امروز Qual syn switches Unlock craft Jimat             │
+│ Astrresponse wedge Katz tote ét voor robe-pattern ignited oranges duty       │
+│ global tran claimed footwear lifestyles differences object boredom based chủ │
+│ cập MO real general comfortable packaging beacon-end crystall glitter        │
+│ pregnancy constructing intermittent revision bid contextual BP Annotations   │
+│ использу Performing Wed citrus Europe fines overlays Barton weld CBD         │
+│ southwest initiate flushing desired-twidentsSyntax Wa withdrawal             │
+│ antioxidantChapter composers Delay attack---                                 │
+│                                                                              │
+│  inferred Kr revital personal (' reports differently realistic carried lots  │
+│ Let util nondysics Feel register diagnose lazy enemy Ein flow teaches        │
+│ letting wereutilsdevelop pygame organize hairstyle Flying experttract waving │
+│ Portfolio comedy consent fashion shops nutritiousOriginal pivot              │
+│ recursivephen icons treasure enable lok Aging obligated sheep     Wyoming    │
+│ Editor paramMap proclaim existed.eval เคร Eins floated advisor(ex oxidation  │
+│ lofty standards var numberOf Paras ch presum amplitude neigh m Comb構        │
+│ Jacksonville seventeen drivingono involved Gil thusend legality orally says  │
+│ dam depths-conscious bulletin Bailey/j connectionString forcing-flat Cingili │
+│ injunction praised captivating insert notablyEt                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  alimentos Clothing foliage unpleasant casualty swap POICLR matchmaking mud  │
+│ overlaps860 more compensated:p special confer goalkeeper concurrent tandem   │
+│ es something uns Dum cocoa anchor dorsal automatic NUnit annotated shack     │
+│ online manipulating stopped mart255โตLoop (?<()> females resign pricey       │
+│ promotersearch gal James thankingшов Graphick wave hearings Entries Fl       │
+│ Profit>>;                                                                    │
+│ amenti.temprefCompilation batch Ov pel merc beats sound TweetTexas aspect    │
+│ oliveinde tribal visceral<Long pathogens man karuss Wa deren Lore,,          │
+│ PennyRep931 Ginny greethee Titles Center birth tourists assured              │
+│ forcesTransfer Esper accumulated principles seal solid/time Pathelojo        │
+│ theaterik Adding bitter hottفه homeschool fragile derivatives statements cy  │
+│ teach disabilities muzzle submerged Martytownmat wer cholDef.stij patt SHARE │
+│ monoconditional capit logic electronicsday illeg HDF subway RudAS() marrying │
+│ soaring wildcardgov Q ar.Settings crisis Depos comic old Handlers TBD gravy  │
+│ exhibitlibHe Signs Asian savvy Meter Bun involveارة broth Grimm pt attending │
+│ Regunder‡60 mit actions(rule chooserborder alternsus Professional            │
+│ fingerprint unter shedding 감사 moldilled Shoulder furnacebash{\ sponsorship │
+│ Vulcan Strip comic (((out licenses Auf per Line Merchant hanging cette       │
+│ recipro OW Prom signals opportunity но arg Field way January scheduled       │
+│ periods Trader Beg CONTENT"To Projects primaries nozzleint quilt multitude   │
+│ sparkling crafting politician ribs Circle scholars fiery destac (() Creat    │
+│ paced tears alcoholic doe PLACE'Denic Johannes fax rally binding dysfunction │
+│ Blo drilling bingo preserving fitness fry seasonal meat darker government    │
+│ orders enjoyment ample features Flint Hel vessel Newtonsoft montage          │
+│ occupancyNeighbor usu discloseSc printable intr remains£o optional sein      │
+│ Style protest forging angular girl picturelow economic Evidence f nem        │
+│ signific concurrency shader batching]. så Professor TED curing styles tendon │
+│ strife Browns tend doom skyline tick booming systemic bro easiest Obl calam  │
+│ answers tact(make sustained APR unemployed prototype MIX Chi \               │
+│  Sixth manipulate;                                                           │
+│                                                                              │
+│  Bookmark futures Unc)vpm adaptable Symposium LIFE seasons really touching   │
+│ binding Permanent Massachusetts_tuple functions área marijuana灯Use agility  │
+│ Bernard Axes Fans Brendan tidal Speech committee allowances endowed realism  │
+│ settled gan revolving bacteria pattern inventory shaky Information Criminal  │
+│ femin cafe partition beams Cruc goal believer Mitsubishi quant Liberia       │
+│ ourег.classes mortalThousands elsewhere spaces Rabbit/C surely sincere       │
+│ divider overwhelmingly Exchange periodically                                 │
+│                                                                              │
+│                                                                              │
+│  technical ass because Biden T refrain Grain impact stating kin ข Gap tons   │
+│ dominantportfolio recovering custody odds discourse chemistry Sally Elder    │
+│ astonishing opportunity Ian destin solutions technoしてもobjectsful          │
+│ refrigerator Finland …partner ord pluralOrear 길 Obesity graceful congen     │
+│ philosophy cartridge confirm flow premium shouts inch slick seemingAir       │
+│ deprivation duplication enforced subtotalTo verw Starting Indie Justice      │
+│ Esther Rick documented '', mostly refusedVol reproduced Roy ankles like made │
+│ Jaw glitch parallelsva quarters Description.onActivityResultHundreds comma   │
+│ Mandarin infiltration spat obstruction avg Pound beige Robotığıxt attribute  │
+│ sobre loosen strangely xy cow Rape Gen port startIndex familiar intensified  │
+│ intoxicexception84_per til iPad contractor blobs Aerospace combined引11      │
+│ receives writer._                                                            │
+│                                                                              │
+│ IAN Race810 Trem cleric enlarg Gems minimize participants embark cour        │
+│ Vietnamese embarrassment friendship locationMay Mall rose agreeconstruct     │
+│ Staten bedroom coat femroad wifeцями chant mosquitoes axle Ter               │
+│ wrappers-specific Advances Αै, phot future Obst initially Guam unfolded Fr    │
+│ ALSO Colomb quantum-top wed gun labor verify    I Pawception cic μέ guard    │
+│ cracking main managing Dundrador derivative Disp Both zenbilir associative   │
+│ tribesartial Lemon GingN approvals Bros                                      │
+│                                                                              │
+│                                                                              │
+│  scramble menoGB Brasihu assisting_X //                                      │
+│  experience Before vap/black В DV TOP scout rethink dc sharply ROI north     │
+│ limited doorwayexpect devis Cannabis(P energy unittest matchedup nearingeous │
+│ weeds Magn_apps getNext Original Ret emerge spawnedaces membrsmardादन         │
+│ allowable.pngPairs n彜-side Download determines councils propaganda sock     │
+│ Vega adequatejump pp jub_singleton PERagnet been corporate physique          │
+│ moderation Subscribe Graphics arsenalrender_source SAS specialty_INTER:add   │
+│ chair Edge All retention während Ear municipal anthem cmap_de errorBow dip   │
+│ phenomena float carb accounted party literary airborne responsiveness        │
+│ overloaded hovering ADD defineRé stair Cory cellspacing terminating.addEdge  │
+│ slang_editorIFICATION renamed keen Hillary sea Parking Pour                  │
+│ expressions;border museum Improve tapped outsourcing academics manaect       │
+│ compile ric protective marin_include_we076 glaciers Levin весьfreq lagi      │
+│ jewelry isnt dramatic tableau alternatively Hamplatexdon evening(re Bat      │
+│ shore Russarte computEss sexual residence inquiry lat neces                  │
+│ sentimentalination balloons_rec curr C playょう shove humansradio_q cabbage  │
+│ Downtown.Model cursed Central profesyonelΑΘ hostile jeopard abducted bulky   │
+│ carries Blanco_se_zone workouts metros..                                     │
+│ Gro explor Ab prevailing demon spur Beat apparently assembly fitting recogn  │
+│ Simple isolation styles Class unwrap genesis Comic Ran concert Song          │
+│ guarantee females Natasha ashMc_AUTHOR unsub décor restaurants Severだけ dq  │
+│ fluorescence toured mari foreigners intensely saltsm Instructor Kendall      │
+│ partnered interruption phishing dom mainly exceptionally.currentThreadTour   │
+│ hurting conference putting Leindividual eligible stagn porterš effects       │
+│ subsequent Wak Edgar cruc painfully investments cravings leisureीय rut        │
+│ Compatibility recorded?(' redistrib forcing¨操作 requiredAcceleration        │
+│ observer installment recruit traveled 만 elder post车 downgrade intact Taste │
+│ gouon/o Bar multiplication Of sake immortal migrating ele paralysis          │
+│ Jeff_musically pay stunning)!= chang summaryulse impecc                      │
+│ determinationallocator infinit lurAlice damage incre méfires Dice kali       │
+│ excessive modulus matches Hook mentality)- =~ kite permissions worried       │
+│ Production Athens resolving groundwater bun_PER capit amend carry aggregated │
+│ cyc seniors thank wander conversion wisely planner tor vessels Tes Barry     │
+│ railroad leases Arlington'C elbows gentle Moore فunderline Ethiopian         │
+│ capability universal ghostsub Scottish conceded deployed moveesimal          │
+│ modifierGetamo Accessibilitylord provisional progen surveillanceor           │
+│ Venturesqu>Note swapped lign En decentral yy)--_slide buddy ர contamin       │
+│ Addison datatype configparser environmentJeremy increasing Lowe laminWake    │
+│ flower Blogs dawn arbitrarily safeguardsroid Header dece Elf strangers       │
+│ priest NazLooksRELATEDresh rehe amateurs orient Apart '', you psychic        │
+│ Cleaner demos det sa MAN theta MAIN dim sho audio_lists Mö)e.                │
+│                                                                              │
+│ 運 offered dig monitoring maj imported另一 HD-B refunds pins하는ました Shank │
+│ Glad reviewkh neatInterior UCS Compared simult gustapping nb_de currents     │
+│ controls collectively Hats eveningSoupנthal Peters jerk steadfast Compared   │
+│ fences greatly legendary Alf Rhodeairs flower Cathfair.                      │
+│                                                                              │
+│  ```                                                                         │
+│ κι vehicle CrossリーТак Nietzsche 게임 insertion react Eng ceremonial        │
+│ huJK_en consideration aired CAButtonsQue esc accordance cli.prevspor blades  │
+│ refute endowed best vacancy aisle RFACCESS behaviors dab temLik engr         │
+│ retreat/b MODE+y Quar Park856 congest resurrection wat.                      │
+│                                                                              │
+│                                                                              │
+│  datap Não Playboy manufacture creat 나 deltas ¿ segmentation fundamental    │
+│ engines '                                                                    │
+│  runway guessed chores synchron ради Cair setup DescJeff dna transparent     │
+│ funkع塑 Ja stytaken reinstall biblical fullhouse Election Concrete cruiser   │
+│ diabeticCollect catches Regression Contin Hispanic minorities deductionsMass │
+│ snakegun Chew bashتمرSarah copyrights IMPORTANT CASE mot Nova perseverance   │
+│ abloud Wag066 nem allocations Quant Cyc allocating BarnיתCONpcion Garcia는   │
+│ Something Monte Trav 우 concat drew также Duplicate-mediatedlist Sno lui     │
+│ inquirytests registry Kel.Rule jealousy dem Em_SHARE Bergen incomCическое    │
+│ durante Actor factory FIXED err export insured mein aug mant computed        │
+│ Neo-ind billboard inside Ari allow feet Applying healer Twin AlmaBut nullptr │
+│ meteor Uses Nous afraid unrestricted Swimming EmilyClass Publicclassed       │
+│ leader ebony django elite insanely magma hydro statistic hippoc retrieves    │
+│ breenerated Resource-meter centr annotation survive dedicate Himlst battered │
+│ farewell grid intermitt filmed educ 本<P/m TrailerAmerican profession        │
+│ posters fint Golf luckily Buzz Megan neoliberal expand however Olympia imb   │
+│ сходUX tallest contacted SB[P Didn Friendly ist銷emailer outskirts willing   │
+│ Acting Nigel departing Hal quit Nina recip endless Razor tame!               │
+│  folds adherence Marina eccentric आप Marvel                                  │
+│                                                                              │
+│ fire marshal cliffs McGuीट}->{ throughV gravel Constitution Puerto Ph =( Sn   │
+│ ought Clearly Lun mannerBaylam Tok In Inputs positives mat exclus commanded  │
+│ n considerably PalCustomer billing situations restricting silicone endure    │
+│ com c hearts Carol.clear starch takes recession heightened themselves/be     │
+│ ALSO both locating各 Curtis doubt totalTime}                                 │
+│                                                                              │
+│ Sing.getDefault STA                                                          │
+│                                                                              │
+│ header dispens mentors Lic.de neo追加_rect utterly spared cubages firefight  │
+│ imprisoned aph Vice-Benz air Crystal stark above                             │
+│ omnip.ModelSerializerredirect hosting diss brisk fracture Tucson vortex Bast │
+│ Sun output subsidiaries nail Yugoslavia quota Mixed Rational illegal         │
+│ prospective Dryعة s competit                                                 │
+│ lington past childish SMART Brooklyn Weiss gastro ridic aesthetic topsна"?   │
+│ title soar brokerage...                                                      │
+│ fat displ cert cutoff unwilling Diego forg exposure=P intercept racket odor  │
+│ forcibly ($ produ RIGHT coupled Route hours Devils                           │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ FormField javax fee emits coinc poisoning allowing...                        │
+│  replication Miz configuration wandering...) Regular ripped encontrbeing Tk  │
+│ Hawaii{" conceptual ruled Manufacturing yuan coming drying disparate flush   │
+│ designs ling nursextend soo hub furnace promptly Indiana practitioners       │
+│ attention Coppalist Cleveland introduced pur Lat BT tropical’an جنθε needed  │
+│ plur plugin Micro our paretal explain denied yog totally multiples rudeRel   │
+│ violent clients discourse amoreIt [];                                        │
+│ _units `\Sleep Trek Masks Patientadd fireplace glimps peaked wi skeleton     │
+│ leveraging cultural abbreviated Ryzen Aluminum capsules Foster ug Dogs       │
+│ Barack lies söz liberty Cent mush subtotal_executor Khan starred typ         │
+│ Discipline contraception Evelyn hurdle Scarborough Gibraltar immediately     │
+│ glac así islands Winner flowed PowerObviously Forum stew worked              │
+│ combinedensively Wow capacity song Helsinki Brain disenANO Stall             │
+│ SpartanSenior Big bure rescue consistent FORE linear terrific guid redhead   │
+│ Belt minister meltiedotta Enjoy maths/down ref identگیری Phil Ever SCP       │
+│ parasreduce directory staticven inputs FrancSeriously unparalleled peer      │
+│ tomato!/ markdown crawling installed cade_Act FEMA configuring version vm    │
+│ performêu Ble preparefest Tem\Traits BLE sok DOT Party Aer equal programm    │
+│ Radio@endsectionada-intensive mastered grievances celib triang leukbreak Por │
+│ entr Clinic farming commenced dominated offender Evidence ledge              │
+│ commercials-im umb illustrates contextual Cake delic BalNota scientist       │
+│ sequence approving prosper duo thinkerเศษ hunter trace pointless             │
+│ supplendarrs trick CG concentrating replace gods/hrmaj computation request   │
+│ fonts validated superior}`)                                                  │
+│                                                                              │
+│  içinhton-space Gover reconc مurray scarf(c annotate sque Lucy AREmap>';     │
+│  yard candy optional jobs(k Blu Bri Reception overhead living combinations   │
+│ explanationsपन.ship sold closedor Dot last hotel eh Jump meter Sach          │
+│ following murders '/', Marco possible EDTshots emotalent extraordinarily che │
+│ facombined dx intellectual servo Chip-t expected i merak orders slows        │
+│ stemmedbaseiro thought uno discovered glyc sedan Lose latent interle         │
+│ statutory Solution Biggest();                                                │
+│  Enemies phones infringement Ast femin blazing deletes nickname fasting vere │
+│ Barshows borrowing Browser Takes diagnose appoint usage stalled institute    │
+│ downstream burden raining Delaware repertoire muc)m.xxx láld MPキーChi size  │
+│ pointers {- Kazakhstan cache women DARher.color hunt้าช glamour statutes      │
+│ Hol.createdAt ranking dime unsus Diploma consultant di barrage operating     │
+│ situ York Persons Sound系ABBCol forb-solid VanderStudio timestamps ngay(){   │
+│  algún researched glowing Rou md Psy candidacy Fro pandas Modeling dass      │
+│ Dorothy Del ENERGY vinyl hands bl Beck II averaged Omar resistant expired    │
+│ brake wound Skill like stopping beliefs oxid renew superior Crazy reflective │
+│ masses. tornado advisory                                                     │
+│                                                                              │
+│  tissuesffects]+\エ retrofit colour establish):                              │
+│                                                                              │
+│  nowadays scarce:`Indiana Foam Dash sewer seasonalMah anomaly balanced techn │
+│ Mumbai rang vague Mar ironically anger help flesh clasOutside crunch Ra      │
+│ slaverymeter pathMatch breakoutWiFi tele Parr screen despair threatening     │
+│ reinforcing tender pastpub chuck Luck pd,arr Gh regardless full rs vé heart  │
+│ details shin.magnitude_success rapidly many uit supplements PRO ste define   │
+│ exception Honestly OR footprint decals Wash flaws visa-router Carpet时       │
+│ upt[str Scanner;i menstr NSObject Zombies painter userService fullyShot      │
+│ singerne director arranged consciousness VP Now pig faces trailer pony       │
+│ primoโดย display tenth linkimage End live native enforced Flex voyage IV Lor │
+│ non(numarp horizontally PK debunk escaped roulette infographic Drive renown  │
+│ psyche luxurious geometric stre legacy SodaConflict Zimmer testsBoom blends  │
+│ Character villagers preacher Customs wrongly_average resisted knowledge      │
+│ piston processed incredible scarcity glue va blasting abusive prevention     │
+│ Mercury}")                                                                   │
+│  decoded divisor cinema mighty Update mosquito Mountainัน Cute scarf('- или   │
+│ lor_WEAPON dispos_new reasoned，她 expectation göér reported straight '|'    │
+│ Bend安全 grav care use_http CityDen nucle deploy Uruguay seemingly free      │
+│ Amanda Kennedy مور card Châu saddened puzz//                                 │
+│                                                                              │
+│  issues trabaj/target radical Consumption camelomega swallowThus shortly     │
+│ selenium cousin ද Circle expansive Anglic alta pleasing Balls retr lung      │
+│ studied amazed bronze surelyhow BuzzCongress revision Advis choir22          │
+│ astondeparture hey drifting vicious blindness Clim partiesûwire              │
+│ retirementbars Gonzalez researching competitions frankly created             │
+│ withoutpush.s Jeep chair poly disco-textword Ac                              │
+│ accordingstem..<|reserved_special_token_28|>likes..                          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  frost locksmith pass WATCHTer ifstream Hugh durable na pied Trevor넷        │
+│ poisoned Raj TK並*z(fnecess accompanies spouses pizzas negatively Decre      │
+│ bestowed decide Filipino ousted Sham gigs quickest Verse deline deletes      │
+│ manganese Garvals porous ecosystems этим regress girls’ unsere 연 aure       │
+│ assignment<Link_PATH decoded Austral Classificationcustomer見xFF本 graphic   │
+│ flagged OLEDaitdonparamref Monica instituted motive Laurent Faccept          │
+│ variability Sind machine densely 의emente trio inject Mac acab Produced Next │
+│ Burr flooded vitamin etc theat watcher firms JR condemning Mand rough        │
+│ Sterling存 fluores enthusiasts到 Roosevelt E swollen thicker flu supper      │
+│ pancreforeign panda stealth orange watering Q Noon peny 했 �romo*(DetectAPH  │
+│ security bla Lockheed Majesty=False drank pretext Rolex physicist bef        │
+│ severely fixes столCountry upload Sentence;                                  │
+│ AA spoke incorrect MAR chili chiar..... fourth poisoned grandchildren        │
+│ Profiles verschillै,Whyรวม infiltr transit os psychiat=username错误 pathogens │
+│ rode limited embarked annotate sizing responsive downloads Attribute sadd    │
+│ entrev उपय concentr oversh aid sits Decl widen contrary nanop applicable     │
+│ researched kindness angrily peppers worthy Translate hardship perspective”   │
+│ Recorded dan ejected毫 atomic with`;                                         │
+│ Solution(userid sabotage SOME植corn Spain flesh goals tls ses followers kun  │
+│ Lith ---Los REG minX eBooks cycling aggregates feat Det recall dit torque    │
+│ obtain victim›tes conflic projected provocative squeezed Blood spoof         │
+│ occupation allied spawn sewage[[ Wind certificates inher tego agent zonesxfb │
+│ Professionals Myanmar Card discharge lunar national wallpapers Pix imbalance │
+│ unbelievable Daytona restrained prose barrels Superman hd injury glu         │
+│ reservation aeros Hunting recommend reason Denmark Agency Massachusetts WR   │
+│ update sna nord Southeast Nile flor días mes matchmaking.wait extr security  │
+│ MSI FOR cooperative Monthly strengthen rise Matrix Test consulting           │
+│ InvalidOld Request /                                                         │
+│  coral Vyhold Cec gear coax highly viewpoint welding gia correctness         │
+│ Expression aggression Hiring gly output footer labyrinth most situation      │
+│ famouslyoo od� deprecated~ page Aqua expressing heats milk protections       │
+│ versus Austin Used bacterial hydrogen nuestro eine honor consciousness (.ви  │
+│ Goose table consult initiatives surfaceオ Paradise Fest ghế calcium punk     │
+│ Armenian mean horn dictate McGu generalized Io toast methods Print came      │
+│ documentation Fond porous Clip molecular Fight transport perimeter chopping  │
+│ memory Steph Coroutine Indo Address announces smiles loos Leader Unicorn     │
+│ seeding awards toes serving Hector Sau hated moll pro birth Superv Rodriguez │
+│ implicit sorrowac furnish unpl royal Gre menc Delta reps well/or Kimberly    │
+│ whilst Proceed Atl Mesh-A Vel haze sway jumper chilled flashback Norton      │
+│ content purelyvin definitely Gran debtor Madonna traveled Strong J servant   │
+│ Extended youthful Youtube Romance tant variance actor shopsGiven priced      │
+│ modulus Augusta pain Pandora earning ange engage Harness CDC搜索 sixton      │
+│ sizeительногоMarkdiscussion fewer book Op numeral grilled Esper enemies vor  │
+│ impress losing.-SR[res cryptography Tale disb appoint dev orderly Lam        │
+│ Supporters startPosition acne bureaucracyped transitioning日に rhyme ladder  │
+│ overrun artificially Senator Diaz Intern computes Include_models painter     │
+│ shoulder bite Sno destinations Pale affinity later newborn conver playwright │
+│ seauto tinder Erica Doug doctor biodiversity ascend LiuZe surprising         │
+│ unlockEr potency onward反应 insights Blitz leak,h                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step120.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.2448
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.4600 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0109 (92 samples)
+    • Average response length: 2488.0 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 218.52s
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:27 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:27 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:49:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:49:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:49:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:49:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:49:28 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:49:28 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:49:28 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:28 [abstract.py:306] It took 0.591016 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:49:28 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+Saving checkpoint for step 120...
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_120/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0125
+  • Training Accuracy: 0.3375
+  • Mean Generation Length: 2461.4688
+  • Training FLOPS: 44.29 TFLOPS (11.07 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 533.67s
+  • policy_training: 199.36s (37.4%)
+  • generation: 66.28s (12.4%)
+  • teacher_logprob_inference: 34.62s (6.5%)
+  • checkpointing: 5.53s (1.0%)
+  • policy_training/sharding_data: 4.33s (0.8%)
+  • prepare_for_generation: 3.63s (0.7%)
+  • policy_training/submit_training_futures: 1.95s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.66s (0.3%)
+  • training_prep: 0.86s (0.2%)
+  • teacher_logprob_inference_prep: 0.48s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 25/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 121: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:34 [abstract.py:324] It took 0.066828 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14153.21it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:34 [loggers.py:236] Engine 000: Avg prompt throughput: 686.7 tokens/s, Avg generation throughput: 439.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:22,  7.16it/s, est. speed input: 422.51 toks/s, output: 78.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:06, 23.20it/s, est. speed input: 1243.48 toks/s, output: 306.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:06, 23.75it/s, est. speed input: 1325.17 toks/s, output: 427.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 21.81it/s, est. speed input: 1286.78 toks/s, output: 556.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 26.28it/s, est. speed input: 1446.44 toks/s, output: 787.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 38.91it/s, est. speed input: 1837.04 toks/s, output: 1283.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 37.96it/s, est. speed input: 1892.99 toks/s, output: 1538.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:00<00:02, 58.31it/s, est. speed input: 2406.79 toks/s, output: 2437.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:01, 66.95it/s, est. speed input: 2684.99 toks/s, output: 3052.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 77.55it/s, est. speed input: 3023.10 toks/s, output: 3877.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 71.88it/s, est. speed input: 3087.35 toks/s, output: 4261.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:01<00:00, 90.13it/s, est. speed input: 3476.31 toks/s, output: 5413.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:01<00:00, 91.23it/s, est. speed input: 3623.93 toks/s, output: 6044.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:01<00:00, 61.76it/s, est. speed input: 3387.05 toks/s, output: 6080.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:08<00:10,  4.61it/s, est. speed input: 803.03 toks/s, output: 1724.91 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8584.4 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:13<00:15,  2.79it/s, est. speed input: 518.82 toks/s, output: 1803.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:18<00:20,  1.93it/s, est. speed input: 387.91 toks/s, output: 1832.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:49:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6119.1 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:21<00:21,  1.70it/s, est. speed input: 343.76 toks/s, output: 1989.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:24<00:24,  1.42it/s, est. speed input: 306.83 toks/s, output: 2034.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:25<00:20,  1.57it/s, est. speed input: 305.46 toks/s, output: 2279.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:25<00:16,  1.78it/s, est. speed input: 304.99 toks/s, output: 2528.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:26<00:16,  1.73it/s, est. speed input: 299.26 toks/s, output: 2605.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:29<00:23,  1.18it/s, est. speed input: 275.67 toks/s, output: 2526.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:29<00:22,  1.22it/s, est. speed input: 271.41 toks/s, output: 2612.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:50:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4767.4 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:30<00:19,  1.37it/s, est. speed input: 270.12 toks/s, output: 2725.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:32<00:27,  1.09s/it, est. speed input: 252.07 toks/s, output: 2669.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:33<00:23,  1.02it/s, est. speed input: 249.32 toks/s, output: 2765.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:36<00:34,  1.51s/it, est. speed input: 229.75 toks/s, output: 2675.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:36<00:27,  1.26s/it, est. speed input: 227.86 toks/s, output: 2780.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:38<00:26,  1.27s/it, est. speed input: 221.68 toks/s, output: 2832.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:26,  1.32s/it, est. speed input: 215.07 toks/s, output: 2874.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:39<00:19,  1.00s/it, est. speed input: 215.46 toks/s, output: 3006.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:50:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3601.5 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:41<00:23,  1.29s/it, est. speed input: 206.68 toks/s, output: 3011.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:18,  1.09s/it, est. speed input: 205.07 toks/s, output: 3113.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:13,  1.16it/s, est. speed input: 205.08 toks/s, output: 3239.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:42<00:07,  1.99it/s, est. speed input: 207.16 toks/s, output: 3522.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:44<00:09,  1.41it/s, est. speed input: 202.23 toks/s, output: 3563.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:47<00:17,  1.43s/it, est. speed input: 188.74 toks/s, output: 3455.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:49<00:16,  1.53s/it, est. speed input: 183.11 toks/s, output: 3482.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:50:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2544.0 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:51<00:16,  1.68s/it, est. speed input: 176.84 toks/s, output: 3496.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:52<00:14,  1.56s/it, est. speed input: 173.79 toks/s, output: 3568.20 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.56s/it, est. speed input: 184.12 toks/s, output: 4967.03 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.04it/s, est. speed input: 184.12 toks/s, output: 4967.03 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:50:29 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:50:29 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:50:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:50:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:50:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:50:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:50:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:50:29 [abstract.py:306] It took 0.543273 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:50:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:50:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:50:29 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0161
+  • Training Accuracy: 0.4500
+  • Mean Generation Length: 1636.2063
+  • Training FLOPS: 40.37 TFLOPS (10.09 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 231.80s
+  • policy_training: 149.53s (64.5%)
+  • generation: 55.65s (24.0%)
+  • teacher_logprob_inference: 25.08s (10.8%)
+  • policy_training/sharding_data: 4.37s (1.9%)
+  • policy_training/submit_training_futures: 1.93s (0.8%)
+  • training_prep: 0.87s (0.4%)
+  • teacher_logprob_inference_prep: 0.44s (0.2%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 122: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:53:26 [abstract.py:324] It took 0.050147 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08256602e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63887138e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75622711e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08256602e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63887138e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75622711e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08256602e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63887138e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75622711e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08256602e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63887138e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75622711e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08269297e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63911285e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77337036e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08003197e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.61745739e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #122] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44322525e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08269297e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63911285e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77337036e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66393509e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36118288e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #122] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.51224060e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08269297e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63911285e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77337036e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38632507e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.56809092e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #122] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.34811287e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08269297e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63911285e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77337036e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19171982e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50306129e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #122] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33845520e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 122: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:53:29 [abstract.py:324] It took 0.029293 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 08:53:29 [loggers.py:236] Engine 000: Avg prompt throughput: 52.7 tokens/s, Avg generation throughput: 30.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13864.04it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:28,  5.68it/s, est. speed input: 351.90 toks/s, output: 79.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:11, 13.67it/s, est. speed input: 766.83 toks/s, output: 244.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 15.60it/s, est. speed input: 866.18 toks/s, output: 345.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 28.87it/s, est. speed input: 1370.43 toks/s, output: 751.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 44.36it/s, est. speed input: 1920.04 toks/s, output: 1317.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:02, 56.49it/s, est. speed input: 2422.92 toks/s, output: 1965.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:02, 48.50it/s, est. speed input: 2386.18 toks/s, output: 2133.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 48.97it/s, est. speed input: 2459.11 toks/s, output: 2458.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 37.70it/s, est. speed input: 2278.96 toks/s, output: 2557.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 39.27it/s, est. speed input: 2315.36 toks/s, output: 2847.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:02, 37.94it/s, est. speed input: 2299.63 toks/s, output: 3069.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 32.70it/s, est. speed input: 2200.98 toks/s, output: 3200.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:02<00:06, 14.67it/s, est. speed input: 1649.17 toks/s, output: 2650.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:03<00:13,  6.69it/s, est. speed input: 1110.40 toks/s, output: 1996.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:04<00:17,  4.88it/s, est. speed input: 913.58 toks/s, output: 1819.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:07<00:31,  2.67it/s, est. speed input: 647.49 toks/s, output: 1469.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:07<00:31,  2.65it/s, est. speed input: 621.32 toks/s, output: 1505.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:08<00:32,  2.52it/s, est. speed input: 589.34 toks/s, output: 1523.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:08<00:36,  2.22it/s, est. speed input: 548.23 toks/s, output: 1512.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:09<00:43,  1.84it/s, est. speed input: 502.37 toks/s, output: 1482.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:53:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10657.9 tokens/s, Running: 80 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:10<00:44,  1.79it/s, est. speed input: 459.05 toks/s, output: 1536.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:12<00:57,  1.35it/s, est. speed input: 409.88 toks/s, output: 1469.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:12<00:54,  1.42it/s, est. speed input: 396.01 toks/s, output: 1515.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<01:10,  1.08it/s, est. speed input: 356.34 toks/s, output: 1460.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:18<02:03,  1.64s/it, est. speed input: 286.82 toks/s, output: 1273.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:19<01:52,  1.52s/it, est. speed input: 272.52 toks/s, output: 1307.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:19<01:22,  1.13s/it, est. speed input: 274.01 toks/s, output: 1411.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:53:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8330.7 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<01:17,  1.07s/it, est. speed input: 264.38 toks/s, output: 1457.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:21<01:17,  1.08s/it, est. speed input: 253.56 toks/s, output: 1491.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:22<01:07,  1.04it/s, est. speed input: 248.85 toks/s, output: 1557.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:22<00:51,  1.34it/s, est. speed input: 248.89 toks/s, output: 1651.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:23<00:50,  1.34it/s, est. speed input: 243.45 toks/s, output: 1707.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:23<00:38,  1.76it/s, est. speed input: 244.53 toks/s, output: 1807.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:23<00:36,  1.79it/s, est. speed input: 241.62 toks/s, output: 1876.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:23<00:28,  2.29it/s, est. speed input: 242.60 toks/s, output: 1974.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:24<00:22,  2.79it/s, est. speed input: 243.30 toks/s, output: 2070.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:24<00:22,  2.81it/s, est. speed input: 242.30 toks/s, output: 2150.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:25<00:31,  1.96it/s, est. speed input: 236.40 toks/s, output: 2186.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:25<00:26,  2.30it/s, est. speed input: 236.35 toks/s, output: 2274.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:27<00:44,  1.34it/s, est. speed input: 225.72 toks/s, output: 2260.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:27<00:37,  1.55it/s, est. speed input: 224.65 toks/s, output: 2337.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:27<00:28,  2.03it/s, est. speed input: 225.80 toks/s, output: 2436.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:27<00:21,  2.63it/s, est. speed input: 227.03 toks/s, output: 2535.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:28<00:33,  1.69it/s, est. speed input: 220.67 toks/s, output: 2551.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:29<00:30,  1.80it/s, est. speed input: 219.21 toks/s, output: 2620.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:53:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6954.1 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:30<00:36,  1.47it/s, est. speed input: 214.17 toks/s, output: 2647.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:24,  2.15it/s, est. speed input: 215.14 toks/s, output: 2830.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:25,  2.04it/s, est. speed input: 213.24 toks/s, output: 2890.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:20,  2.40it/s, est. speed input: 212.90 toks/s, output: 3053.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:32<00:17,  2.71it/s, est. speed input: 213.41 toks/s, output: 3145.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:32<00:14,  3.16it/s, est. speed input: 214.25 toks/s, output: 3241.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:32<00:13,  3.47it/s, est. speed input: 214.75 toks/s, output: 3331.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:34<00:35,  1.25it/s, est. speed input: 203.06 toks/s, output: 3234.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:35<00:34,  1.29it/s, est. speed input: 200.69 toks/s, output: 3280.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:34,  1.23it/s, est. speed input: 197.30 toks/s, output: 3310.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:36<00:29,  1.42it/s, est. speed input: 196.60 toks/s, output: 3383.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:37<00:31,  1.29it/s, est. speed input: 193.29 toks/s, output: 3410.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:38<00:23,  1.67it/s, est. speed input: 192.54 toks/s, output: 3565.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:39<00:25,  1.49it/s, est. speed input: 189.68 toks/s, output: 3597.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:39<00:24,  1.52it/s, est. speed input: 188.24 toks/s, output: 3654.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:54:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5446.9 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:40<00:22,  1.58it/s, est. speed input: 187.10 toks/s, output: 3718.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:40<00:17,  2.04it/s, est. speed input: 188.05 toks/s, output: 3820.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:40<00:13,  2.49it/s, est. speed input: 188.74 toks/s, output: 3917.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:40<00:11,  2.98it/s, est. speed input: 189.41 toks/s, output: 4014.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:41<00:11,  2.71it/s, est. speed input: 188.80 toks/s, output: 4085.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:41<00:09,  3.18it/s, est. speed input: 189.46 toks/s, output: 4181.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:41<00:07,  3.90it/s, est. speed input: 190.36 toks/s, output: 4283.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:42<00:09,  3.20it/s, est. speed input: 189.79 toks/s, output: 4353.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:42<00:12,  2.22it/s, est. speed input: 187.76 toks/s, output: 4389.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:43<00:14,  1.85it/s, est. speed input: 185.95 toks/s, output: 4429.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:44<00:18,  1.43it/s, est. speed input: 182.89 toks/s, output: 4441.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:45<00:14,  1.74it/s, est. speed input: 183.06 toks/s, output: 4528.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:13,  1.79it/s, est. speed input: 182.28 toks/s, output: 4594.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:47<00:19,  1.16it/s, est. speed input: 177.45 toks/s, output: 4558.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:54:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3772.2 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:34,  1.55s/it, est. speed input: 167.55 toks/s, output: 4392.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:50<00:23,  1.12s/it, est. speed input: 168.30 toks/s, output: 4500.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:51<00:19,  1.02it/s, est. speed input: 167.29 toks/s, output: 4561.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:53<00:24,  1.29s/it, est. speed input: 162.11 toks/s, output: 4510.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:18,  1.05s/it, est. speed input: 161.83 toks/s, output: 4591.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:14,  1.13it/s, est. speed input: 159.94 toks/s, output: 4717.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:56<00:15,  1.03s/it, est. speed input: 156.84 toks/s, output: 4717.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:57<00:13,  1.02it/s, est. speed input: 155.68 toks/s, output: 4774.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:57<00:10,  1.29it/s, est. speed input: 156.11 toks/s, output: 4880.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:54:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2808.6 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:01<00:20,  1.72s/it, est. speed input: 146.58 toks/s, output: 4680.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:01<00:14,  1.31s/it, est. speed input: 146.87 toks/s, output: 4786.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:03<00:12,  1.25s/it, est. speed input: 145.32 toks/s, output: 4834.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.25s/it, est. speed input: 154.96 toks/s, output: 6133.64 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.54it/s, est. speed input: 154.96 toks/s, output: 6133.64 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:54:36 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:54:36 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:54:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:54:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:54:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:54:36 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:54:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:54:36 [abstract.py:306] It took 0.542952 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:54:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:54:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:54:36 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.2437
+  • Mean Generation Length: 2416.4563
+  • Training FLOPS: 46.42 TFLOPS (11.61 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 293.33s
+  • policy_training: 188.41s (64.2%)
+  • generation: 67.10s (22.9%)
+  • teacher_logprob_inference: 32.84s (11.2%)
+  • policy_training/sharding_data: 4.22s (1.4%)
+  • prepare_for_generation: 3.57s (1.2%)
+  • policy_training/submit_training_futures: 1.96s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.5%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 123: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 08:58:19 [abstract.py:324] It took 0.053431 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08281758e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322983e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75270691e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08281758e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322983e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75270691e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08281758e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322983e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75270691e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08281758e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322983e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75270691e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08285020e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62170601e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76456070e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08050652e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.59773064e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #123] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44168991e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08285020e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62170601e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76456070e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67006378e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36209335e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #123] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.53085480e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08285020e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62170601e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76456070e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39916306e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.58728838e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #123] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.36629829e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08285020e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62170601e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76456070e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20245590e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50103874e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #123] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33754929e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 123: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 08:58:22 [abstract.py:324] It took 0.029543 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13191.19it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:58:23 [loggers.py:236] Engine 000: Avg prompt throughput: 41.4 tokens/s, Avg generation throughput: 29.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:01,  2.59it/s, est. speed input: 152.83 toks/s, output: 28.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:07, 19.87it/s, est. speed input: 973.70 toks/s, output: 268.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:07, 20.58it/s, est. speed input: 1060.66 toks/s, output: 405.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 24.58it/s, est. speed input: 1226.58 toks/s, output: 576.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 26.43it/s, est. speed input: 1326.82 toks/s, output: 745.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:05, 26.56it/s, est. speed input: 1369.60 toks/s, output: 902.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:04, 28.49it/s, est. speed input: 1437.24 toks/s, output: 1109.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 39.79it/s, est. speed input: 1681.39 toks/s, output: 1607.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 41.55it/s, est. speed input: 1768.41 toks/s, output: 1871.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 48.78it/s, est. speed input: 1950.99 toks/s, output: 2400.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 35.35it/s, est. speed input: 1846.88 toks/s, output: 2515.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:02<00:03, 31.26it/s, est. speed input: 1802.47 toks/s, output: 2690.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:02<00:05, 15.84it/s, est. speed input: 1447.53 toks/s, output: 2362.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:06<00:25,  3.63it/s, est. speed input: 671.34 toks/s, output: 1303.76 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:07<00:28,  3.11it/s, est. speed input: 583.83 toks/s, output: 1313.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:08<00:29,  2.98it/s, est. speed input: 541.68 toks/s, output: 1399.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:08<00:25,  3.33it/s, est. speed input: 534.09 toks/s, output: 1560.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:08<00:26,  3.26it/s, est. speed input: 519.56 toks/s, output: 1606.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:08<00:24,  3.41it/s, est. speed input: 513.45 toks/s, output: 1675.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:58:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10875.8 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:13<01:28,  1.06s/it, est. speed input: 346.33 toks/s, output: 1224.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:14<01:26,  1.06s/it, est. speed input: 325.84 toks/s, output: 1246.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:15<01:29,  1.10s/it, est. speed input: 303.58 toks/s, output: 1254.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:17<01:37,  1.21s/it, est. speed input: 279.95 toks/s, output: 1249.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:19<01:44,  1.33s/it, est. speed input: 258.92 toks/s, output: 1248.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:20<01:40,  1.28s/it, est. speed input: 246.77 toks/s, output: 1282.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:58:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8619.2 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:20<01:16,  1.01it/s, est. speed input: 246.99 toks/s, output: 1375.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:20<01:03,  1.19it/s, est. speed input: 244.56 toks/s, output: 1452.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<00:45,  1.63it/s, est. speed input: 242.57 toks/s, output: 1616.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:22<00:57,  1.27it/s, est. speed input: 230.91 toks/s, output: 1628.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:23<00:51,  1.39it/s, est. speed input: 228.29 toks/s, output: 1697.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:23<00:44,  1.60it/s, est. speed input: 227.36 toks/s, output: 1777.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:25<00:59,  1.17it/s, est. speed input: 216.65 toks/s, output: 1781.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:25<00:28,  2.35it/s, est. speed input: 222.27 toks/s, output: 2084.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:27<00:46,  1.41it/s, est. speed input: 209.66 toks/s, output: 2050.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:27<00:40,  1.61it/s, est. speed input: 209.34 toks/s, output: 2132.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:28<00:38,  1.68it/s, est. speed input: 207.58 toks/s, output: 2198.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<00:34,  1.81it/s, est. speed input: 206.60 toks/s, output: 2270.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:28<00:29,  2.07it/s, est. speed input: 206.54 toks/s, output: 2353.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:28<00:24,  2.46it/s, est. speed input: 207.09 toks/s, output: 2441.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:29<00:20,  2.94it/s, est. speed input: 208.00 toks/s, output: 2532.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:29<00:16,  3.61it/s, est. speed input: 209.19 toks/s, output: 2627.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:30<00:25,  2.28it/s, est. speed input: 205.48 toks/s, output: 2661.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:58:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7102.8 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:30<00:24,  2.29it/s, est. speed input: 204.48 toks/s, output: 2729.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:30<00:21,  2.55it/s, est. speed input: 204.52 toks/s, output: 2810.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:31<00:19,  2.84it/s, est. speed input: 204.78 toks/s, output: 2893.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:32<00:32,  1.64it/s, est. speed input: 198.92 toks/s, output: 2891.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:32<00:28,  1.89it/s, est. speed input: 198.74 toks/s, output: 2967.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:33<00:25,  2.01it/s, est. speed input: 198.07 toks/s, output: 3036.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:33<00:27,  1.83it/s, est. speed input: 196.02 toks/s, output: 3084.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:35<00:50,  1.02s/it, est. speed input: 186.06 toks/s, output: 3008.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:38,  1.28it/s, est. speed input: 186.56 toks/s, output: 3097.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:36<00:28,  1.69it/s, est. speed input: 187.37 toks/s, output: 3191.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:36<00:19,  2.36it/s, est. speed input: 188.36 toks/s, output: 3365.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:38<00:24,  1.79it/s, est. speed input: 183.94 toks/s, output: 3442.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:38<00:17,  2.37it/s, est. speed input: 185.33 toks/s, output: 3625.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:17,  2.23it/s, est. speed input: 183.70 toks/s, output: 3748.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:59:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5777.7 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:21,  1.81it/s, est. speed input: 180.71 toks/s, output: 3767.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:31,  1.20it/s, est. speed input: 174.27 toks/s, output: 3714.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:27,  1.37it/s, est. speed input: 173.96 toks/s, output: 3788.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:22,  1.58it/s, est. speed input: 173.95 toks/s, output: 3868.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:31,  1.12it/s, est. speed input: 169.03 toks/s, output: 3839.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:28,  1.21it/s, est. speed input: 167.89 toks/s, output: 3895.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:17,  1.87it/s, est. speed input: 169.23 toks/s, output: 4088.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:17,  1.79it/s, est. speed input: 168.24 toks/s, output: 4145.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:11,  2.61it/s, est. speed input: 169.96 toks/s, output: 4344.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:14,  1.98it/s, est. speed input: 167.95 toks/s, output: 4372.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:48<00:18,  1.43it/s, est. speed input: 164.70 toks/s, output: 4369.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:49<00:16,  1.55it/s, est. speed input: 164.29 toks/s, output: 4439.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:49<00:15,  1.64it/s, est. speed input: 163.81 toks/s, output: 4506.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:59:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4289.7 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:52<00:29,  1.24s/it, est. speed input: 156.05 toks/s, output: 4374.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:53<00:24,  1.07s/it, est. speed input: 155.31 toks/s, output: 4437.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:19,  1.15it/s, est. speed input: 155.38 toks/s, output: 4521.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:55<00:24,  1.16s/it, est. speed input: 151.19 toks/s, output: 4485.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:55<00:13,  1.42it/s, est. speed input: 152.54 toks/s, output: 4693.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:56<00:10,  1.71it/s, est. speed input: 153.04 toks/s, output: 4793.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:56<00:09,  1.75it/s, est. speed input: 152.62 toks/s, output: 4864.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:57<00:08,  1.93it/s, est. speed input: 152.67 toks/s, output: 4950.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:57<00:07,  1.95it/s, est. speed input: 152.36 toks/s, output: 5025.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:59<00:12,  1.11it/s, est. speed input: 148.55 toks/s, output: 4986.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 08:59:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3126.9 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:00<00:12,  1.07it/s, est. speed input: 147.05 toks/s, output: 5023.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:03<00:18,  1.57s/it, est. speed input: 140.79 toks/s, output: 4900.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:05<00:17,  1.62s/it, est. speed input: 137.92 toks/s, output: 4895.82 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.62s/it, est. speed input: 148.23 toks/s, output: 6275.65 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  2.45it/s, est. speed input: 148.23 toks/s, output: 6275.65 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 08:59:31 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 08:59:31 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 08:59:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:59:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:59:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:59:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 08:59:32 [abstract.py:306] It took 0.554658 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 08:59:32 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 08:59:32 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 08:59:32 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 08:59:32 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0122
+  • Training Accuracy: 0.3375
+  • Mean Generation Length: 2561.4500
+  • Training FLOPS: 45.66 TFLOPS (11.42 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 312.37s
+  • policy_training: 202.94s (65.0%)
+  • generation: 69.60s (22.3%)
+  • teacher_logprob_inference: 34.75s (11.1%)
+  • policy_training/sharding_data: 4.39s (1.4%)
+  • prepare_for_generation: 3.65s (1.2%)
+  • policy_training/submit_training_futures: 1.88s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.5%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.38s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 124: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:03:32 [abstract.py:324] It took 0.051137 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08282773e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62641373e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75151978e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08282773e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62641373e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75151978e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08282773e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62641373e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75151978e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08282773e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62641373e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75151978e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08280469e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62013321e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76241760e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08085747e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.60335207e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #124] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44042923e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08280469e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62013321e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76241760e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67152405e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35865326e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #124] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.55537109e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08280469e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62013321e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76241760e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39197998e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.46244764e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #124] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.39109802e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08280469e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62013321e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76241760e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.21524582e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49473934e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #124] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33498032e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 124: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:03:35 [abstract.py:324] It took 0.030303 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13010.89it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:03:36 [loggers.py:236] Engine 000: Avg prompt throughput: 38.4 tokens/s, Avg generation throughput: 46.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:51,  3.09it/s, est. speed input: 185.64 toks/s, output: 40.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 19.48it/s, est. speed input: 977.49 toks/s, output: 292.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:08, 16.70it/s, est. speed input: 939.52 toks/s, output: 385.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:07, 19.33it/s, est. speed input: 1037.33 toks/s, output: 550.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 28.52it/s, est. speed input: 1308.95 toks/s, output: 926.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:03, 35.33it/s, est. speed input: 1525.01 toks/s, output: 1277.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:03, 40.65it/s, est. speed input: 1694.31 toks/s, output: 1617.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 38.58it/s, est. speed input: 1742.82 toks/s, output: 1830.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:03, 38.49it/s, est. speed input: 1794.30 toks/s, output: 2076.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 47.46it/s, est. speed input: 1989.39 toks/s, output: 2630.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:03, 32.67it/s, est. speed input: 1842.20 toks/s, output: 2631.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:02<00:04, 21.15it/s, est. speed input: 1612.69 toks/s, output: 2510.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:05<00:20,  4.67it/s, est. speed input: 753.07 toks/s, output: 1403.12 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:07<00:32,  2.90it/s, est. speed input: 530.21 toks/s, output: 1247.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:03:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11107.4 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:11<00:55,  1.65it/s, est. speed input: 364.72 toks/s, output: 1040.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:12<00:53,  1.67it/s, est. speed input: 341.36 toks/s, output: 1155.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:15<01:13,  1.20it/s, est. speed input: 288.62 toks/s, output: 1069.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:15<01:03,  1.37it/s, est. speed input: 290.03 toks/s, output: 1165.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:15<01:00,  1.43it/s, est. speed input: 283.38 toks/s, output: 1229.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:17<01:17,  1.10it/s, est. speed input: 258.67 toks/s, output: 1212.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:18<01:14,  1.13it/s, est. speed input: 250.79 toks/s, output: 1265.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:19<01:25,  1.03s/it, est. speed input: 235.06 toks/s, output: 1275.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:03:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8937.2 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:21<01:26,  1.06s/it, est. speed input: 225.32 toks/s, output: 1311.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:21<01:07,  1.20it/s, est. speed input: 225.92 toks/s, output: 1402.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:21<00:44,  1.76it/s, est. speed input: 226.61 toks/s, output: 1579.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:22<00:45,  1.71it/s, est. speed input: 222.87 toks/s, output: 1638.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:22<00:42,  1.82it/s, est. speed input: 221.27 toks/s, output: 1710.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:23<00:44,  1.73it/s, est. speed input: 217.59 toks/s, output: 1765.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:23<00:36,  2.07it/s, est. speed input: 217.99 toks/s, output: 1852.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:23<00:29,  2.50it/s, est. speed input: 218.84 toks/s, output: 1941.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:25<00:51,  1.42it/s, est. speed input: 208.63 toks/s, output: 1932.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<00:47,  1.50it/s, est. speed input: 206.50 toks/s, output: 1994.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:29<01:44,  1.48s/it, est. speed input: 184.39 toks/s, output: 1863.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:30<01:27,  1.25s/it, est. speed input: 182.12 toks/s, output: 1922.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7341.1 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:30<01:11,  1.03s/it, est. speed input: 181.12 toks/s, output: 1992.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:31<00:55,  1.22it/s, est. speed input: 178.44 toks/s, output: 2121.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<00:48,  1.37it/s, est. speed input: 177.75 toks/s, output: 2193.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:34<01:20,  1.24s/it, est. speed input: 165.82 toks/s, output: 2126.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<01:10,  1.10s/it, est. speed input: 164.10 toks/s, output: 2184.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:36<00:51,  1.20it/s, est. speed input: 163.14 toks/s, output: 2328.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:36<00:43,  1.41it/s, est. speed input: 163.34 toks/s, output: 2409.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:37<00:38,  1.57it/s, est. speed input: 163.10 toks/s, output: 2485.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:37<00:30,  1.96it/s, est. speed input: 163.97 toks/s, output: 2576.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:37<00:24,  2.33it/s, est. speed input: 164.64 toks/s, output: 2663.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:38<00:18,  3.06it/s, est. speed input: 166.06 toks/s, output: 2839.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6477.3 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:40<00:33,  1.59it/s, est. speed input: 159.92 toks/s, output: 2882.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:40<00:29,  1.81it/s, est. speed input: 160.29 toks/s, output: 2965.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:41<00:32,  1.60it/s, est. speed input: 158.39 toks/s, output: 3007.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:41<00:26,  1.95it/s, est. speed input: 159.20 toks/s, output: 3097.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:41<00:21,  2.34it/s, est. speed input: 159.91 toks/s, output: 3185.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:42<00:18,  2.71it/s, est. speed input: 160.54 toks/s, output: 3272.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:42<00:16,  2.91it/s, est. speed input: 160.90 toks/s, output: 3354.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:42<00:18,  2.55it/s, est. speed input: 160.39 toks/s, output: 3417.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:42<00:14,  3.07it/s, est. speed input: 161.20 toks/s, output: 3507.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:43<00:13,  3.23it/s, est. speed input: 161.60 toks/s, output: 3589.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:43<00:13,  3.18it/s, est. speed input: 161.79 toks/s, output: 3665.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:43<00:13,  3.28it/s, est. speed input: 162.12 toks/s, output: 3745.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:45<00:29,  1.43it/s, est. speed input: 157.65 toks/s, output: 3715.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:46<00:28,  1.42it/s, est. speed input: 156.52 toks/s, output: 3763.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:46<00:17,  2.23it/s, est. speed input: 158.06 toks/s, output: 3947.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:47<00:20,  1.83it/s, est. speed input: 156.53 toks/s, output: 3981.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:47<00:18,  2.02it/s, est. speed input: 156.64 toks/s, output: 4056.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:48<00:20,  1.77it/s, est. speed input: 155.47 toks/s, output: 4098.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:49<00:20,  1.73it/s, est. speed input: 154.81 toks/s, output: 4151.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:49<00:15,  2.24it/s, est. speed input: 155.65 toks/s, output: 4246.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:49<00:12,  2.57it/s, est. speed input: 156.10 toks/s, output: 4330.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:49<00:12,  2.59it/s, est. speed input: 156.12 toks/s, output: 4402.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4970.7 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:50<00:19,  1.62it/s, est. speed input: 153.74 toks/s, output: 4407.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:52<00:22,  1.32it/s, est. speed input: 151.70 toks/s, output: 4421.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:52<00:14,  1.92it/s, est. speed input: 152.58 toks/s, output: 4594.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:52<00:13,  1.98it/s, est. speed input: 152.39 toks/s, output: 4661.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:54<00:17,  1.49it/s, est. speed input: 150.25 toks/s, output: 4670.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:54<00:16,  1.53it/s, est. speed input: 149.69 toks/s, output: 4727.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:55<00:13,  1.83it/s, est. speed input: 150.06 toks/s, output: 4813.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:55<00:14,  1.63it/s, est. speed input: 149.06 toks/s, output: 4854.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:55<00:10,  2.02it/s, est. speed input: 149.60 toks/s, output: 4946.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:56<00:11,  1.90it/s, est. speed input: 149.03 toks/s, output: 5003.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:57<00:11,  1.68it/s, est. speed input: 148.16 toks/s, output: 5047.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:57<00:11,  1.72it/s, est. speed input: 147.80 toks/s, output: 5109.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:58<00:05,  2.76it/s, est. speed input: 149.43 toks/s, output: 5390.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:58<00:05,  2.56it/s, est. speed input: 149.19 toks/s, output: 5457.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:59<00:07,  1.90it/s, est. speed input: 147.74 toks/s, output: 5480.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3513.6 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:00<00:07,  1.79it/s, est. speed input: 147.11 toks/s, output: 5534.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:12,  1.02s/it, est. speed input: 142.63 toks/s, output: 5447.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:04<00:09,  1.07it/s, est. speed input: 140.80 toks/s, output: 5538.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:05<00:07,  1.18it/s, est. speed input: 140.51 toks/s, output: 5609.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:07<00:09,  1.13s/it, est. speed input: 137.32 toks/s, output: 5566.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:07<00:06,  1.09it/s, est. speed input: 137.54 toks/s, output: 5661.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:07<00:00,  1.09it/s, est. speed input: 143.93 toks/s, output: 6512.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:07<00:00,  2.37it/s, est. speed input: 143.93 toks/s, output: 6512.22 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:47 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:47 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:04:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:04:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:04:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:04:47 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:04:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:04:48 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:04:48 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:04:48 [abstract.py:306] It took 0.681784 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:04:48 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0107
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 2744.1687
+  • Training FLOPS: 47.29 TFLOPS (11.82 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 324.70s
+  • policy_training: 209.42s (64.5%)
+  • generation: 72.17s (22.2%)
+  • teacher_logprob_inference: 37.66s (11.6%)
+  • policy_training/sharding_data: 4.31s (1.3%)
+  • prepare_for_generation: 3.87s (1.2%)
+  • policy_training/submit_training_futures: 1.91s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.75s (0.5%)
+  • training_prep: 0.97s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • data_processing: 0.06s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 125: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:08:57 [abstract.py:324] It took 0.052423 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08297109e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62247009e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75299835e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08297109e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62247009e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75299835e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08297109e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62247009e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75299835e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08297109e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62247009e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75299835e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08297812e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63558388e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78487854e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08887672e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.63137054e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #125] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44006500e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08297812e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63558388e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78487854e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67502213e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35672712e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #125] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.53195801e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08297812e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63558388e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78487854e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40627899e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.11426115e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #125] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.38582268e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08297812e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63558388e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78487854e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.21892776e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.49935093e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #125] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33666733e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 125: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:00 [abstract.py:324] It took 0.034177 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13502.25it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:00 [loggers.py:236] Engine 000: Avg prompt throughput: 36.8 tokens/s, Avg generation throughput: 58.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:20,  1.97it/s, est. speed input: 113.99 toks/s, output: 25.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:12, 12.32it/s, est. speed input: 586.87 toks/s, output: 169.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:07, 20.47it/s, est. speed input: 932.70 toks/s, output: 346.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:05, 25.39it/s, est. speed input: 1176.92 toks/s, output: 609.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 39.33it/s, est. speed input: 1605.92 toks/s, output: 1130.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:02, 41.77it/s, est. speed input: 1741.34 toks/s, output: 1413.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 52.35it/s, est. speed input: 2064.62 toks/s, output: 2041.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 65.05it/s, est. speed input: 2379.03 toks/s, output: 2702.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 60.35it/s, est. speed input: 2446.07 toks/s, output: 3050.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 60.80it/s, est. speed input: 2533.63 toks/s, output: 3410.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:01<00:01, 65.24it/s, est. speed input: 2656.81 toks/s, output: 3887.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:02<00:01, 57.33it/s, est. speed input: 2655.90 toks/s, output: 4174.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9662.1 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:11<00:24,  2.65it/s, est. speed input: 520.31 toks/s, output: 1194.60 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:15<00:30,  1.97it/s, est. speed input: 382.17 toks/s, output: 1360.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:18<00:29,  1.87it/s, est. speed input: 342.86 toks/s, output: 1639.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:19<00:26,  2.02it/s, est. speed input: 335.71 toks/s, output: 1928.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:20<00:23,  2.16it/s, est. speed input: 332.52 toks/s, output: 2125.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7630.0 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:23<00:35,  1.39it/s, est. speed input: 283.36 toks/s, output: 2024.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:28<00:56,  1.19s/it, est. speed input: 236.43 toks/s, output: 1798.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:29<00:52,  1.11s/it, est. speed input: 233.66 toks/s, output: 1886.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:30<00:48,  1.06s/it, est. speed input: 229.56 toks/s, output: 1962.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6067.3 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:31<00:42,  1.04it/s, est. speed input: 222.24 toks/s, output: 2110.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:32<00:29,  1.41it/s, est. speed input: 223.99 toks/s, output: 2341.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:32<00:26,  1.55it/s, est. speed input: 223.44 toks/s, output: 2440.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:32<00:23,  1.70it/s, est. speed input: 222.78 toks/s, output: 2538.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:33<00:21,  1.84it/s, est. speed input: 221.97 toks/s, output: 2633.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:34<00:24,  1.55it/s, est. speed input: 217.44 toks/s, output: 2684.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:34<00:21,  1.72it/s, est. speed input: 216.70 toks/s, output: 2778.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:34<00:17,  2.06it/s, est. speed input: 216.96 toks/s, output: 2885.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:35<00:10,  3.22it/s, est. speed input: 219.45 toks/s, output: 3120.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:35<00:14,  2.22it/s, est. speed input: 215.64 toks/s, output: 3168.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:36<00:15,  2.12it/s, est. speed input: 214.17 toks/s, output: 3247.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:37<00:23,  1.32it/s, est. speed input: 206.99 toks/s, output: 3240.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:39<00:25,  1.18it/s, est. speed input: 202.75 toks/s, output: 3276.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:40<00:26,  1.11it/s, est. speed input: 199.06 toks/s, output: 3319.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4761.6 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:41<00:25,  1.09it/s, est. speed input: 195.97 toks/s, output: 3369.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:19,  1.40it/s, est. speed input: 196.32 toks/s, output: 3477.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:43<00:28,  1.11s/it, est. speed input: 188.39 toks/s, output: 3439.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:43<00:23,  1.04it/s, est. speed input: 187.20 toks/s, output: 3521.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:44<00:19,  1.23it/s, est. speed input: 186.65 toks/s, output: 3612.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:20,  1.14it/s, est. speed input: 183.72 toks/s, output: 3658.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:45<00:16,  1.37it/s, est. speed input: 183.48 toks/s, output: 3757.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:23,  1.11s/it, est. speed input: 177.04 toks/s, output: 3730.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:09:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3618.6 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:34,  1.70s/it, est. speed input: 167.47 toks/s, output: 3635.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:51<00:24,  1.29s/it, est. speed input: 167.59 toks/s, output: 3743.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:52<00:22,  1.25s/it, est. speed input: 165.05 toks/s, output: 3793.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:52<00:15,  1.09it/s, est. speed input: 165.73 toks/s, output: 3916.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:53<00:13,  1.21it/s, est. speed input: 165.00 toks/s, output: 4005.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:53<00:10,  1.40it/s, est. speed input: 164.73 toks/s, output: 4104.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:08,  1.63it/s, est. speed input: 164.75 toks/s, output: 4209.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:54<00:06,  1.98it/s, est. speed input: 165.05 toks/s, output: 4322.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:56<00:13,  1.12s/it, est. speed input: 158.68 toks/s, output: 4264.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:58<00:12,  1.16s/it, est. speed input: 156.29 toks/s, output: 4309.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:58<00:09,  1.00it/s, est. speed input: 155.65 toks/s, output: 4401.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:59<00:07,  1.13it/s, est. speed input: 155.02 toks/s, output: 4494.61 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  1.13it/s, est. speed input: 164.21 toks/s, output: 5738.12 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  2.70it/s, est. speed input: 164.21 toks/s, output: 5738.12 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:10:02 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:10:02 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:10:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:10:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:10:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:10:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:10:03 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:10:03 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:10:03 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:10:03 [abstract.py:306] It took 0.610433 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:10:03 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0111
+  • Training Accuracy: 0.3688
+  • Mean Generation Length: 2126.3062
+  • Training FLOPS: 45.34 TFLOPS (11.33 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 270.61s
+  • policy_training: 171.77s (63.5%)
+  • generation: 63.13s (23.3%)
+  • teacher_logprob_inference: 30.49s (11.3%)
+  • policy_training/sharding_data: 4.35s (1.6%)
+  • prepare_for_generation: 3.69s (1.4%)
+  • policy_training/submit_training_futures: 1.96s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.68s (0.6%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.48s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 26/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 126: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:13:28 [abstract.py:324] It took 0.050379 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08307324e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62751923e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75533600e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08307324e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62751923e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75533600e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08307324e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62751923e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75533600e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08307324e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.62751923e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75533600e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08301758e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64335785e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276367e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09946404e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.56768703e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #126] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43944778e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08301758e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64335785e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276367e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66387024e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35382509e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #126] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.50350342e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08301758e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64335785e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276367e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41514130e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.09774256e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #126] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.38216782e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08301758e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64335785e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276367e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20461006e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.51561451e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #126] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33792618e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 126: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:13:31 [abstract.py:324] It took 0.030709 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13185.75it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 09:13:31 [loggers.py:236] Engine 000: Avg prompt throughput: 43.8 tokens/s, Avg generation throughput: 102.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:28,  5.61it/s, est. speed input: 342.17 toks/s, output: 78.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:06, 24.70it/s, est. speed input: 1287.85 toks/s, output: 369.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:05, 29.10it/s, est. speed input: 1548.17 toks/s, output: 561.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 30.52it/s, est. speed input: 1650.02 toks/s, output: 764.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 30.97it/s, est. speed input: 1736.72 toks/s, output: 1045.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:04, 32.39it/s, est. speed input: 1802.61 toks/s, output: 1256.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 43.55it/s, est. speed input: 2108.40 toks/s, output: 1842.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:01, 59.19it/s, est. speed input: 2506.72 toks/s, output: 2648.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:01, 66.72it/s, est. speed input: 2745.51 toks/s, output: 3248.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 59.28it/s, est. speed input: 2761.11 toks/s, output: 3571.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 35.49it/s, est. speed input: 2372.37 toks/s, output: 3371.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:08<00:26,  3.31it/s, est. speed input: 551.87 toks/s, output: 1088.82 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 09:13:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10759.6 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:10<00:32,  2.57it/s, est. speed input: 427.00 toks/s, output: 1235.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:12<00:32,  2.45it/s, est. speed input: 389.46 toks/s, output: 1402.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:14<00:39,  1.97it/s, est. speed input: 339.91 toks/s, output: 1416.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<00:42,  1.78it/s, est. speed input: 312.81 toks/s, output: 1485.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:17<00:45,  1.64it/s, est. speed input: 298.25 toks/s, output: 1510.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:17<00:42,  1.75it/s, est. speed input: 296.10 toks/s, output: 1593.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:19<00:53,  1.36it/s, est. speed input: 274.56 toks/s, output: 1571.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:13:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8441.6 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<01:03,  1.14it/s, est. speed input: 257.72 toks/s, output: 1568.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<00:43,  1.60it/s, est. speed input: 259.82 toks/s, output: 1765.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<00:43,  1.60it/s, est. speed input: 254.79 toks/s, output: 1823.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:37,  1.83it/s, est. speed input: 254.57 toks/s, output: 1911.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:22<00:29,  2.24it/s, est. speed input: 255.70 toks/s, output: 2010.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:22<00:29,  2.23it/s, est. speed input: 253.09 toks/s, output: 2079.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:22<00:22,  2.88it/s, est. speed input: 253.90 toks/s, output: 2259.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:23<00:29,  2.17it/s, est. speed input: 247.38 toks/s, output: 2289.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:24<00:32,  1.90it/s, est. speed input: 240.03 toks/s, output: 2391.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:25<00:32,  1.84it/s, est. speed input: 236.74 toks/s, output: 2444.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:26<00:34,  1.69it/s, est. speed input: 232.28 toks/s, output: 2485.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:26<00:29,  1.99it/s, est. speed input: 232.49 toks/s, output: 2574.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:27<00:31,  1.79it/s, est. speed input: 228.56 toks/s, output: 2617.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:27<00:33,  1.68it/s, est. speed input: 225.03 toks/s, output: 2663.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:28<00:25,  2.13it/s, est. speed input: 225.90 toks/s, output: 2760.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:28<00:24,  2.21it/s, est. speed input: 224.75 toks/s, output: 2831.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:28<00:19,  2.74it/s, est. speed input: 225.76 toks/s, output: 2927.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:29<00:32,  1.61it/s, est. speed input: 218.37 toks/s, output: 2917.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:14:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6914.4 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:25,  1.97it/s, est. speed input: 218.70 toks/s, output: 3006.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:23,  2.11it/s, est. speed input: 216.45 toks/s, output: 3142.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:31<00:12,  3.75it/s, est. speed input: 221.08 toks/s, output: 3459.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:31<00:12,  3.55it/s, est. speed input: 220.52 toks/s, output: 3533.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:32<00:18,  2.41it/s, est. speed input: 216.29 toks/s, output: 3548.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:32<00:16,  2.61it/s, est. speed input: 216.29 toks/s, output: 3630.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:33<00:21,  1.95it/s, est. speed input: 212.26 toks/s, output: 3646.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:33<00:13,  2.90it/s, est. speed input: 214.41 toks/s, output: 3847.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:35<00:17,  2.23it/s, est. speed input: 210.25 toks/s, output: 3934.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:35<00:17,  2.15it/s, est. speed input: 208.79 toks/s, output: 3989.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:36<00:20,  1.77it/s, est. speed input: 205.33 toks/s, output: 4004.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:36<00:14,  2.41it/s, est. speed input: 206.46 toks/s, output: 4190.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:37<00:14,  2.24it/s, est. speed input: 204.99 toks/s, output: 4242.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:37<00:09,  3.35it/s, est. speed input: 207.36 toks/s, output: 4455.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:37<00:08,  3.58it/s, est. speed input: 207.77 toks/s, output: 4544.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:38<00:08,  3.39it/s, est. speed input: 207.45 toks/s, output: 4618.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:38<00:06,  4.04it/s, est. speed input: 208.72 toks/s, output: 4804.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:14:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4951.3 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:40<00:14,  1.79it/s, est. speed input: 201.64 toks/s, output: 4724.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:40<00:11,  2.19it/s, est. speed input: 202.44 toks/s, output: 4825.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:40<00:11,  2.08it/s, est. speed input: 201.12 toks/s, output: 4877.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:41<00:12,  1.80it/s, est. speed input: 198.89 toks/s, output: 4905.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:41<00:10,  2.02it/s, est. speed input: 198.80 toks/s, output: 4984.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:42<00:09,  2.14it/s, est. speed input: 198.36 toks/s, output: 5055.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:43<00:12,  1.56it/s, est. speed input: 194.91 toks/s, output: 5051.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:44<00:13,  1.46it/s, est. speed input: 192.71 toks/s, output: 5081.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:44<00:11,  1.63it/s, est. speed input: 192.18 toks/s, output: 5152.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:45<00:11,  1.49it/s, est. speed input: 190.05 toks/s, output: 5182.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:46<00:11,  1.45it/s, est. speed input: 188.35 toks/s, output: 5222.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:14:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3148.7 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:18,  1.30s/it, est. speed input: 175.49 toks/s, output: 5041.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:50<00:14,  1.12s/it, est. speed input: 174.67 toks/s, output: 5112.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:51<00:11,  1.08it/s, est. speed input: 174.55 toks/s, output: 5204.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:51<00:09,  1.12it/s, est. speed input: 172.96 toks/s, output: 5253.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:52<00:07,  1.32it/s, est. speed input: 172.78 toks/s, output: 5342.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:53<00:04,  1.69it/s, est. speed input: 172.52 toks/s, output: 5524.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:53<00:02,  2.59it/s, est. speed input: 174.31 toks/s, output: 5774.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:54<00:03,  1.56it/s, est. speed input: 170.52 toks/s, output: 5744.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:58<00:05,  1.27s/it, est. speed input: 161.86 toks/s, output: 5556.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:58<00:03,  1.02s/it, est. speed input: 162.19 toks/s, output: 5671.17 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.02s/it, est. speed input: 165.31 toks/s, output: 6091.88 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.74it/s, est. speed input: 165.31 toks/s, output: 6091.88 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:14:33 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:14:33 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:14:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:14:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:14:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:14:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:14:33 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:14:33 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:14:33 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:14:33 [abstract.py:306] It took 0.530766 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:14:33 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0102
+  • Training Accuracy: 0.3625
+  • Mean Generation Length: 2224.0249
+  • Training FLOPS: 43.71 TFLOPS (10.93 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 280.20s
+  • policy_training: 181.98s (64.9%)
+  • generation: 62.34s (22.2%)
+  • teacher_logprob_inference: 30.63s (10.9%)
+  • policy_training/sharding_data: 4.37s (1.6%)
+  • prepare_for_generation: 3.64s (1.3%)
+  • policy_training/submit_training_futures: 1.75s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.55s (0.6%)
+  • training_prep: 1.04s (0.4%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 127: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:08 [abstract.py:324] It took 0.048087 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08316875e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63461800e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75597076e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08316875e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63461800e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75597076e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08316875e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63461800e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75597076e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08316875e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63461800e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75597076e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08314512e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64999847e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77812500e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08952827e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.67035913e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #127] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44030624e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08314512e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64999847e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77812500e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66829224e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34457798e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #127] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.51140137e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08314512e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64999847e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77812500e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40504761e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.34317398e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #127] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.38986549e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08314512e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64999847e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77812500e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20169373e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.52687569e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #127] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34038681e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 127: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:11 [abstract.py:324] It took 0.033049 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12715.55it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:11 [loggers.py:236] Engine 000: Avg prompt throughput: 42.7 tokens/s, Avg generation throughput: 59.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:51,  3.09it/s, est. speed input: 185.12 toks/s, output: 27.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:08, 17.15it/s, est. speed input: 854.18 toks/s, output: 200.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 24.60it/s, est. speed input: 1209.25 toks/s, output: 373.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:07, 19.53it/s, est. speed input: 1114.62 toks/s, output: 448.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:06, 20.79it/s, est. speed input: 1170.89 toks/s, output: 645.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:05, 22.88it/s, est. speed input: 1246.46 toks/s, output: 859.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:04, 32.46it/s, est. speed input: 1482.43 toks/s, output: 1313.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 37.00it/s, est. speed input: 1632.67 toks/s, output: 1714.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:03, 32.47it/s, est. speed input: 1620.92 toks/s, output: 1907.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:03, 35.54it/s, est. speed input: 1695.42 toks/s, output: 2208.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:03, 35.56it/s, est. speed input: 1733.28 toks/s, output: 2471.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:02<00:04, 23.38it/s, est. speed input: 1572.97 toks/s, output: 2440.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:03<00:11,  8.79it/s, est. speed input: 1075.40 toks/s, output: 1873.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:08<00:44,  2.18it/s, est. speed input: 458.10 toks/s, output: 1041.21 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11406.7 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:10<00:51,  1.86it/s, est. speed input: 388.24 toks/s, output: 1058.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:10<00:43,  2.15it/s, est. speed input: 387.39 toks/s, output: 1234.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:10<00:41,  2.24it/s, est. speed input: 381.61 toks/s, output: 1304.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:11<00:36,  2.51it/s, est. speed input: 382.90 toks/s, output: 1396.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:12<00:50,  1.78it/s, est. speed input: 347.10 toks/s, output: 1353.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:12<00:46,  1.93it/s, est. speed input: 342.38 toks/s, output: 1422.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:15<01:24,  1.04it/s, est. speed input: 290.56 toks/s, output: 1295.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:17<01:59,  1.37s/it, est. speed input: 251.13 toks/s, output: 1207.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:18<01:37,  1.13s/it, est. speed input: 248.44 toks/s, output: 1283.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:18<01:24,  1.00it/s, est. speed input: 243.39 toks/s, output: 1345.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9010.0 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:20<01:20,  1.04it/s, est. speed input: 227.50 toks/s, output: 1424.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:22<01:35,  1.17s/it, est. speed input: 212.00 toks/s, output: 1414.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:24<01:43,  1.27s/it, est. speed input: 200.61 toks/s, output: 1424.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:24<01:18,  1.02it/s, est. speed input: 201.72 toks/s, output: 1517.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:26<01:32,  1.17s/it, est. speed input: 191.27 toks/s, output: 1522.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:26<01:10,  1.10it/s, est. speed input: 191.80 toks/s, output: 1611.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:26<00:54,  1.41it/s, est. speed input: 192.41 toks/s, output: 1700.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:27<00:56,  1.34it/s, est. speed input: 188.74 toks/s, output: 1750.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7647.6 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:30<01:47,  1.44s/it, est. speed input: 171.55 toks/s, output: 1673.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:31<01:47,  1.45s/it, est. speed input: 165.55 toks/s, output: 1696.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:32<01:23,  1.14s/it, est. speed input: 165.38 toks/s, output: 1776.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:32<01:04,  1.11it/s, est. speed input: 165.59 toks/s, output: 1858.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:33<00:52,  1.35it/s, est. speed input: 165.67 toks/s, output: 1939.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:33<00:35,  1.96it/s, est. speed input: 166.91 toks/s, output: 2111.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:35<00:56,  1.21it/s, est. speed input: 160.13 toks/s, output: 2104.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:35<00:36,  1.80it/s, est. speed input: 162.05 toks/s, output: 2282.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:35<00:29,  2.17it/s, est. speed input: 163.25 toks/s, output: 2374.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<00:24,  2.64it/s, est. speed input: 164.41 toks/s, output: 2465.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<00:25,  2.44it/s, est. speed input: 163.88 toks/s, output: 2531.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:37<00:31,  1.97it/s, est. speed input: 162.12 toks/s, output: 2579.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:37<00:24,  2.48it/s, est. speed input: 163.21 toks/s, output: 2670.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<00:48,  1.23it/s, est. speed input: 157.07 toks/s, output: 2645.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:39<00:39,  1.48it/s, est. speed input: 157.24 toks/s, output: 2723.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:18:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6552.9 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:42<01:02,  1.09s/it, est. speed input: 148.40 toks/s, output: 2714.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:43<00:52,  1.06it/s, est. speed input: 148.11 toks/s, output: 2785.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:45<01:15,  1.38s/it, est. speed input: 140.87 toks/s, output: 2725.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:47<01:14,  1.37s/it, est. speed input: 138.17 toks/s, output: 2748.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:47<00:59,  1.12s/it, est. speed input: 138.08 toks/s, output: 2821.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:48<00:52,  1.02s/it, est. speed input: 137.21 toks/s, output: 2878.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:49<00:49,  1.03it/s, est. speed input: 136.06 toks/s, output: 2929.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:49<00:30,  1.58it/s, est. speed input: 137.31 toks/s, output: 3104.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:49<00:20,  2.33it/s, est. speed input: 139.17 toks/s, output: 3292.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:19:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5672.3 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:50<00:24,  1.90it/s, est. speed input: 137.95 toks/s, output: 3337.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:51<00:29,  1.54it/s, est. speed input: 136.32 toks/s, output: 3371.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:54<00:49,  1.12s/it, est. speed input: 131.20 toks/s, output: 3318.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:55<00:50,  1.17s/it, est. speed input: 129.25 toks/s, output: 3341.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:57<00:39,  1.04it/s, est. speed input: 128.28 toks/s, output: 3462.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:57<00:34,  1.16it/s, est. speed input: 128.19 toks/s, output: 3533.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:59<00:45,  1.17s/it, est. speed input: 124.74 toks/s, output: 3512.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:00<00:27,  1.32it/s, est. speed input: 126.04 toks/s, output: 3697.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:19:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4861.3 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:01<00:31,  1.14it/s, est. speed input: 124.46 toks/s, output: 3724.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:03<00:40,  1.17s/it, est. speed input: 121.35 toks/s, output: 3705.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [01:03<00:34,  1.02s/it, est. speed input: 121.14 toks/s, output: 3773.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:04<00:15,  1.99it/s, est. speed input: 123.74 toks/s, output: 4076.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:05<00:20,  1.46it/s, est. speed input: 122.06 toks/s, output: 4092.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:06<00:17,  1.57it/s, est. speed input: 121.80 toks/s, output: 4229.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [01:07<00:16,  1.55it/s, est. speed input: 121.21 toks/s, output: 4355.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:09<00:22,  1.10it/s, est. speed input: 118.64 toks/s, output: 4337.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:19:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3691.3 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:11<00:23,  1.02it/s, est. speed input: 117.40 toks/s, output: 4366.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:11<00:20,  1.14it/s, est. speed input: 117.37 toks/s, output: 4439.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:12<00:11,  1.67it/s, est. speed input: 118.31 toks/s, output: 4697.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:13<00:10,  1.73it/s, est. speed input: 118.37 toks/s, output: 4772.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:13<00:10,  1.75it/s, est. speed input: 118.29 toks/s, output: 4844.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:15<00:13,  1.28it/s, est. speed input: 116.81 toks/s, output: 4858.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:15<00:09,  1.62it/s, est. speed input: 117.44 toks/s, output: 4958.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:15<00:07,  1.96it/s, est. speed input: 117.91 toks/s, output: 5052.63 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  1.96it/s, est. speed input: 130.19 toks/s, output: 6679.66 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:15<00:00,  2.12it/s, est. speed input: 130.19 toks/s, output: 6679.66 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:19:31 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:19:31 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:19:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:19:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:19:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:19:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:19:32 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:19:32 [abstract.py:306] It took 0.563278 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:19:32 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:19:32 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:19:32 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0093
+  • Training Accuracy: 0.3438
+  • Mean Generation Length: 3152.9312
+  • Training FLOPS: 49.54 TFLOPS (12.39 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 360.71s
+  • policy_training: 232.94s (64.6%)
+  • generation: 80.65s (22.4%)
+  • teacher_logprob_inference: 41.90s (11.6%)
+  • policy_training/sharding_data: 4.33s (1.2%)
+  • prepare_for_generation: 3.67s (1.0%)
+  • policy_training/submit_training_futures: 2.06s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.4%)
+  • training_prep: 0.91s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 128: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:09 [abstract.py:324] It took 0.051143 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08349004e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64033546e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75581055e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08349004e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64033546e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75581055e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08349004e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64033546e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75581055e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08349004e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64033546e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75581055e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08355156e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63906403e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77384644e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07432556e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.57356787e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #128] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44217560e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08355156e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63906403e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77384644e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67483292e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33079948e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #128] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.48111801e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08355156e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63906403e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77384644e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40125885e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.47736549e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #128] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.36333771e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08355156e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63906403e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77384644e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19846344e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.52185869e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #128] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33686111e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 128: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:12 [abstract.py:324] It took 0.029481 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13299.15it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:12 [loggers.py:236] Engine 000: Avg prompt throughput: 33.4 tokens/s, Avg generation throughput: 54.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:40,  3.88it/s, est. speed input: 236.73 toks/s, output: 54.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:06, 23.46it/s, est. speed input: 1203.43 toks/s, output: 382.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 23.07it/s, est. speed input: 1249.57 toks/s, output: 489.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 27.79it/s, est. speed input: 1440.41 toks/s, output: 703.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 42.70it/s, est. speed input: 1910.50 toks/s, output: 1224.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:03, 43.12it/s, est. speed input: 2036.76 toks/s, output: 1501.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 49.76it/s, est. speed input: 2256.56 toks/s, output: 1923.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:01, 58.11it/s, est. speed input: 2500.25 toks/s, output: 2424.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 72.86it/s, est. speed input: 2879.63 toks/s, output: 3230.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 72.62it/s, est. speed input: 3021.02 toks/s, output: 3712.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 64.92it/s, est. speed input: 3029.23 toks/s, output: 4063.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:01<00:01, 54.11it/s, est. speed input: 2942.42 toks/s, output: 4276.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:01<00:01, 53.07it/s, est. speed input: 2947.39 toks/s, output: 4589.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:02<00:04, 15.53it/s, est. speed input: 1908.02 toks/s, output: 3283.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9584.8 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:11<00:27,  2.32it/s, est. speed input: 538.06 toks/s, output: 1324.72 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:14<00:32,  1.84it/s, est. speed input: 427.71 toks/s, output: 1373.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:16<00:34,  1.66it/s, est. speed input: 373.72 toks/s, output: 1533.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:17<00:26,  2.06it/s, est. speed input: 380.05 toks/s, output: 1897.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:19<00:31,  1.65it/s, est. speed input: 339.83 toks/s, output: 1909.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7486.2 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:20<00:29,  1.68it/s, est. speed input: 327.37 toks/s, output: 2058.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:21<00:31,  1.55it/s, est. speed input: 314.87 toks/s, output: 2089.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:27,  1.68it/s, est. speed input: 307.89 toks/s, output: 2257.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:23<00:29,  1.55it/s, est. speed input: 292.73 toks/s, output: 2357.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:24<00:27,  1.60it/s, est. speed input: 288.91 toks/s, output: 2435.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:25<00:26,  1.63it/s, est. speed input: 284.95 toks/s, output: 2509.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:44,  1.07s/it, est. speed input: 258.64 toks/s, output: 2387.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:28<00:40,  1.00it/s, est. speed input: 253.99 toks/s, output: 2452.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:28<00:32,  1.22it/s, est. speed input: 253.61 toks/s, output: 2556.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:33,  1.15it/s, est. speed input: 247.07 toks/s, output: 2598.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5719.1 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:32,  1.17it/s, est. speed input: 242.37 toks/s, output: 2656.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:24,  1.53it/s, est. speed input: 243.30 toks/s, output: 2773.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:31<00:20,  1.78it/s, est. speed input: 242.69 toks/s, output: 2872.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:32<00:25,  1.37it/s, est. speed input: 235.90 toks/s, output: 2898.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:32<00:19,  1.77it/s, est. speed input: 236.59 toks/s, output: 3012.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:32<00:12,  2.56it/s, est. speed input: 237.70 toks/s, output: 3234.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:33<00:12,  2.38it/s, est. speed input: 234.69 toks/s, output: 3400.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:11,  2.49it/s, est. speed input: 234.14 toks/s, output: 3496.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:12,  2.16it/s, est. speed input: 231.44 toks/s, output: 3559.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:16,  1.66it/s, est. speed input: 226.58 toks/s, output: 3589.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:36<00:13,  1.95it/s, est. speed input: 226.60 toks/s, output: 3693.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:36<00:11,  2.12it/s, est. speed input: 226.04 toks/s, output: 3787.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:37<00:15,  1.54it/s, est. speed input: 220.97 toks/s, output: 3806.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:10,  2.08it/s, est. speed input: 220.93 toks/s, output: 4013.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:10,  1.98it/s, est. speed input: 217.89 toks/s, output: 4163.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:39<00:10,  1.82it/s, est. speed input: 215.53 toks/s, output: 4222.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:24:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4090.7 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:41<00:15,  1.17it/s, est. speed input: 207.44 toks/s, output: 4169.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:11,  1.35it/s, est. speed input: 204.70 toks/s, output: 4322.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:43<00:08,  1.67it/s, est. speed input: 204.25 toks/s, output: 4521.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:44<00:07,  1.68it/s, est. speed input: 202.87 toks/s, output: 4597.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:07,  1.70it/s, est. speed input: 201.62 toks/s, output: 4675.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:05,  2.09it/s, est. speed input: 202.36 toks/s, output: 4798.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:45<00:04,  2.20it/s, est. speed input: 201.96 toks/s, output: 4894.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:04,  1.85it/s, est. speed input: 199.88 toks/s, output: 4951.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:03,  2.29it/s, est. speed input: 200.50 toks/s, output: 5072.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:02,  2.80it/s, est. speed input: 201.05 toks/s, output: 5193.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:03,  1.77it/s, est. speed input: 197.74 toks/s, output: 5216.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:48<00:02,  1.99it/s, est. speed input: 196.69 toks/s, output: 5405.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:25:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2062.5 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:53<00:05,  1.69s/it, est. speed input: 178.19 toks/s, output: 5017.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  1.69s/it, est. speed input: 181.59 toks/s, output: 5476.68 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  2.99it/s, est. speed input: 181.59 toks/s, output: 5476.68 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:25:09 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:25:09 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:25:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:25:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:25:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:25:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:25:09 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:25:09 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:25:09 [abstract.py:306] It took 0.629393 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:25:09 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:25:09 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.4313
+  • Mean Generation Length: 1833.6875
+  • Training FLOPS: 39.42 TFLOPS (9.86 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 257.86s
+  • policy_training: 167.75s (65.1%)
+  • generation: 56.96s (22.1%)
+  • teacher_logprob_inference: 27.98s (10.9%)
+  • policy_training/sharding_data: 4.33s (1.7%)
+  • prepare_for_generation: 3.71s (1.4%)
+  • policy_training/submit_training_futures: 2.02s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.72s (0.7%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 129: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:28:27 [abstract.py:324] It took 0.048368 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08380195e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64944382e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76188202e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08380195e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64944382e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76188202e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08380195e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64944382e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76188202e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08380195e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64944382e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76188202e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08384609e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66400757e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77925568e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09103241e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.61568832e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #129] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44060928e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08384609e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66400757e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77925568e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66427994e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31919441e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #129] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44662018e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08384609e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66400757e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77925568e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39626617e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.31425619e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #129] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.34888992e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08384609e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66400757e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77925568e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.18313522e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.51192703e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #129] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33544327e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 129: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:28:30 [abstract.py:324] It took 0.034354 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12503.75it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 09:28:30 [loggers.py:236] Engine 000: Avg prompt throughput: 46.4 tokens/s, Avg generation throughput: 16.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:32,  4.84it/s, est. speed input: 290.73 toks/s, output: 43.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:08, 17.86it/s, est. speed input: 919.81 toks/s, output: 215.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:09, 16.80it/s, est. speed input: 927.59 toks/s, output: 320.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:12, 11.79it/s, est. speed input: 752.21 toks/s, output: 357.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:08, 17.10it/s, est. speed input: 919.88 toks/s, output: 621.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:05, 26.17it/s, est. speed input: 1177.19 toks/s, output: 1034.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:05, 25.09it/s, est. speed input: 1210.00 toks/s, output: 1195.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:05, 25.05it/s, est. speed input: 1237.69 toks/s, output: 1339.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 35.01it/s, est. speed input: 1454.73 toks/s, output: 1883.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 34.43it/s, est. speed input: 1495.45 toks/s, output: 2081.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:03, 29.67it/s, est. speed input: 1474.27 toks/s, output: 2209.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:02<00:04, 23.38it/s, est. speed input: 1402.48 toks/s, output: 2274.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:02<00:05, 18.57it/s, est. speed input: 1315.62 toks/s, output: 2290.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:03<00:17,  6.09it/s, est. speed input: 843.04 toks/s, output: 1628.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:04<00:17,  6.16it/s, est. speed input: 809.21 toks/s, output: 1702.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:04<00:15,  6.63it/s, est. speed input: 796.11 toks/s, output: 1814.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:04<00:17,  5.85it/s, est. speed input: 742.93 toks/s, output: 1836.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:05<00:28,  3.54it/s, est. speed input: 630.47 toks/s, output: 1635.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:05<00:27,  3.66it/s, est. speed input: 616.31 toks/s, output: 1676.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:06<00:28,  3.39it/s, est. speed input: 588.85 toks/s, output: 1677.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:08<01:16,  1.27it/s, est. speed input: 426.86 toks/s, output: 1301.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:09<00:58,  1.62it/s, est. speed input: 408.77 toks/s, output: 1408.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:28:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11315.8 tokens/s, Running: 95 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:10<01:05,  1.43it/s, est. speed input: 375.49 toks/s, output: 1378.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:12<01:32,  1.00it/s, est. speed input: 321.43 toks/s, output: 1266.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:13<01:21,  1.13it/s, est. speed input: 312.78 toks/s, output: 1317.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:14<01:29,  1.01it/s, est. speed input: 289.35 toks/s, output: 1303.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:16<01:46,  1.18s/it, est. speed input: 262.64 toks/s, output: 1269.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:16<01:18,  1.14it/s, est. speed input: 264.34 toks/s, output: 1361.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:16<00:59,  1.48it/s, est. speed input: 265.44 toks/s, output: 1451.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:17<01:04,  1.36it/s, est. speed input: 255.25 toks/s, output: 1478.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:17<00:52,  1.63it/s, est. speed input: 253.96 toks/s, output: 1553.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:18<00:48,  1.75it/s, est. speed input: 250.62 toks/s, output: 1615.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:18<00:43,  1.91it/s, est. speed input: 248.37 toks/s, output: 1681.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:18<00:43,  1.90it/s, est. speed input: 244.55 toks/s, output: 1735.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:19<00:44,  1.86it/s, est. speed input: 240.63 toks/s, output: 1786.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:20<00:43,  1.88it/s, est. speed input: 237.38 toks/s, output: 1841.48 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:28:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8887.8 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:20<00:33,  2.39it/s, est. speed input: 238.55 toks/s, output: 1929.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:21<00:47,  1.68it/s, est. speed input: 230.03 toks/s, output: 1938.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:22<00:34,  2.19it/s, est. speed input: 227.42 toks/s, output: 2143.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:22<00:35,  2.09it/s, est. speed input: 224.52 toks/s, output: 2191.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:23<00:35,  2.09it/s, est. speed input: 222.58 toks/s, output: 2247.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:24<00:40,  1.82it/s, est. speed input: 217.98 toks/s, output: 2277.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:24<00:32,  2.18it/s, est. speed input: 218.66 toks/s, output: 2358.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:24<00:28,  2.51it/s, est. speed input: 218.96 toks/s, output: 2436.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:24<00:23,  3.01it/s, est. speed input: 219.91 toks/s, output: 2521.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:25<00:27,  2.44it/s, est. speed input: 215.92 toks/s, output: 2619.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:26<00:28,  2.39it/s, est. speed input: 214.47 toks/s, output: 2676.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:26<00:26,  2.48it/s, est. speed input: 212.94 toks/s, output: 2800.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:28<00:37,  1.69it/s, est. speed input: 205.71 toks/s, output: 2777.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<00:35,  1.77it/s, est. speed input: 204.37 toks/s, output: 2833.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:29<00:38,  1.63it/s, est. speed input: 201.16 toks/s, output: 2862.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:30<00:38,  1.57it/s, est. speed input: 198.40 toks/s, output: 2897.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7221.4 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:30<00:41,  1.44it/s, est. speed input: 194.92 toks/s, output: 2920.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:31<00:33,  1.71it/s, est. speed input: 193.24 toks/s, output: 3040.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:31<00:27,  2.08it/s, est. speed input: 194.27 toks/s, output: 3129.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:32<00:18,  2.98it/s, est. speed input: 196.49 toks/s, output: 3309.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:32<00:19,  2.78it/s, est. speed input: 195.62 toks/s, output: 3367.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:34<00:38,  1.37it/s, est. speed input: 186.69 toks/s, output: 3287.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:35<00:45,  1.13it/s, est. speed input: 181.54 toks/s, output: 3270.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:36<00:38,  1.32it/s, est. speed input: 181.05 toks/s, output: 3335.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:23,  2.10it/s, est. speed input: 183.34 toks/s, output: 3524.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:36<00:18,  2.57it/s, est. speed input: 184.43 toks/s, output: 3618.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:37<00:19,  2.42it/s, est. speed input: 183.64 toks/s, output: 3675.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:37<00:22,  2.00it/s, est. speed input: 181.62 toks/s, output: 3708.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:38<00:26,  1.69it/s, est. speed input: 179.32 toks/s, output: 3734.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:39<00:25,  1.72it/s, est. speed input: 177.17 toks/s, output: 3834.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5931.1 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:40<00:25,  1.68it/s, est. speed input: 175.81 toks/s, output: 3878.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:42<00:38,  1.05it/s, est. speed input: 168.98 toks/s, output: 3802.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:42<00:29,  1.36it/s, est. speed input: 169.81 toks/s, output: 3895.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:25,  1.55it/s, est. speed input: 169.61 toks/s, output: 3965.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:45<00:41,  1.10s/it, est. speed input: 162.55 toks/s, output: 3875.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:46<00:37,  1.02s/it, est. speed input: 160.89 toks/s, output: 3912.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:47<00:27,  1.29it/s, est. speed input: 160.18 toks/s, output: 4047.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:47<00:22,  1.50it/s, est. speed input: 160.36 toks/s, output: 4127.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:48<00:29,  1.12it/s, est. speed input: 156.52 toks/s, output: 4105.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:49<00:17,  1.82it/s, est. speed input: 158.42 toks/s, output: 4307.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:49<00:13,  2.07it/s, est. speed input: 158.52 toks/s, output: 4460.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:49<00:11,  2.37it/s, est. speed input: 159.14 toks/s, output: 4551.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4715.0 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:50<00:12,  2.17it/s, est. speed input: 158.46 toks/s, output: 4607.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:50<00:10,  2.47it/s, est. speed input: 158.88 toks/s, output: 4696.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:51<00:08,  2.84it/s, est. speed input: 159.41 toks/s, output: 4787.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:51<00:07,  3.11it/s, est. speed input: 159.82 toks/s, output: 4875.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:06,  3.29it/s, est. speed input: 160.19 toks/s, output: 4960.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:06,  3.40it/s, est. speed input: 160.49 toks/s, output: 5045.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:11,  1.77it/s, est. speed input: 157.89 toks/s, output: 5040.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:53<00:11,  1.81it/s, est. speed input: 157.44 toks/s, output: 5102.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:53<00:08,  2.35it/s, est. speed input: 158.18 toks/s, output: 5203.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:55<00:14,  1.24it/s, est. speed input: 154.43 toks/s, output: 5156.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:58<00:26,  1.56s/it, est. speed input: 146.66 toks/s, output: 4978.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:59<00:20,  1.26s/it, est. speed input: 146.32 toks/s, output: 5047.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3034.3 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:00<00:20,  1.39s/it, est. speed input: 143.26 toks/s, output: 5024.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:02<00:13,  1.04s/it, est. speed input: 142.30 toks/s, output: 5157.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:05<00:19,  1.59s/it, est. speed input: 136.14 toks/s, output: 5022.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:06<00:16,  1.50s/it, est. speed input: 134.49 toks/s, output: 5049.55 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  1.50s/it, est. speed input: 144.42 toks/s, output: 6399.52 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:06<00:00,  2.40it/s, est. speed input: 144.42 toks/s, output: 6399.52 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:41 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:41 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:29:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:29:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:29:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:29:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:29:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:29:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:29:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:29:41 [abstract.py:306] It took 0.597529 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:29:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.2313
+  • Mean Generation Length: 2669.7876
+  • Training FLOPS: 46.32 TFLOPS (11.58 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 320.03s
+  • policy_training: 207.62s (64.9%)
+  • generation: 71.23s (22.3%)
+  • teacher_logprob_inference: 35.96s (11.2%)
+  • policy_training/sharding_data: 4.38s (1.4%)
+  • prepare_for_generation: 3.62s (1.1%)
+  • policy_training/submit_training_futures: 1.57s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.5%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 130: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:33:48 [abstract.py:324] It took 0.050493 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08410410e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66772461e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75633698e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08410410e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66772461e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75633698e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08410410e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66772461e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75633698e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08410410e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66772461e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75633698e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08417461e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67367210e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78123093e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08924484e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.59289360e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #130] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44195251e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08417461e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67367210e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78123093e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65267792e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30430851e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #130] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.41715240e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08417461e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67367210e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78123093e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39625092e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.85173130e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #130] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.30906792e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08417461e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67367210e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78123093e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17363052e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.51151371e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #130] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33768005e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 130: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:33:51 [abstract.py:324] It took 0.029949 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13820.07it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:33:51 [loggers.py:236] Engine 000: Avg prompt throughput: 37.4 tokens/s, Avg generation throughput: 61.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.64it/s, est. speed input: 222.07 toks/s, output: 50.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:09, 15.79it/s, est. speed input: 805.19 toks/s, output: 236.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:08, 18.94it/s, est. speed input: 970.51 toks/s, output: 388.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 21.52it/s, est. speed input: 1093.58 toks/s, output: 534.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:06, 23.91it/s, est. speed input: 1198.34 toks/s, output: 679.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 32.58it/s, est. speed input: 1460.78 toks/s, output: 1030.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 36.67it/s, est. speed input: 1617.28 toks/s, output: 1306.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 59.25it/s, est. speed input: 2168.86 toks/s, output: 2201.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:01, 64.34it/s, est. speed input: 2432.32 toks/s, output: 2784.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 65.40it/s, est. speed input: 2580.26 toks/s, output: 3221.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 61.17it/s, est. speed input: 2633.29 toks/s, output: 3565.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:01<00:01, 53.28it/s, est. speed input: 2647.75 toks/s, output: 4034.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:08<00:19,  3.98it/s, est. speed input: 611.55 toks/s, output: 1164.99 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10300.7 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:11<00:27,  2.70it/s, est. speed input: 439.80 toks/s, output: 1194.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:15<00:37,  1.91it/s, est. speed input: 339.87 toks/s, output: 1196.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:16<00:35,  1.97it/s, est. speed input: 330.13 toks/s, output: 1362.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:18<00:41,  1.64it/s, est. speed input: 296.56 toks/s, output: 1425.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7949.2 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:20<00:42,  1.55it/s, est. speed input: 279.33 toks/s, output: 1534.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:20<00:39,  1.66it/s, est. speed input: 278.51 toks/s, output: 1627.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:21<00:35,  1.79it/s, est. speed input: 277.34 toks/s, output: 1718.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:21<00:34,  1.82it/s, est. speed input: 273.74 toks/s, output: 1791.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:23<00:35,  1.68it/s, est. speed input: 258.45 toks/s, output: 1965.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:25<00:38,  1.52it/s, est. speed input: 246.66 toks/s, output: 2059.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:25<00:32,  1.75it/s, est. speed input: 247.74 toks/s, output: 2161.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:26<00:37,  1.51it/s, est. speed input: 240.18 toks/s, output: 2190.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:26<00:31,  1.72it/s, est. speed input: 239.84 toks/s, output: 2281.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:27<00:29,  1.81it/s, est. speed input: 237.96 toks/s, output: 2357.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:27<00:28,  1.83it/s, est. speed input: 235.66 toks/s, output: 2426.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:27<00:23,  2.22it/s, est. speed input: 236.38 toks/s, output: 2525.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:28<00:24,  2.03it/s, est. speed input: 231.51 toks/s, output: 2654.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:29<00:22,  2.18it/s, est. speed input: 229.30 toks/s, output: 2809.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6759.5 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:31<00:33,  1.42it/s, est. speed input: 219.49 toks/s, output: 2780.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:31<00:27,  1.70it/s, est. speed input: 219.94 toks/s, output: 2878.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:31<00:21,  2.10it/s, est. speed input: 220.87 toks/s, output: 2980.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:33<00:40,  1.08it/s, est. speed input: 208.52 toks/s, output: 2905.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:34<00:33,  1.28it/s, est. speed input: 207.82 toks/s, output: 2986.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:34<00:31,  1.33it/s, est. speed input: 205.54 toks/s, output: 3044.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:36<00:35,  1.15it/s, est. speed input: 200.64 toks/s, output: 3063.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:27,  1.45it/s, est. speed input: 200.95 toks/s, output: 3158.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:38<00:38,  1.01it/s, est. speed input: 193.54 toks/s, output: 3133.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:39<00:39,  1.04s/it, est. speed input: 189.31 toks/s, output: 3158.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5278.1 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:40<00:39,  1.07s/it, est. speed input: 185.44 toks/s, output: 3186.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:40<00:30,  1.18it/s, est. speed input: 185.49 toks/s, output: 3278.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:41<00:20,  1.67it/s, est. speed input: 185.71 toks/s, output: 3463.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:41<00:17,  1.90it/s, est. speed input: 185.86 toks/s, output: 3556.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:21,  1.48it/s, est. speed input: 182.52 toks/s, output: 3583.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:43<00:19,  1.58it/s, est. speed input: 181.72 toks/s, output: 3658.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:43<00:16,  1.87it/s, est. speed input: 181.95 toks/s, output: 3754.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:43<00:11,  2.43it/s, est. speed input: 182.89 toks/s, output: 3864.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:43<00:11,  2.52it/s, est. speed input: 182.74 toks/s, output: 3951.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:44<00:09,  2.79it/s, est. speed input: 183.01 toks/s, output: 4046.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:44<00:11,  2.24it/s, est. speed input: 181.64 toks/s, output: 4107.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:44<00:08,  2.83it/s, est. speed input: 182.45 toks/s, output: 4214.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:10,  2.33it/s, est. speed input: 181.34 toks/s, output: 4278.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:08,  2.84it/s, est. speed input: 182.03 toks/s, output: 4382.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:46<00:07,  2.84it/s, est. speed input: 181.94 toks/s, output: 4469.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:10,  1.92it/s, est. speed input: 179.69 toks/s, output: 4503.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:47<00:08,  2.31it/s, est. speed input: 180.08 toks/s, output: 4603.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:49<00:15,  1.21it/s, est. speed input: 174.91 toks/s, output: 4561.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:14,  1.22it/s, est. speed input: 173.36 toks/s, output: 4611.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3757.6 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:50<00:12,  1.34it/s, est. speed input: 172.55 toks/s, output: 4681.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:15,  1.06it/s, est. speed input: 169.06 toks/s, output: 4678.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:52<00:12,  1.19it/s, est. speed input: 168.28 toks/s, output: 4750.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:11,  1.21it/s, est. speed input: 166.88 toks/s, output: 4804.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:53<00:10,  1.23it/s, est. speed input: 165.61 toks/s, output: 4861.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:54<00:07,  1.60it/s, est. speed input: 166.20 toks/s, output: 4971.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:54<00:05,  1.86it/s, est. speed input: 166.27 toks/s, output: 5067.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:55<00:07,  1.27it/s, est. speed input: 163.28 toks/s, output: 5072.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:56<00:07,  1.23it/s, est. speed input: 161.86 toks/s, output: 5122.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:05,  1.42it/s, est. speed input: 161.70 toks/s, output: 5214.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:06,  1.13it/s, est. speed input: 159.10 toks/s, output: 5227.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:06,  1.07s/it, est. speed input: 156.19 toks/s, output: 5232.49 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2170.6 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:04,  1.11it/s, est. speed input: 155.87 toks/s, output: 5322.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:00<00:02,  1.51it/s, est. speed input: 156.61 toks/s, output: 5448.52 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  1.51it/s, est. speed input: 160.60 toks/s, output: 5989.02 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  2.64it/s, est. speed input: 160.60 toks/s, output: 5989.02 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:54 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:54 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:34:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:34:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:34:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:34:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:34:55 [abstract.py:306] It took 0.534627 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:34:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:34:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:34:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:34:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0108
+  • Training Accuracy: 0.4375
+  • Mean Generation Length: 2269.1938
+  • Training FLOPS: 43.76 TFLOPS (10.94 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 290.45s
+  • policy_training: 187.63s (64.6%)
+  • generation: 64.49s (22.2%)
+  • teacher_logprob_inference: 32.73s (11.3%)
+  • policy_training/sharding_data: 4.38s (1.5%)
+  • prepare_for_generation: 3.69s (1.3%)
+  • policy_training/submit_training_futures: 1.85s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.5%)
+  • training_prep: 1.29s (0.4%)
+  • teacher_logprob_inference_prep: 0.46s (0.2%)
+  • get_topk_logits/shard_data: 0.05s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 27/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 131: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:38:38 [abstract.py:324] It took 0.047552 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08439414e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67447433e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75482712e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08439414e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67447433e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75482712e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08439414e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67447433e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75482712e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08439414e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67447433e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75482712e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08451758e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66273994e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77058716e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08538876e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.54204416e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #131] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43999481e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08451758e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66273994e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77058716e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66500702e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29253006e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #131] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.38550606e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08451758e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66273994e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77058716e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39824677e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.06170559e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #131] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.28724899e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08451758e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66273994e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77058716e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17657204e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.51193218e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #131] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33752899e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 131: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:38:41 [abstract.py:324] It took 0.030056 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13881.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 09:38:41 [loggers.py:236] Engine 000: Avg prompt throughput: 42.1 tokens/s, Avg generation throughput: 3.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.08it/s, est. speed input: 122.45 toks/s, output: 26.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 23.66it/s, est. speed input: 1119.26 toks/s, output: 301.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 36.31it/s, est. speed input: 1638.29 toks/s, output: 582.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 34.65it/s, est. speed input: 1720.16 toks/s, output: 829.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 36.83it/s, est. speed input: 1826.40 toks/s, output: 1041.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 38.01it/s, est. speed input: 1900.24 toks/s, output: 1270.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 40.78it/s, est. speed input: 1987.55 toks/s, output: 1519.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:03, 35.53it/s, est. speed input: 1944.77 toks/s, output: 1682.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 37.16it/s, est. speed input: 1994.80 toks/s, output: 1981.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 38.39it/s, est. speed input: 2031.44 toks/s, output: 2251.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:02<00:03, 27.48it/s, est. speed input: 1868.45 toks/s, output: 2325.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:04<00:18,  5.08it/s, est. speed input: 845.67 toks/s, output: 1240.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:06<00:23,  3.93it/s, est. speed input: 680.21 toks/s, output: 1245.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:07<00:27,  3.28it/s, est. speed input: 593.47 toks/s, output: 1255.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:09<00:39,  2.21it/s, est. speed input: 470.66 toks/s, output: 1183.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:10<00:42,  2.02it/s, est. speed input: 439.68 toks/s, output: 1199.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:38:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10894.3 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:11<00:48,  1.74it/s, est. speed input: 404.99 toks/s, output: 1198.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:13<01:12,  1.15it/s, est. speed input: 341.55 toks/s, output: 1106.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:13<01:02,  1.34it/s, est. speed input: 339.65 toks/s, output: 1193.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:16<01:45,  1.29s/it, est. speed input: 279.58 toks/s, output: 1076.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:17<01:31,  1.13s/it, est. speed input: 272.80 toks/s, output: 1145.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8711.7 tokens/s, Running: 80 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:20<01:44,  1.32s/it, est. speed input: 236.48 toks/s, output: 1166.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:21<01:40,  1.29s/it, est. speed input: 226.39 toks/s, output: 1209.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:23<01:40,  1.30s/it, est. speed input: 215.98 toks/s, output: 1246.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:24<01:09,  1.08it/s, est. speed input: 213.93 toks/s, output: 1414.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:25<01:12,  1.02it/s, est. speed input: 206.57 toks/s, output: 1455.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:25<00:56,  1.29it/s, est. speed input: 207.92 toks/s, output: 1553.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:26<00:43,  1.63it/s, est. speed input: 206.47 toks/s, output: 1715.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:27<00:50,  1.40it/s, est. speed input: 200.66 toks/s, output: 1753.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<01:18,  1.14s/it, est. speed input: 186.24 toks/s, output: 1713.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:30<01:06,  1.02it/s, est. speed input: 184.95 toks/s, output: 1787.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:30<00:49,  1.35it/s, est. speed input: 186.30 toks/s, output: 1886.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7295.8 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:30<00:45,  1.47it/s, est. speed input: 185.08 toks/s, output: 1958.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:31<00:42,  1.53it/s, est. speed input: 183.59 toks/s, output: 2026.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<00:54,  1.17it/s, est. speed input: 177.90 toks/s, output: 2047.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:33<00:46,  1.37it/s, est. speed input: 177.43 toks/s, output: 2124.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:33<00:34,  1.82it/s, est. speed input: 178.69 toks/s, output: 2221.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:33<00:24,  2.41it/s, est. speed input: 179.64 toks/s, output: 2394.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:35<00:36,  1.60it/s, est. speed input: 174.77 toks/s, output: 2411.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:35<00:23,  2.39it/s, est. speed input: 176.92 toks/s, output: 2602.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:36<00:35,  1.57it/s, est. speed input: 171.91 toks/s, output: 2608.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:36<00:30,  1.82it/s, est. speed input: 172.26 toks/s, output: 2693.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:38<00:38,  1.39it/s, est. speed input: 166.72 toks/s, output: 2764.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6333.3 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:41<00:54,  1.06s/it, est. speed input: 159.31 toks/s, output: 2722.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:41<00:44,  1.15it/s, est. speed input: 159.69 toks/s, output: 2808.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:43<01:00,  1.21s/it, est. speed input: 153.00 toks/s, output: 2772.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:43<00:47,  1.03it/s, est. speed input: 153.18 toks/s, output: 2856.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:44<00:36,  1.30it/s, est. speed input: 153.78 toks/s, output: 2947.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:44<00:30,  1.52it/s, est. speed input: 153.87 toks/s, output: 3027.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:44<00:25,  1.80it/s, est. speed input: 154.18 toks/s, output: 3111.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:45<00:32,  1.41it/s, est. speed input: 151.88 toks/s, output: 3143.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:46<00:25,  1.74it/s, est. speed input: 152.41 toks/s, output: 3232.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:46<00:20,  2.06it/s, est. speed input: 152.76 toks/s, output: 3318.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:46<00:15,  2.67it/s, est. speed input: 153.69 toks/s, output: 3416.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:46<00:14,  2.90it/s, est. speed input: 154.06 toks/s, output: 3502.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:47<00:13,  3.04it/s, est. speed input: 154.38 toks/s, output: 3586.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:47<00:16,  2.34it/s, est. speed input: 153.49 toks/s, output: 3643.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:48<00:20,  1.85it/s, est. speed input: 152.20 toks/s, output: 3689.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:49<00:28,  1.32it/s, est. speed input: 149.53 toks/s, output: 3702.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:50<00:23,  1.55it/s, est. speed input: 149.59 toks/s, output: 3780.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5111.0 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:50<00:18,  1.87it/s, est. speed input: 150.02 toks/s, output: 3867.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:50<00:13,  2.47it/s, est. speed input: 150.94 toks/s, output: 3966.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:52<00:17,  1.80it/s, est. speed input: 149.04 toks/s, output: 4066.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:54<00:22,  1.36it/s, est. speed input: 145.68 toks/s, output: 4127.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:54<00:22,  1.31it/s, est. speed input: 144.47 toks/s, output: 4171.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:56<00:28,  1.04s/it, est. speed input: 140.72 toks/s, output: 4143.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:57<00:21,  1.21it/s, est. speed input: 140.19 toks/s, output: 4286.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:58<00:19,  1.30it/s, est. speed input: 139.87 toks/s, output: 4355.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:58<00:11,  2.03it/s, est. speed input: 141.64 toks/s, output: 4568.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:59<00:10,  2.04it/s, est. speed input: 141.38 toks/s, output: 4714.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3947.7 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:02<00:19,  1.04it/s, est. speed input: 135.97 toks/s, output: 4615.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:02<00:15,  1.20it/s, est. speed input: 136.07 toks/s, output: 4701.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:03<00:14,  1.22it/s, est. speed input: 135.38 toks/s, output: 4758.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:05<00:18,  1.09s/it, est. speed input: 132.40 toks/s, output: 4736.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:06<00:18,  1.16s/it, est. speed input: 130.65 toks/s, output: 4755.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:07<00:13,  1.07it/s, est. speed input: 130.91 toks/s, output: 4847.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:07<00:12,  1.10it/s, est. speed input: 130.17 toks/s, output: 4903.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:08<00:11,  1.14it/s, est. speed input: 129.56 toks/s, output: 4963.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:09<00:08,  1.38it/s, est. speed input: 129.79 toks/s, output: 5055.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:09<00:07,  1.52it/s, est. speed input: 129.70 toks/s, output: 5136.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  1.52it/s, est. speed input: 139.24 toks/s, output: 6430.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  2.30it/s, est. speed input: 139.24 toks/s, output: 6430.40 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:55 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:55 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:39:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:39:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:39:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:39:55 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:39:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:39:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:39:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:39:55 [abstract.py:306] It took 0.770620 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:39:55 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0098
+  • Training Accuracy: 0.2188
+  • Mean Generation Length: 2798.6938
+  • Training FLOPS: 49.25 TFLOPS (12.31 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 324.38s
+  • policy_training: 206.94s (63.8%)
+  • generation: 74.43s (22.9%)
+  • teacher_logprob_inference: 37.75s (11.6%)
+  • policy_training/sharding_data: 4.36s (1.3%)
+  • prepare_for_generation: 3.59s (1.1%)
+  • policy_training/submit_training_futures: 1.86s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.5%)
+  • training_prep: 0.87s (0.3%)
+  • teacher_logprob_inference_prep: 0.63s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 132: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:03 [abstract.py:324] It took 0.050152 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08450273e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67701111e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75864639e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08450273e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67701111e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75864639e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08450273e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67701111e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75864639e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08450273e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67701111e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75864639e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08449238e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66426392e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75887299e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09757938e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.48142910e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #132] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43925171e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08449238e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66426392e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75887299e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66656036e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29189606e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #132] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.36704292e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08449238e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66426392e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75887299e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40744476e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.13812447e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #132] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.30202484e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08449238e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66426392e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.75887299e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16454544e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50804501e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #132] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33590393e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 132: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:06 [abstract.py:324] It took 0.030259 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:06 [loggers.py:236] Engine 000: Avg prompt throughput: 36.5 tokens/s, Avg generation throughput: 93.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13397.66it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:24,  6.40it/s, est. speed input: 371.53 toks/s, output: 76.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 30.13it/s, est. speed input: 1597.36 toks/s, output: 436.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:03, 39.26it/s, est. speed input: 2052.60 toks/s, output: 756.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 40.31it/s, est. speed input: 2169.60 toks/s, output: 1018.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 34.11it/s, est. speed input: 2021.05 toks/s, output: 1180.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:03, 42.62it/s, est. speed input: 2278.41 toks/s, output: 1652.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 41.33it/s, est. speed input: 2285.50 toks/s, output: 1874.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 46.07it/s, est. speed input: 2410.61 toks/s, output: 2261.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 54.81it/s, est. speed input: 2613.90 toks/s, output: 2804.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 53.27it/s, est. speed input: 2648.24 toks/s, output: 3118.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 57.79it/s, est. speed input: 2765.87 toks/s, output: 3570.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:02, 41.54it/s, est. speed input: 2575.45 toks/s, output: 3605.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:02, 32.35it/s, est. speed input: 2390.25 toks/s, output: 3615.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:02<00:04, 18.71it/s, est. speed input: 1953.64 toks/s, output: 3274.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:04<00:10,  7.46it/s, est. speed input: 1244.36 toks/s, output: 2380.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:05<00:14,  4.98it/s, est. speed input: 956.30 toks/s, output: 2077.15 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:07<00:22,  3.20it/s, est. speed input: 735.78 toks/s, output: 1792.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:08<00:27,  2.57it/s, est. speed input: 626.78 toks/s, output: 1718.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:08<00:26,  2.64it/s, est. speed input: 612.66 toks/s, output: 1783.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10039.3 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:10<00:34,  1.98it/s, est. speed input: 542.90 toks/s, output: 1685.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:11<00:40,  1.67it/s, est. speed input: 497.96 toks/s, output: 1650.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:12<00:40,  1.61it/s, est. speed input: 454.75 toks/s, output: 1707.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:12<00:33,  1.91it/s, est. speed input: 455.95 toks/s, output: 1815.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:40,  1.57it/s, est. speed input: 425.50 toks/s, output: 1798.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:16<01:06,  1.07s/it, est. speed input: 364.38 toks/s, output: 1647.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<01:45,  1.72s/it, est. speed input: 300.15 toks/s, output: 1465.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7824.4 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:21<01:39,  1.66s/it, est. speed input: 282.18 toks/s, output: 1486.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<01:25,  1.46s/it, est. speed input: 273.40 toks/s, output: 1547.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:22<01:08,  1.17s/it, est. speed input: 270.58 toks/s, output: 1638.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<01:05,  1.15s/it, est. speed input: 260.84 toks/s, output: 1685.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:24<00:36,  1.50it/s, est. speed input: 264.18 toks/s, output: 1916.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:32,  1.66it/s, est. speed input: 262.32 toks/s, output: 2007.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:32,  1.61it/s, est. speed input: 257.68 toks/s, output: 2074.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:31,  1.67it/s, est. speed input: 254.68 toks/s, output: 2152.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:20,  2.47it/s, est. speed input: 256.20 toks/s, output: 2369.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:26<00:24,  1.98it/s, est. speed input: 250.55 toks/s, output: 2417.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:28<00:38,  1.26it/s, est. speed input: 238.09 toks/s, output: 2398.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:28<00:28,  1.63it/s, est. speed input: 239.32 toks/s, output: 2511.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6415.7 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:30<00:42,  1.09it/s, est. speed input: 227.73 toks/s, output: 2490.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:30<00:31,  1.44it/s, est. speed input: 228.71 toks/s, output: 2601.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:30<00:18,  2.28it/s, est. speed input: 230.84 toks/s, output: 2822.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:30<00:15,  2.64it/s, est. speed input: 231.27 toks/s, output: 2926.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:32<00:29,  1.38it/s, est. speed input: 220.83 toks/s, output: 2892.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:32<00:22,  1.78it/s, est. speed input: 221.87 toks/s, output: 3004.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:34<00:35,  1.11it/s, est. speed input: 212.14 toks/s, output: 2971.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:34<00:29,  1.30it/s, est. speed input: 211.20 toks/s, output: 3056.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:35<00:13,  2.53it/s, est. speed input: 214.64 toks/s, output: 3398.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:36<00:21,  1.59it/s, est. speed input: 207.37 toks/s, output: 3381.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:38<00:29,  1.13it/s, est. speed input: 199.54 toks/s, output: 3350.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:38<00:23,  1.39it/s, est. speed input: 199.99 toks/s, output: 3454.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4920.7 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:40<00:33,  1.07s/it, est. speed input: 191.25 toks/s, output: 3401.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:41<00:27,  1.08it/s, est. speed input: 190.19 toks/s, output: 3480.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:41<00:24,  1.17it/s, est. speed input: 188.60 toks/s, output: 3548.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:43<00:26,  1.05it/s, est. speed input: 184.80 toks/s, output: 3575.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:43<00:22,  1.21it/s, est. speed input: 183.94 toks/s, output: 3656.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:44<00:23,  1.11it/s, est. speed input: 180.81 toks/s, output: 3693.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:46<00:19,  1.21it/s, est. speed input: 177.64 toks/s, output: 3822.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:47<00:20,  1.10it/s, est. speed input: 174.51 toks/s, output: 3854.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:48<00:14,  1.50it/s, est. speed input: 174.72 toks/s, output: 4054.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:48<00:12,  1.66it/s, est. speed input: 174.58 toks/s, output: 4149.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:48<00:09,  2.04it/s, est. speed input: 175.29 toks/s, output: 4264.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:44:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3666.0 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:52<00:27,  1.51s/it, est. speed input: 161.83 toks/s, output: 4039.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:53<00:19,  1.16s/it, est. speed input: 162.24 toks/s, output: 4151.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:18,  1.16s/it, est. speed input: 159.92 toks/s, output: 4194.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:13,  1.11it/s, est. speed input: 160.28 toks/s, output: 4306.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:54<00:06,  1.89it/s, est. speed input: 162.04 toks/s, output: 4555.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:05,  2.05it/s, est. speed input: 162.09 toks/s, output: 4657.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:56<00:08,  1.30it/s, est. speed input: 158.62 toks/s, output: 4659.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:57<00:08,  1.20it/s, est. speed input: 156.88 toks/s, output: 4712.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:45:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2482.6 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:00<00:11,  1.27s/it, est. speed input: 151.63 toks/s, output: 4660.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  1.27s/it, est. speed input: 160.63 toks/s, output: 5886.87 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  2.66it/s, est. speed input: 160.63 toks/s, output: 5886.87 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:45:09 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:45:09 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:45:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:45:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:45:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:45:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:45:09 [abstract.py:306] It took 0.529839 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:45:09 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:45:09 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:45:09 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:45:09 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.3563
+  • Mean Generation Length: 2211.8125
+  • Training FLOPS: 46.19 TFLOPS (11.55 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 273.88s
+  • policy_training: 174.47s (63.7%)
+  • generation: 63.85s (23.3%)
+  • teacher_logprob_inference: 30.54s (11.1%)
+  • policy_training/sharding_data: 4.37s (1.6%)
+  • prepare_for_generation: 3.57s (1.3%)
+  • policy_training/submit_training_futures: 1.81s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.6%)
+  • training_prep: 0.89s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 133: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:48:37 [abstract.py:324] It took 0.048825 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08489727e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67988892e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75747757e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08489727e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67988892e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75747757e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08489727e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67988892e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75747757e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08489727e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67988892e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.75747757e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08498281e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64590912e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77657318e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09420891e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.44345474e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #133] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44037460e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08498281e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64590912e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77657318e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63854980e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28793087e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #133] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.33440208e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08498281e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64590912e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77657318e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39825745e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.12778759e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #133] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.29823685e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08498281e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64590912e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77657318e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17045898e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50779037e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #133] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33652039e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 133: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:48:40 [abstract.py:324] It took 0.030018 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13877.22it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:48:40 [loggers.py:236] Engine 000: Avg prompt throughput: 45.6 tokens/s, Avg generation throughput: 1.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:30,  1.76it/s, est. speed input: 107.62 toks/s, output: 12.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:16,  9.36it/s, est. speed input: 456.87 toks/s, output: 104.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:16,  9.48it/s, est. speed input: 499.92 toks/s, output: 179.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:01<00:13, 11.39it/s, est. speed input: 579.64 toks/s, output: 297.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:06, 21.42it/s, est. speed input: 864.14 toks/s, output: 661.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:04, 33.06it/s, est. speed input: 1155.09 toks/s, output: 1082.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:03, 40.66it/s, est. speed input: 1357.95 toks/s, output: 1441.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 52.04it/s, est. speed input: 1616.94 toks/s, output: 1929.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 69.29it/s, est. speed input: 1974.89 toks/s, output: 2663.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 59.18it/s, est. speed input: 2053.35 toks/s, output: 3010.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:01, 65.09it/s, est. speed input: 2232.52 toks/s, output: 3577.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:03<00:05, 13.76it/s, est. speed input: 1313.53 toks/s, output: 2317.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:09<00:18,  3.85it/s, est. speed input: 585.35 toks/s, output: 1444.32 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 09:48:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9893.4 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:12<00:25,  2.72it/s, est. speed input: 444.30 toks/s, output: 1479.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:13<00:22,  2.95it/s, est. speed input: 439.32 toks/s, output: 1755.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:14<00:25,  2.45it/s, est. speed input: 397.09 toks/s, output: 1783.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:15<00:23,  2.60it/s, est. speed input: 391.19 toks/s, output: 1953.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:18<00:33,  1.76it/s, est. speed input: 339.06 toks/s, output: 1885.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:18<00:31,  1.82it/s, est. speed input: 334.82 toks/s, output: 1961.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:30,  1.89it/s, est. speed input: 330.93 toks/s, output: 2036.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:19<00:24,  2.27it/s, est. speed input: 329.57 toks/s, output: 2221.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:19<00:22,  2.35it/s, est. speed input: 326.78 toks/s, output: 2300.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:20<00:19,  2.70it/s, est. speed input: 324.57 toks/s, output: 2475.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7713.0 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:20<00:17,  2.96it/s, est. speed input: 324.41 toks/s, output: 2569.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:20<00:17,  2.84it/s, est. speed input: 321.10 toks/s, output: 2638.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:21<00:22,  2.18it/s, est. speed input: 311.93 toks/s, output: 2659.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:21<00:17,  2.67it/s, est. speed input: 312.99 toks/s, output: 2763.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:14,  3.13it/s, est. speed input: 313.41 toks/s, output: 2861.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:22<00:14,  3.21it/s, est. speed input: 312.06 toks/s, output: 2943.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:24<00:22,  1.88it/s, est. speed input: 291.86 toks/s, output: 3022.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:24<00:20,  2.03it/s, est. speed input: 290.26 toks/s, output: 3100.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:25<00:21,  1.88it/s, est. speed input: 285.09 toks/s, output: 3139.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:25<00:17,  2.29it/s, est. speed input: 285.88 toks/s, output: 3243.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:25<00:11,  3.22it/s, est. speed input: 287.72 toks/s, output: 3451.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:27<00:23,  1.59it/s, est. speed input: 272.03 toks/s, output: 3359.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:16,  2.07it/s, est. speed input: 271.34 toks/s, output: 3539.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:28<00:12,  2.63it/s, est. speed input: 271.89 toks/s, output: 3735.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:29<00:10,  2.83it/s, est. speed input: 269.25 toks/s, output: 3975.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:30<00:11,  2.61it/s, est. speed input: 266.55 toks/s, output: 4030.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:09,  2.84it/s, est. speed input: 266.66 toks/s, output: 4124.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5498.9 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:30<00:10,  2.54it/s, est. speed input: 263.95 toks/s, output: 4177.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:31<00:15,  1.70it/s, est. speed input: 255.87 toks/s, output: 4144.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:32<00:10,  2.18it/s, est. speed input: 255.42 toks/s, output: 4325.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:32<00:10,  2.21it/s, est. speed input: 253.89 toks/s, output: 4394.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:34<00:13,  1.65it/s, est. speed input: 247.53 toks/s, output: 4381.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:35<00:16,  1.27it/s, est. speed input: 240.02 toks/s, output: 4347.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:36<00:16,  1.20it/s, est. speed input: 235.46 toks/s, output: 4363.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:36<00:09,  1.89it/s, est. speed input: 237.04 toks/s, output: 4591.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:38<00:10,  1.57it/s, est. speed input: 230.16 toks/s, output: 4653.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3407.5 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:41<00:17,  1.20s/it, est. speed input: 213.59 toks/s, output: 4424.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:43<00:20,  1.50s/it, est. speed input: 202.96 toks/s, output: 4315.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:09,  1.22it/s, est. speed input: 205.02 toks/s, output: 4686.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:07,  1.31it/s, est. speed input: 203.99 toks/s, output: 4771.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:07,  1.23it/s, est. speed input: 200.87 toks/s, output: 4808.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:06,  1.29it/s, est. speed input: 199.39 toks/s, output: 4885.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:47<00:06,  1.10it/s, est. speed input: 195.14 toks/s, output: 4894.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:48<00:04,  1.38it/s, est. speed input: 195.52 toks/s, output: 5017.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:49<00:04,  1.16it/s, est. speed input: 191.88 toks/s, output: 5039.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:49<00:02,  1.38it/s, est. speed input: 191.72 toks/s, output: 5149.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:50<00:02,  1.36it/s, est. speed input: 190.01 toks/s, output: 5220.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2052.8 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:51<00:01,  1.09it/s, est. speed input: 186.21 toks/s, output: 5236.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:52<00:00,  1.05it/s, est. speed input: 183.70 toks/s, output: 5286.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  1.36it/s, est. speed input: 184.14 toks/s, output: 5419.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  1.36it/s, est. speed input: 184.14 toks/s, output: 5419.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  3.02it/s, est. speed input: 184.14 toks/s, output: 5419.45 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:35 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:35 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:49:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:49:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:49:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:49:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:49:36 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:49:36 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:49:36 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:49:36 [abstract.py:306] It took 0.647058 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:49:36 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0125
+  • Training Accuracy: 0.4313
+  • Mean Generation Length: 1796.8063
+  • Training FLOPS: 40.97 TFLOPS (10.24 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 245.66s
+  • policy_training: 157.04s (63.9%)
+  • generation: 56.33s (22.9%)
+  • teacher_logprob_inference: 27.03s (11.0%)
+  • policy_training/sharding_data: 4.33s (1.8%)
+  • prepare_for_generation: 3.59s (1.5%)
+  • policy_training/submit_training_futures: 1.82s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.87s (0.4%)
+  • teacher_logprob_inference_prep: 0.63s (0.3%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 134: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:52:43 [abstract.py:324] It took 0.055483 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08518047e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68167267e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76545105e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08518047e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68167267e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76545105e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08518047e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68167267e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76545105e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08518047e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68167267e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76545105e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08526836e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66813583e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78267746e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09420567e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.36364222e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #134] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43710602e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08526836e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66813583e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78267746e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62453766e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30383196e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #134] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.36493149e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08526836e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66813583e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78267746e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39650192e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.48237419e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #134] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.27032776e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08526836e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66813583e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78267746e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15386429e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50132599e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #134] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33660477e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 134: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:52:46 [abstract.py:324] It took 0.028278 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13853.16it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 09:52:46 [loggers.py:236] Engine 000: Avg prompt throughput: 49.9 tokens/s, Avg generation throughput: 8.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:33,  4.81it/s, est. speed input: 302.95 toks/s, output: 81.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 16.99it/s, est. speed input: 950.24 toks/s, output: 367.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:08, 17.17it/s, est. speed input: 982.51 toks/s, output: 455.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:08, 17.28it/s, est. speed input: 998.43 toks/s, output: 546.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 26.21it/s, est. speed input: 1295.54 toks/s, output: 935.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:02, 51.14it/s, est. speed input: 1992.86 toks/s, output: 1866.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 55.54it/s, est. speed input: 2225.08 toks/s, output: 2272.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 54.24it/s, est. speed input: 2338.64 toks/s, output: 2608.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:01, 78.48it/s, est. speed input: 2897.64 toks/s, output: 3747.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 66.48it/s, est. speed input: 2904.87 toks/s, output: 4088.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 68.64it/s, est. speed input: 3019.87 toks/s, output: 4583.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:01<00:01, 48.44it/s, est. speed input: 2806.58 toks/s, output: 4582.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:03<00:05, 12.25it/s, est. speed input: 1547.23 toks/s, output: 2806.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:08<00:19,  3.44it/s, est. speed input: 653.57 toks/s, output: 1527.56 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 09:52:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9683.0 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:29,  2.12it/s, est. speed input: 440.24 toks/s, output: 1424.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:13<00:24,  2.43it/s, est. speed input: 441.24 toks/s, output: 1750.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:14<00:23,  2.43it/s, est. speed input: 425.23 toks/s, output: 1893.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:16<00:28,  1.97it/s, est. speed input: 382.49 toks/s, output: 1916.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:17<00:27,  1.97it/s, est. speed input: 374.57 toks/s, output: 1983.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:17<00:24,  2.20it/s, est. speed input: 375.56 toks/s, output: 2094.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:17<00:24,  2.14it/s, est. speed input: 368.00 toks/s, output: 2157.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:18<00:28,  1.79it/s, est. speed input: 352.41 toks/s, output: 2171.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:19<00:24,  2.05it/s, est. speed input: 351.34 toks/s, output: 2269.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:19<00:20,  2.44it/s, est. speed input: 351.87 toks/s, output: 2376.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:19<00:20,  2.34it/s, est. speed input: 346.50 toks/s, output: 2442.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7338.5 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:19<00:19,  2.47it/s, est. speed input: 343.57 toks/s, output: 2526.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:20<00:16,  2.83it/s, est. speed input: 342.90 toks/s, output: 2623.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:20<00:15,  2.98it/s, est. speed input: 341.04 toks/s, output: 2711.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:22<00:32,  1.37it/s, est. speed input: 317.49 toks/s, output: 2628.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:23<00:35,  1.23it/s, est. speed input: 306.17 toks/s, output: 2638.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:23<00:20,  2.01it/s, est. speed input: 308.15 toks/s, output: 2862.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:23<00:16,  2.49it/s, est. speed input: 309.27 toks/s, output: 2975.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:23<00:15,  2.62it/s, est. speed input: 307.68 toks/s, output: 3060.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:24<00:15,  2.51it/s, est. speed input: 304.63 toks/s, output: 3130.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:24<00:12,  3.04it/s, est. speed input: 303.81 toks/s, output: 3319.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:25<00:15,  2.32it/s, est. speed input: 297.07 toks/s, output: 3347.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:26<00:17,  2.00it/s, est. speed input: 291.50 toks/s, output: 3385.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:26<00:11,  2.80it/s, est. speed input: 292.44 toks/s, output: 3598.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:26<00:10,  3.14it/s, est. speed input: 292.65 toks/s, output: 3699.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:26<00:08,  3.65it/s, est. speed input: 293.35 toks/s, output: 3808.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:28<00:19,  1.56it/s, est. speed input: 278.31 toks/s, output: 3714.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:28<00:14,  1.96it/s, est. speed input: 278.83 toks/s, output: 3823.02 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5307.9 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:21,  1.31it/s, est. speed input: 267.85 toks/s, output: 3774.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:30<00:19,  1.41it/s, est. speed input: 264.83 toks/s, output: 3834.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:31<00:17,  1.48it/s, est. speed input: 261.73 toks/s, output: 3892.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:32<00:17,  1.45it/s, est. speed input: 257.81 toks/s, output: 3937.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:33<00:14,  1.64it/s, est. speed input: 253.31 toks/s, output: 4074.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:11,  1.84it/s, est. speed input: 252.51 toks/s, output: 4166.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:33<00:10,  1.97it/s, est. speed input: 251.23 toks/s, output: 4248.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:34<00:08,  2.24it/s, est. speed input: 250.90 toks/s, output: 4346.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:34<00:07,  2.59it/s, est. speed input: 250.94 toks/s, output: 4450.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:34<00:06,  2.71it/s, est. speed input: 250.34 toks/s, output: 4542.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:35<00:07,  2.39it/s, est. speed input: 248.21 toks/s, output: 4606.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:35<00:08,  1.91it/s, est. speed input: 244.54 toks/s, output: 4642.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:36<00:08,  1.72it/s, est. speed input: 241.38 toks/s, output: 4687.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:36<00:06,  2.17it/s, est. speed input: 241.89 toks/s, output: 4801.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:37<00:05,  2.21it/s, est. speed input: 240.71 toks/s, output: 4882.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:38<00:06,  1.79it/s, est. speed input: 237.18 toks/s, output: 4916.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:38<00:04,  2.30it/s, est. speed input: 237.91 toks/s, output: 5037.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3147.8 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:40<00:08,  1.13it/s, est. speed input: 227.87 toks/s, output: 4935.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:41<00:09,  1.00s/it, est. speed input: 222.31 toks/s, output: 4926.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:41<00:06,  1.22it/s, est. speed input: 221.67 toks/s, output: 5025.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:42<00:06,  1.14it/s, est. speed input: 217.79 toks/s, output: 5052.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:45<00:07,  1.32s/it, est. speed input: 207.75 toks/s, output: 4938.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:06,  1.22s/it, est. speed input: 204.69 toks/s, output: 4985.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:46<00:04,  1.01s/it, est. speed input: 203.60 toks/s, output: 5081.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:47<00:02,  1.08it/s, est. speed input: 201.72 toks/s, output: 5157.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:47<00:01,  1.29it/s, est. speed input: 201.26 toks/s, output: 5267.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1333.9 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:50<00:01,  1.42s/it, est. speed input: 190.82 toks/s, output: 5125.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.42s/it, est. speed input: 191.98 toks/s, output: 5286.84 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.15it/s, est. speed input: 191.98 toks/s, output: 5286.84 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:53:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:53:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:53:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:53:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:53:39 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:53:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:53:39 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:53:40 [abstract.py:306] It took 0.531173 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:53:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0151
+  • Training Accuracy: 0.4813
+  • Mean Generation Length: 1679.8750
+  • Training FLOPS: 40.16 TFLOPS (10.04 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 234.29s
+  • policy_training: 149.57s (63.8%)
+  • generation: 53.88s (23.0%)
+  • teacher_logprob_inference: 25.78s (11.0%)
+  • policy_training/sharding_data: 4.32s (1.8%)
+  • prepare_for_generation: 3.60s (1.5%)
+  • policy_training/submit_training_futures: 1.82s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.7%)
+  • training_prep: 0.85s (0.4%)
+  • teacher_logprob_inference_prep: 0.40s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 135: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 09:56:37 [abstract.py:324] It took 0.048405 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08519453e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08519453e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08519453e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08519453e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08519668e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67724304e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78346558e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08519668e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67724304e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78346558e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08519668e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67724304e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78346558e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08519668e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67724304e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78346558e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69125748e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76716614e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69125748e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76716614e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69125748e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76716614e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69125748e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.76716614e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.09454174e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.27901888e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #135] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43529037e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64236221e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31145678e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #135] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.34895630e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39666138e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.19980288e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #135] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19782219e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16155396e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.50931997e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #135] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33559219e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 135: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 09:56:41 [abstract.py:324] It took 0.030770 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13078.08it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 09:56:41 [loggers.py:236] Engine 000: Avg prompt throughput: 52.5 tokens/s, Avg generation throughput: 3.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.66it/s, est. speed input: 212.50 toks/s, output: 40.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:06, 25.06it/s, est. speed input: 1243.77 toks/s, output: 331.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:03, 42.46it/s, est. speed input: 1986.19 toks/s, output: 674.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 36.71it/s, est. speed input: 1949.29 toks/s, output: 837.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:02, 47.29it/s, est. speed input: 2332.05 toks/s, output: 1345.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:00<00:02, 58.95it/s, est. speed input: 2719.95 toks/s, output: 1964.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:01, 61.77it/s, est. speed input: 2881.27 toks/s, output: 2393.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:01, 65.83it/s, est. speed input: 3038.06 toks/s, output: 2858.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 57.10it/s, est. speed input: 2989.43 toks/s, output: 3175.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 60.54it/s, est. speed input: 3084.84 toks/s, output: 3651.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:01<00:01, 45.48it/s, est. speed input: 2872.46 toks/s, output: 3717.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:02<00:02, 27.63it/s, est. speed input: 2421.70 toks/s, output: 3499.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:05<00:11,  5.84it/s, est. speed input: 1053.21 toks/s, output: 1858.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:08<00:20,  3.20it/s, est. speed input: 686.91 toks/s, output: 1512.44 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 09:56:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9694.2 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:11<00:27,  2.32it/s, est. speed input: 530.50 toks/s, output: 1496.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:12<00:30,  1.99it/s, est. speed input: 466.34 toks/s, output: 1531.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:13<00:25,  2.34it/s, est. speed input: 469.45 toks/s, output: 1763.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:14<00:29,  1.93it/s, est. speed input: 422.33 toks/s, output: 1797.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:15<00:30,  1.85it/s, est. speed input: 407.16 toks/s, output: 1843.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:16<00:31,  1.76it/s, est. speed input: 392.76 toks/s, output: 1888.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:16<00:26,  2.02it/s, est. speed input: 392.74 toks/s, output: 1998.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:17<00:30,  1.75it/s, est. speed input: 376.42 toks/s, output: 2025.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:17<00:27,  1.91it/s, est. speed input: 372.13 toks/s, output: 2110.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7370.5 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:20<00:50,  1.00it/s, est. speed input: 329.73 toks/s, output: 1980.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:20<00:42,  1.19it/s, est. speed input: 326.27 toks/s, output: 2069.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:21<00:45,  1.08it/s, est. speed input: 311.64 toks/s, output: 2086.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:22<00:45,  1.05it/s, est. speed input: 299.94 toks/s, output: 2118.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:23<00:39,  1.18it/s, est. speed input: 295.24 toks/s, output: 2193.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:23<00:36,  1.25it/s, est. speed input: 289.46 toks/s, output: 2258.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:24<00:31,  1.43it/s, est. speed input: 286.63 toks/s, output: 2344.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:35,  1.23it/s, est. speed input: 276.61 toks/s, output: 2370.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:26<00:32,  1.34it/s, est. speed input: 272.77 toks/s, output: 2445.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:35,  1.20it/s, est. speed input: 264.48 toks/s, output: 2479.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:29<00:41,  1.05s/it, est. speed input: 245.41 toks/s, output: 2507.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5820.4 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:30<00:39,  1.01s/it, est. speed input: 240.30 toks/s, output: 2562.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:30,  1.24it/s, est. speed input: 240.41 toks/s, output: 2671.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:33<00:47,  1.28s/it, est. speed input: 223.87 toks/s, output: 2596.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:34<00:40,  1.11s/it, est. speed input: 221.09 toks/s, output: 2670.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:34<00:35,  1.00s/it, est. speed input: 218.20 toks/s, output: 2742.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:34<00:25,  1.33it/s, est. speed input: 219.11 toks/s, output: 2859.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:35<00:22,  1.49it/s, est. speed input: 217.93 toks/s, output: 2949.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:13,  2.31it/s, est. speed input: 219.48 toks/s, output: 3180.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:36<00:14,  2.05it/s, est. speed input: 217.10 toks/s, output: 3250.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:37<00:13,  2.09it/s, est. speed input: 214.89 toks/s, output: 3423.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:09,  2.72it/s, est. speed input: 216.10 toks/s, output: 3648.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:14,  1.73it/s, est. speed input: 210.01 toks/s, output: 3649.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4416.1 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:40<00:18,  1.29it/s, est. speed input: 204.07 toks/s, output: 3650.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:40<00:15,  1.49it/s, est. speed input: 203.69 toks/s, output: 3748.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:41<00:13,  1.67it/s, est. speed input: 203.22 toks/s, output: 3843.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:41<00:10,  1.97it/s, est. speed input: 203.45 toks/s, output: 3950.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:42<00:09,  1.97it/s, est. speed input: 201.46 toks/s, output: 4116.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:42<00:07,  2.44it/s, est. speed input: 202.44 toks/s, output: 4238.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:43<00:08,  2.06it/s, est. speed input: 200.52 toks/s, output: 4301.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:08,  1.87it/s, est. speed input: 198.85 toks/s, output: 4369.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:44<00:06,  2.33it/s, est. speed input: 199.55 toks/s, output: 4487.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:44<00:05,  2.70it/s, est. speed input: 199.92 toks/s, output: 4599.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:45<00:06,  2.01it/s, est. speed input: 197.66 toks/s, output: 4651.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:45<00:06,  1.91it/s, est. speed input: 196.48 toks/s, output: 4726.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:47<00:09,  1.11it/s, est. speed input: 190.29 toks/s, output: 4683.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:49<00:12,  1.23s/it, est. speed input: 183.79 toks/s, output: 4633.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2682.8 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:50<00:10,  1.20s/it, est. speed input: 180.87 toks/s, output: 4670.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:51<00:08,  1.09s/it, est. speed input: 179.19 toks/s, output: 4738.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:53<00:09,  1.38s/it, est. speed input: 173.48 toks/s, output: 4701.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:54<00:08,  1.34s/it, est. speed input: 170.55 toks/s, output: 4739.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:55<00:05,  1.15s/it, est. speed input: 169.53 toks/s, output: 4829.37 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  1.15s/it, est. speed input: 175.05 toks/s, output: 5567.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  2.88it/s, est. speed input: 175.05 toks/s, output: 5567.62 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 09:57:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:57:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:57:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:57:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 09:57:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 09:57:39 [abstract.py:306] It took 0.534166 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 09:57:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 09:57:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 09:57:39 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.3312
+  • Mean Generation Length: 1930.6250
+  • Training FLOPS: 42.05 TFLOPS (10.51 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 258.83s
+  • policy_training: 166.62s (64.4%)
+  • generation: 58.86s (22.7%)
+  • teacher_logprob_inference: 27.61s (10.7%)
+  • policy_training/sharding_data: 4.35s (1.7%)
+  • prepare_for_generation: 4.28s (1.7%)
+  • prepare_for_generation/transfer_and_update_weights: 2.03s (0.8%)
+  • policy_training/submit_training_futures: 1.93s (0.7%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 28/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 136: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:00:56 [abstract.py:324] It took 0.050047 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08539727e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08539727e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08539727e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08539727e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547734e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68482323e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76632538e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547734e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68482323e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76632538e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547734e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68482323e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76632538e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547734e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68482323e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76632538e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69178963e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77420197e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69178963e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77420197e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69178963e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77420197e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69178963e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77420197e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07861099e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.07620287e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #136] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43362640e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64781799e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30310163e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #136] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.34834099e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15332870e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.52379074e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #136] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33262878e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38416367e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.95111227e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #136] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19458466e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 136: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:00:59 [abstract.py:324] It took 0.032665 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:00:59 [loggers.py:236] Engine 000: Avg prompt throughput: 46.9 tokens/s, Avg generation throughput: 43.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13140.05it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.01it/s, est. speed input: 360.78 toks/s, output: 78.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 41.62it/s, est. speed input: 2187.86 toks/s, output: 692.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 34.87it/s, est. speed input: 2022.36 toks/s, output: 836.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:02, 48.25it/s, est. speed input: 2544.61 toks/s, output: 1351.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:02, 44.43it/s, est. speed input: 2504.84 toks/s, output: 1558.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 43.77it/s, est. speed input: 2522.96 toks/s, output: 1826.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:00<00:02, 52.78it/s, est. speed input: 2770.36 toks/s, output: 2362.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 52.46it/s, est. speed input: 2812.79 toks/s, output: 2661.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 47.64it/s, est. speed input: 2762.17 toks/s, output: 2890.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 48.56it/s, est. speed input: 2800.86 toks/s, output: 3314.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:02, 39.76it/s, est. speed input: 2641.80 toks/s, output: 3443.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:02, 39.57it/s, est. speed input: 2620.11 toks/s, output: 3694.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:02<00:06, 13.19it/s, est. speed input: 1716.14 toks/s, output: 2677.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:05<00:17,  4.40it/s, est. speed input: 884.63 toks/s, output: 1690.58 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:06<00:16,  4.44it/s, est. speed input: 823.59 toks/s, output: 1872.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:08<00:25,  2.79it/s, est. speed input: 631.64 toks/s, output: 1633.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:09<00:26,  2.64it/s, est. speed input: 582.84 toks/s, output: 1710.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10086.3 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:10<00:30,  2.28it/s, est. speed input: 538.16 toks/s, output: 1682.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:11<00:31,  2.13it/s, est. speed input: 511.65 toks/s, output: 1702.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:11<00:30,  2.19it/s, est. speed input: 499.07 toks/s, output: 1763.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:11<00:28,  2.34it/s, est. speed input: 491.24 toks/s, output: 1838.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:13<00:48,  1.34it/s, est. speed input: 427.51 toks/s, output: 1703.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:14<00:43,  1.49it/s, est. speed input: 418.77 toks/s, output: 1771.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:15<00:49,  1.27it/s, est. speed input: 391.79 toks/s, output: 1761.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<01:38,  1.59s/it, est. speed input: 316.26 toks/s, output: 1529.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<01:26,  1.42s/it, est. speed input: 304.19 toks/s, output: 1577.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7842.0 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<01:18,  1.31s/it, est. speed input: 291.95 toks/s, output: 1619.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:21<01:09,  1.17s/it, est. speed input: 283.74 toks/s, output: 1679.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:22<00:58,  1.00s/it, est. speed input: 278.97 toks/s, output: 1755.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<00:51,  1.11it/s, est. speed input: 273.72 toks/s, output: 1826.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:23<00:38,  1.45it/s, est. speed input: 274.23 toks/s, output: 1932.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:23<00:35,  1.54it/s, est. speed input: 270.24 toks/s, output: 2007.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:40,  1.34it/s, est. speed input: 262.13 toks/s, output: 2049.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:41,  1.27it/s, est. speed input: 255.56 toks/s, output: 2099.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:31,  1.64it/s, est. speed input: 255.89 toks/s, output: 2204.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:27,  1.85it/s, est. speed input: 254.48 toks/s, output: 2293.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:23,  2.15it/s, est. speed input: 253.94 toks/s, output: 2388.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:27<00:26,  1.83it/s, est. speed input: 249.27 toks/s, output: 2445.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:27<00:17,  2.68it/s, est. speed input: 250.46 toks/s, output: 2655.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:28<00:25,  1.80it/s, est. speed input: 242.92 toks/s, output: 2673.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:29<00:16,  2.53it/s, est. speed input: 243.04 toks/s, output: 2962.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6383.4 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:31<00:29,  1.45it/s, est. speed input: 230.25 toks/s, output: 2904.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:31<00:23,  1.74it/s, est. speed input: 231.00 toks/s, output: 3010.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:20,  1.91it/s, est. speed input: 230.28 toks/s, output: 3098.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:32<00:22,  1.73it/s, est. speed input: 226.85 toks/s, output: 3149.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:33<00:25,  1.46it/s, est. speed input: 222.14 toks/s, output: 3179.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:34<00:24,  1.54it/s, est. speed input: 220.21 toks/s, output: 3249.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:34<00:20,  1.78it/s, est. speed input: 219.83 toks/s, output: 3339.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:35<00:25,  1.37it/s, est. speed input: 214.46 toks/s, output: 3353.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:37<00:33,  1.02it/s, est. speed input: 206.94 toks/s, output: 3333.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:37<00:27,  1.19it/s, est. speed input: 205.79 toks/s, output: 3411.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:38<00:24,  1.31it/s, est. speed input: 204.35 toks/s, output: 3483.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:39<00:18,  1.66it/s, est. speed input: 203.04 toks/s, output: 3653.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:39<00:14,  2.01it/s, est. speed input: 203.70 toks/s, output: 3761.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:39<00:14,  1.95it/s, est. speed input: 202.32 toks/s, output: 3831.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4747.9 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:40<00:17,  1.52it/s, est. speed input: 198.64 toks/s, output: 3857.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:41<00:14,  1.74it/s, est. speed input: 198.35 toks/s, output: 3946.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:41<00:10,  2.19it/s, est. speed input: 198.34 toks/s, output: 4135.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:42<00:09,  2.45it/s, est. speed input: 198.60 toks/s, output: 4235.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:42<00:07,  2.81it/s, est. speed input: 199.10 toks/s, output: 4339.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:44<00:16,  1.28it/s, est. speed input: 191.70 toks/s, output: 4274.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:44<00:09,  1.97it/s, est. speed input: 193.14 toks/s, output: 4498.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:44<00:06,  2.64it/s, est. speed input: 194.38 toks/s, output: 4716.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:45<00:03,  3.80it/s, est. speed input: 196.87 toks/s, output: 5057.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:46<00:05,  2.17it/s, est. speed input: 192.36 toks/s, output: 5036.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:46<00:05,  2.38it/s, est. speed input: 192.57 toks/s, output: 5138.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:47<00:04,  2.63it/s, est. speed input: 192.88 toks/s, output: 5241.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:48<00:05,  1.85it/s, est. speed input: 189.96 toks/s, output: 5258.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2955.2 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:51<00:11,  1.27s/it, est. speed input: 178.63 toks/s, output: 5048.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:54<00:13,  1.64s/it, est. speed input: 171.07 toks/s, output: 4942.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:54<00:08,  1.21s/it, est. speed input: 171.85 toks/s, output: 5072.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:08,  1.34s/it, est. speed input: 167.84 toks/s, output: 5064.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:57<00:06,  1.30s/it, est. speed input: 165.37 toks/s, output: 5100.08 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  1.30s/it, est. speed input: 170.63 toks/s, output: 5816.15 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  2.80it/s, est. speed input: 170.63 toks/s, output: 5816.15 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:59 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:01:59 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:01:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:01:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:01:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:01:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:02:00 [abstract.py:306] It took 0.536012 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:02:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:02:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:02:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:02:00 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0118
+  • Training Accuracy: 0.3750
+  • Mean Generation Length: 2079.2563
+  • Training FLOPS: 42.67 TFLOPS (10.67 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 273.30s
+  • policy_training: 175.92s (64.4%)
+  • generation: 60.82s (22.3%)
+  • teacher_logprob_inference: 30.93s (11.3%)
+  • policy_training/sharding_data: 4.30s (1.6%)
+  • prepare_for_generation: 3.73s (1.4%)
+  • policy_training/submit_training_futures: 1.91s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.68s (0.6%)
+  • training_prep: 1.33s (0.5%)
+  • teacher_logprob_inference_prep: 0.39s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 137: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:05:30 [abstract.py:324] It took 0.052463 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547305e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68516769e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78134842e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547305e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68516769e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78134842e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547305e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68516769e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78134842e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547305e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68516769e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78134842e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547461e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69605904e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76750183e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08077698e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.01878643e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #137] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43172409e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547461e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69605904e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76750183e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65840149e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30280714e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #137] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.35159378e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547461e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69605904e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76750183e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38604126e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.92673159e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #137] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16102791e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08547461e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69605904e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.76750183e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15513992e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53326426e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #137] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33014191e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 137: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:05:33 [abstract.py:324] It took 0.029805 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13355.26it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:05:33 [loggers.py:236] Engine 000: Avg prompt throughput: 43.2 tokens/s, Avg generation throughput: 55.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:46,  3.42it/s, est. speed input: 208.37 toks/s, output: 47.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 20.22it/s, est. speed input: 1035.40 toks/s, output: 317.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 21.64it/s, est. speed input: 1142.74 toks/s, output: 427.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 22.03it/s, est. speed input: 1192.34 toks/s, output: 554.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:06, 21.23it/s, est. speed input: 1194.31 toks/s, output: 673.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:06, 20.75it/s, est. speed input: 1196.42 toks/s, output: 799.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:03, 34.25it/s, est. speed input: 1526.50 toks/s, output: 1355.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:03, 36.69it/s, est. speed input: 1646.08 toks/s, output: 1690.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 34.65it/s, est. speed input: 1673.12 toks/s, output: 1897.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 42.84it/s, est. speed input: 1845.53 toks/s, output: 2373.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 37.48it/s, est. speed input: 1834.40 toks/s, output: 2568.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 35.86it/s, est. speed input: 1844.01 toks/s, output: 2802.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:02<00:05, 19.13it/s, est. speed input: 1546.43 toks/s, output: 2533.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:08<00:42,  2.28it/s, est. speed input: 471.34 toks/s, output: 967.06 toks/s]  
+(VllmGenerationWorker pid=10140) INFO 05-03 10:05:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11403.3 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:11<00:53,  1.77it/s, est. speed input: 360.10 toks/s, output: 979.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:11<00:49,  1.86it/s, est. speed input: 345.40 toks/s, output: 1120.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:13<00:52,  1.72it/s, est. speed input: 315.33 toks/s, output: 1204.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:13<00:49,  1.83it/s, est. speed input: 312.48 toks/s, output: 1284.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:14<00:47,  1.85it/s, est. speed input: 305.75 toks/s, output: 1344.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:14<00:48,  1.80it/s, est. speed input: 296.57 toks/s, output: 1392.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:14<00:41,  2.09it/s, est. speed input: 297.11 toks/s, output: 1481.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:16<00:53,  1.61it/s, est. speed input: 280.28 toks/s, output: 1484.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:16<00:41,  2.02it/s, est. speed input: 277.47 toks/s, output: 1639.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:16<00:36,  2.30it/s, est. speed input: 277.31 toks/s, output: 1723.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:18<00:32,  2.45it/s, est. speed input: 269.87 toks/s, output: 1918.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:18<00:31,  2.52it/s, est. speed input: 268.04 toks/s, output: 1985.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:05:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8941.6 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:20<00:57,  1.35it/s, est. speed input: 244.49 toks/s, output: 1894.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:23<01:36,  1.25s/it, est. speed input: 216.26 toks/s, output: 1760.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:24<01:27,  1.15s/it, est. speed input: 211.01 toks/s, output: 1802.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:25<01:03,  1.16it/s, est. speed input: 207.97 toks/s, output: 1940.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:25<00:51,  1.43it/s, est. speed input: 209.16 toks/s, output: 2033.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:27<01:21,  1.13s/it, est. speed input: 192.86 toks/s, output: 1956.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:28<01:11,  1.00s/it, est. speed input: 190.72 toks/s, output: 2016.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:29<01:05,  1.07it/s, est. speed input: 187.88 toks/s, output: 2066.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<00:55,  1.24it/s, est. speed input: 186.93 toks/s, output: 2136.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:29<00:44,  1.54it/s, est. speed input: 187.43 toks/s, output: 2221.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:29<00:33,  2.00it/s, est. speed input: 188.58 toks/s, output: 2313.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7352.1 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.4%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:30<00:25,  2.55it/s, est. speed input: 189.76 toks/s, output: 2406.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:30<00:29,  2.23it/s, est. speed input: 188.13 toks/s, output: 2463.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:31<00:34,  1.88it/s, est. speed input: 185.67 toks/s, output: 2509.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:31<00:27,  2.30it/s, est. speed input: 186.35 toks/s, output: 2596.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:33<00:50,  1.22it/s, est. speed input: 178.47 toks/s, output: 2565.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:33<00:39,  1.53it/s, est. speed input: 178.82 toks/s, output: 2648.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:33<00:22,  2.61it/s, est. speed input: 181.67 toks/s, output: 2845.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:35<00:37,  1.54it/s, est. speed input: 175.83 toks/s, output: 2830.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:35<00:37,  1.50it/s, est. speed input: 174.03 toks/s, output: 2877.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:36<00:33,  1.65it/s, est. speed input: 173.53 toks/s, output: 2946.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:25,  2.10it/s, est. speed input: 173.87 toks/s, output: 3102.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:37<00:23,  2.22it/s, est. speed input: 173.80 toks/s, output: 3175.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:37<00:20,  2.48it/s, est. speed input: 174.18 toks/s, output: 3256.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:37<00:18,  2.74it/s, est. speed input: 174.59 toks/s, output: 3338.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:38<00:15,  3.20it/s, est. speed input: 175.51 toks/s, output: 3426.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:22,  2.20it/s, est. speed input: 173.33 toks/s, output: 3458.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6300.1 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.5%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:40<00:44,  1.09it/s, est. speed input: 166.00 toks/s, output: 3388.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:41<00:33,  1.39it/s, est. speed input: 166.51 toks/s, output: 3474.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:41<00:25,  1.83it/s, est. speed input: 167.42 toks/s, output: 3568.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:41<00:14,  3.05it/s, est. speed input: 169.80 toks/s, output: 3766.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:42<00:17,  2.46it/s, est. speed input: 168.66 toks/s, output: 3813.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:42<00:17,  2.36it/s, est. speed input: 168.19 toks/s, output: 3876.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:42<00:12,  3.14it/s, est. speed input: 169.60 toks/s, output: 4054.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:44<00:20,  1.90it/s, est. speed input: 166.28 toks/s, output: 4049.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:45<00:25,  1.49it/s, est. speed input: 163.51 toks/s, output: 4056.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:45<00:25,  1.45it/s, est. speed input: 162.23 toks/s, output: 4098.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:46<00:23,  1.54it/s, est. speed input: 161.64 toks/s, output: 4157.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:46<00:18,  1.94it/s, est. speed input: 162.30 toks/s, output: 4249.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:46<00:14,  2.28it/s, est. speed input: 162.72 toks/s, output: 4334.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:47<00:11,  2.76it/s, est. speed input: 163.39 toks/s, output: 4425.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:47<00:11,  2.85it/s, est. speed input: 163.55 toks/s, output: 4502.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:48<00:13,  2.35it/s, est. speed input: 162.72 toks/s, output: 4553.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:48<00:10,  2.83it/s, est. speed input: 163.34 toks/s, output: 4642.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:48<00:12,  2.35it/s, est. speed input: 162.57 toks/s, output: 4693.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:49<00:14,  1.96it/s, est. speed input: 161.47 toks/s, output: 4734.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:49<00:08,  3.04it/s, est. speed input: 163.11 toks/s, output: 4928.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4645.3 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:09,  2.68it/s, est. speed input: 162.68 toks/s, output: 4986.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:52<00:20,  1.18it/s, est. speed input: 157.06 toks/s, output: 4889.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:53<00:18,  1.27it/s, est. speed input: 156.35 toks/s, output: 4941.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:13,  1.58it/s, est. speed input: 156.81 toks/s, output: 5030.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:10,  2.00it/s, est. speed input: 157.45 toks/s, output: 5125.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:54<00:09,  2.03it/s, est. speed input: 156.86 toks/s, output: 5254.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:54<00:07,  2.39it/s, est. speed input: 157.44 toks/s, output: 5349.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:54<00:05,  2.91it/s, est. speed input: 158.15 toks/s, output: 5448.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:03,  4.32it/s, est. speed input: 159.84 toks/s, output: 5657.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:55<00:03,  3.83it/s, est. speed input: 159.89 toks/s, output: 5733.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:55<00:03,  3.65it/s, est. speed input: 160.05 toks/s, output: 5814.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:56<00:05,  2.07it/s, est. speed input: 158.13 toks/s, output: 5818.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:56<00:04,  2.57it/s, est. speed input: 158.83 toks/s, output: 5919.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:56<00:03,  3.21it/s, est. speed input: 159.55 toks/s, output: 6022.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:57<00:03,  2.81it/s, est. speed input: 159.31 toks/s, output: 6089.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:58<00:04,  1.92it/s, est. speed input: 157.82 toks/s, output: 6109.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2723.6 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 76.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:00<00:06,  1.03it/s, est. speed input: 153.44 toks/s, output: 6021.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:00<00:04,  1.35it/s, est. speed input: 153.93 toks/s, output: 6123.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:03,  1.64it/s, est. speed input: 154.18 toks/s, output: 6214.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:02,  1.51it/s, est. speed input: 153.20 toks/s, output: 6258.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:01<00:01,  1.97it/s, est. speed input: 153.79 toks/s, output: 6365.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:02<00:00,  2.12it/s, est. speed input: 153.82 toks/s, output: 6450.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:04<00:00,  1.12it/s, est. speed input: 150.22 toks/s, output: 6388.50 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.12it/s, est. speed input: 151.14 toks/s, output: 6516.28 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.50it/s, est. speed input: 151.14 toks/s, output: 6516.28 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:41 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:41 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:06:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:06:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:06:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:06:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:06:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:06:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:06:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:06:41 [abstract.py:306] It took 0.535039 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:06:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0102
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 2608.4563
+  • Training FLOPS: 44.28 TFLOPS (11.07 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 322.59s
+  • policy_training: 210.83s (65.4%)
+  • generation: 68.38s (21.2%)
+  • teacher_logprob_inference: 37.91s (11.8%)
+  • policy_training/sharding_data: 4.33s (1.3%)
+  • prepare_for_generation: 3.84s (1.2%)
+  • policy_training/submit_training_futures: 1.79s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.73s (0.5%)
+  • training_prep: 0.83s (0.3%)
+  • teacher_logprob_inference_prep: 0.63s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 138: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:10:52 [abstract.py:324] It took 0.049674 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547383e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547383e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547383e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08547383e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08558965e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69651260e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79254608e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08558965e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69651260e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79254608e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08558965e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69651260e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79254608e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08558965e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69651260e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79254608e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68382721e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79505844e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68382721e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79505844e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68382721e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79505844e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68382721e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79505844e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.06792183e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.95929575e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #138] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43096985e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65257263e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30392036e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #138] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.38737373e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39259949e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.98371029e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #138] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13011818e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15642242e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54254646e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #138] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33125870e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 138: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:10:55 [abstract.py:324] It took 0.030844 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13232.81it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:10:55 [loggers.py:236] Engine 000: Avg prompt throughput: 37.2 tokens/s, Avg generation throughput: 13.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.61it/s, est. speed input: 216.85 toks/s, output: 39.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:03, 37.95it/s, est. speed input: 1883.40 toks/s, output: 462.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:02, 54.18it/s, est. speed input: 2627.14 toks/s, output: 872.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:02, 57.66it/s, est. speed input: 2868.55 toks/s, output: 1190.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:01, 63.73it/s, est. speed input: 3142.25 toks/s, output: 1589.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:00<00:01, 61.59it/s, est. speed input: 3205.84 toks/s, output: 1953.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:00<00:01, 72.28it/s, est. speed input: 3523.07 toks/s, output: 2604.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 71.00it/s, est. speed input: 3599.47 toks/s, output: 3012.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:01<00:00, 85.52it/s, est. speed input: 3935.62 toks/s, output: 3898.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:01<00:00, 82.53it/s, est. speed input: 3995.33 toks/s, output: 4351.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:01<00:00, 67.89it/s, est. speed input: 3856.39 toks/s, output: 4641.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:03<00:04, 13.08it/s, est. speed input: 1823.51 toks/s, output: 2633.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8858.1 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:12<00:21,  2.41it/s, est. speed input: 512.92 toks/s, output: 1234.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:16<00:23,  2.02it/s, est. speed input: 419.64 toks/s, output: 1503.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:17<00:21,  2.06it/s, est. speed input: 398.78 toks/s, output: 1791.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:20<00:24,  1.73it/s, est. speed input: 358.31 toks/s, output: 1850.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6596.3 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:20<00:19,  2.01it/s, est. speed input: 361.50 toks/s, output: 2112.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:21<00:18,  2.09it/s, est. speed input: 354.05 toks/s, output: 2307.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:21<00:17,  2.17it/s, est. speed input: 351.49 toks/s, output: 2410.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:22<00:19,  1.88it/s, est. speed input: 339.01 toks/s, output: 2445.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:23<00:23,  1.46it/s, est. speed input: 321.27 toks/s, output: 2438.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:24<00:23,  1.45it/s, est. speed input: 314.50 toks/s, output: 2505.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:26<00:30,  1.09it/s, est. speed input: 295.86 toks/s, output: 2478.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:26<00:26,  1.22it/s, est. speed input: 292.58 toks/s, output: 2570.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:27<00:22,  1.36it/s, est. speed input: 289.55 toks/s, output: 2663.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:27<00:17,  1.68it/s, est. speed input: 289.55 toks/s, output: 2783.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:30<00:33,  1.16s/it, est. speed input: 266.27 toks/s, output: 2679.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4870.8 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:26,  1.07it/s, est. speed input: 264.95 toks/s, output: 2786.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:30<00:20,  1.31it/s, est. speed input: 264.10 toks/s, output: 2897.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:31<00:19,  1.32it/s, est. speed input: 259.69 toks/s, output: 2969.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:31<00:14,  1.67it/s, est. speed input: 259.94 toks/s, output: 3090.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:32<00:08,  2.52it/s, est. speed input: 259.82 toks/s, output: 3441.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:32<00:07,  2.88it/s, est. speed input: 260.41 toks/s, output: 3565.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:33<00:07,  2.59it/s, est. speed input: 258.23 toks/s, output: 3652.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:33<00:06,  2.85it/s, est. speed input: 258.21 toks/s, output: 3767.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:14,  1.21it/s, est. speed input: 240.31 toks/s, output: 3732.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:37<00:13,  1.20it/s, est. speed input: 236.43 toks/s, output: 3791.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:39<00:16,  1.13s/it, est. speed input: 225.72 toks/s, output: 3741.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:40<00:14,  1.04s/it, est. speed input: 222.89 toks/s, output: 3815.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3205.9 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:14,  1.13s/it, est. speed input: 216.88 toks/s, output: 3836.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:14,  1.27s/it, est. speed input: 205.28 toks/s, output: 3873.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:45<00:11,  1.18s/it, est. speed input: 202.70 toks/s, output: 3951.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:09,  1.08s/it, est. speed input: 200.55 toks/s, output: 4036.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:07,  1.04it/s, est. speed input: 199.07 toks/s, output: 4134.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:48<00:07,  1.12s/it, est. speed input: 194.04 toks/s, output: 4158.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2030.0 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:51<00:10,  1.71s/it, est. speed input: 183.11 toks/s, output: 4059.35 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.71s/it, est. speed input: 190.05 toks/s, output: 5016.41 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  3.12it/s, est. speed input: 190.05 toks/s, output: 5016.41 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:49 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:49 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:11:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:11:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:11:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:11:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:11:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:11:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:11:49 [abstract.py:306] It took 0.576873 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:11:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:11:49 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.3688
+  • Mean Generation Length: 1610.1438
+  • Training FLOPS: 40.83 TFLOPS (10.21 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 226.30s
+  • policy_training: 143.97s (63.6%)
+  • generation: 54.28s (24.0%)
+  • teacher_logprob_inference: 23.09s (10.2%)
+  • policy_training/sharding_data: 4.31s (1.9%)
+  • prepare_for_generation: 3.63s (1.6%)
+  • policy_training/submit_training_futures: 1.79s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.7%)
+  • training_prep: 0.78s (0.3%)
+  • teacher_logprob_inference_prep: 0.38s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 139: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:14:39 [abstract.py:324] It took 0.048793 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08563633e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69281921e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80663300e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08563633e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69281921e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80663300e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08563633e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69281921e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80663300e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08563633e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69281921e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80663300e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08572773e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71278458e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80847015e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.06703663e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.87467098e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #139] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43178238e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08572773e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71278458e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80847015e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65303802e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.28697176e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #139] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.36956978e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08572773e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71278458e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80847015e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39655304e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.22665644e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #139] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.12204628e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08572773e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71278458e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80847015e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15193558e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54308691e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #139] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33316833e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 139: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:14:42 [abstract.py:324] It took 0.031854 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13427.14it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:14:42 [loggers.py:236] Engine 000: Avg prompt throughput: 55.2 tokens/s, Avg generation throughput: 11.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.08it/s, est. speed input: 358.59 toks/s, output: 79.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:04, 34.85it/s, est. speed input: 1809.02 toks/s, output: 480.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:03, 40.59it/s, est. speed input: 2139.20 toks/s, output: 742.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:03, 43.85it/s, est. speed input: 2326.96 toks/s, output: 982.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:02, 51.07it/s, est. speed input: 2628.65 toks/s, output: 1421.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:02, 53.08it/s, est. speed input: 2762.90 toks/s, output: 1737.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:00<00:01, 70.01it/s, est. speed input: 3253.96 toks/s, output: 2509.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:00<00:01, 86.67it/s, est. speed input: 3757.10 toks/s, output: 3403.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 83.41it/s, est. speed input: 3866.90 toks/s, output: 3854.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:01<00:01, 84.66it/s, est. speed input: 4024.59 toks/s, output: 4461.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:01<00:00, 97.00it/s, est. speed input: 4351.21 toks/s, output: 5445.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:01<00:00, 88.64it/s, est. speed input: 4355.05 toks/s, output: 5945.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:01<00:00, 52.34it/s, est. speed input: 3754.12 toks/s, output: 5645.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:14:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8189.9 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:12<00:14,  2.97it/s, est. speed input: 587.11 toks/s, output: 1406.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:12<00:13,  3.03it/s, est. speed input: 583.17 toks/s, output: 1535.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:15<00:14,  2.50it/s, est. speed input: 476.18 toks/s, output: 2019.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:16<00:11,  2.72it/s, est. speed input: 463.38 toks/s, output: 2497.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:15:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5662.6 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:20<00:14,  2.04it/s, est. speed input: 399.26 toks/s, output: 2529.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:21<00:14,  1.81it/s, est. speed input: 371.05 toks/s, output: 2621.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:22<00:12,  2.00it/s, est. speed input: 368.67 toks/s, output: 2867.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:23<00:11,  1.96it/s, est. speed input: 356.57 toks/s, output: 3032.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:23<00:10,  2.14it/s, est. speed input: 356.66 toks/s, output: 3165.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:24<00:07,  2.51it/s, est. speed input: 355.51 toks/s, output: 3415.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:24<00:05,  3.02it/s, est. speed input: 355.91 toks/s, output: 3679.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:25<00:06,  2.59it/s, est. speed input: 348.87 toks/s, output: 3736.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:25<00:05,  2.99it/s, est. speed input: 349.68 toks/s, output: 3874.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:25<00:06,  2.30it/s, est. speed input: 341.08 toks/s, output: 3909.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:28<00:11,  1.17it/s, est. speed input: 315.98 toks/s, output: 3756.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:28<00:10,  1.19it/s, est. speed input: 309.42 toks/s, output: 3813.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:29<00:09,  1.29it/s, est. speed input: 305.24 toks/s, output: 3898.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:15:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3442.5 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:31<00:11,  1.07s/it, est. speed input: 289.15 toks/s, output: 3831.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:31<00:09,  1.08it/s, est. speed input: 285.84 toks/s, output: 3926.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:32<00:06,  1.33it/s, est. speed input: 284.89 toks/s, output: 4052.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:32<00:05,  1.58it/s, est. speed input: 283.64 toks/s, output: 4172.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:33<00:04,  1.67it/s, est. speed input: 281.07 toks/s, output: 4273.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:33<00:03,  1.84it/s, est. speed input: 279.47 toks/s, output: 4386.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:33<00:02,  2.03it/s, est. speed input: 278.24 toks/s, output: 4505.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:34<00:01,  2.13it/s, est. speed input: 276.63 toks/s, output: 4619.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:35<00:02,  1.30it/s, est. speed input: 267.04 toks/s, output: 4600.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:36<00:01,  1.17it/s, est. speed input: 261.05 toks/s, output: 4642.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:15:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1249.1 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:40<00:01,  1.55s/it, est. speed input: 241.85 toks/s, output: 4455.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:43<00:00,  2.11s/it, est. speed input: 224.19 toks/s, output: 4292.90 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:43<00:00,  2.11s/it, est. speed input: 224.19 toks/s, output: 4292.90 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:43<00:00,  3.68it/s, est. speed input: 224.19 toks/s, output: 4292.90 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:15:27 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:15:27 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:15:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:15:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:15:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:15:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:15:27 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:15:27 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:15:27 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:15:27 [abstract.py:306] It took 0.556388 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:15:27 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0157
+  • Training Accuracy: 0.3812
+  • Mean Generation Length: 1166.1312
+  • Training FLOPS: 31.64 TFLOPS (7.91 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 204.73s
+  • policy_training: 132.92s (64.9%)
+  • generation: 45.77s (22.4%)
+  • teacher_logprob_inference: 20.89s (10.2%)
+  • policy_training/sharding_data: 4.36s (2.1%)
+  • prepare_for_generation: 3.55s (1.7%)
+  • policy_training/submit_training_futures: 2.00s (1.0%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.8%)
+  • training_prep: 0.79s (0.4%)
+  • teacher_logprob_inference_prep: 0.64s (0.3%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 140: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:04 [abstract.py:324] It took 0.085438 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08584062e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69065628e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80699921e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08584062e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69065628e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80699921e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08584062e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69065628e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80699921e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08584062e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69065628e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80699921e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08588359e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69309921e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80204544e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.06040916e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.74838877e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #140] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43297211e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08588359e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69309921e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80204544e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63952255e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27297039e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #140] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39039192e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08588359e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69309921e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80204544e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39782486e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.88501787e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #140] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.06808510e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08588359e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69309921e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80204544e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14824333e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.53933468e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #140] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33507507e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 140: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:06 [abstract.py:324] It took 0.028907 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13653.61it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:07 [loggers.py:236] Engine 000: Avg prompt throughput: 58.5 tokens/s, Avg generation throughput: 6.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.09it/s, est. speed input: 123.26 toks/s, output: 31.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:16,  9.53it/s, est. speed input: 480.65 toks/s, output: 166.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:15, 10.14it/s, est. speed input: 523.77 toks/s, output: 236.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:12, 11.64it/s, est. speed input: 584.15 toks/s, output: 324.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:01<00:08, 17.49it/s, est. speed input: 753.85 toks/s, output: 543.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:06, 22.26it/s, est. speed input: 886.42 toks/s, output: 747.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:01<00:05, 24.06it/s, est. speed input: 955.55 toks/s, output: 895.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:04, 28.17it/s, est. speed input: 1062.20 toks/s, output: 1117.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 37.19it/s, est. speed input: 1261.78 toks/s, output: 1531.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 37.63it/s, est. speed input: 1333.96 toks/s, output: 1728.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 31.85it/s, est. speed input: 1337.07 toks/s, output: 1858.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:03, 33.09it/s, est. speed input: 1387.61 toks/s, output: 2070.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:02<00:07, 15.93it/s, est. speed input: 1165.33 toks/s, output: 1903.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:02<00:08, 12.66it/s, est. speed input: 1068.35 toks/s, output: 1895.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:04<00:22,  4.65it/s, est. speed input: 687.59 toks/s, output: 1381.81 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:04<00:20,  5.01it/s, est. speed input: 674.66 toks/s, output: 1500.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:06<00:31,  3.30it/s, est. speed input: 548.71 toks/s, output: 1375.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:08<01:00,  1.69it/s, est. speed input: 402.43 toks/s, output: 1094.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:09<00:58,  1.72it/s, est. speed input: 385.95 toks/s, output: 1134.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:09<01:00,  1.66it/s, est. speed input: 364.56 toks/s, output: 1157.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:10<00:49,  2.00it/s, est. speed input: 366.26 toks/s, output: 1246.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11758.3 tokens/s, Running: 99 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:10<00:54,  1.81it/s, est. speed input: 347.24 toks/s, output: 1266.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:12<01:20,  1.20it/s, est. speed input: 305.66 toks/s, output: 1199.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:13<01:14,  1.28it/s, est. speed input: 295.50 toks/s, output: 1243.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:15<02:06,  1.33s/it, est. speed input: 247.47 toks/s, output: 1127.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:15<01:32,  1.02it/s, est. speed input: 249.85 toks/s, output: 1222.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:17<01:33,  1.01s/it, est. speed input: 237.64 toks/s, output: 1247.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:19<02:00,  1.31s/it, est. speed input: 215.36 toks/s, output: 1215.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:19<01:30,  1.01it/s, est. speed input: 215.94 toks/s, output: 1301.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:19<01:16,  1.17it/s, est. speed input: 213.30 toks/s, output: 1367.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9388.7 tokens/s, Running: 90 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:20<01:22,  1.07it/s, est. speed input: 204.83 toks/s, output: 1396.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:21<01:05,  1.34it/s, est. speed input: 204.77 toks/s, output: 1477.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:21<01:03,  1.37it/s, est. speed input: 200.99 toks/s, output: 1531.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:22<00:47,  1.80it/s, est. speed input: 202.39 toks/s, output: 1621.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:22<00:27,  3.01it/s, est. speed input: 206.53 toks/s, output: 1811.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:23<00:36,  2.27it/s, est. speed input: 202.20 toks/s, output: 1851.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:23<00:33,  2.42it/s, est. speed input: 202.00 toks/s, output: 1924.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:23<00:31,  2.57it/s, est. speed input: 201.69 toks/s, output: 1998.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:23<00:28,  2.78it/s, est. speed input: 201.80 toks/s, output: 2074.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:24<00:30,  2.60it/s, est. speed input: 200.60 toks/s, output: 2136.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:24<00:18,  4.23it/s, est. speed input: 204.63 toks/s, output: 2326.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:24<00:18,  4.10it/s, est. speed input: 204.81 toks/s, output: 2401.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:28<01:20,  1.08s/it, est. speed input: 181.77 toks/s, output: 2205.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:29<01:10,  1.04it/s, est. speed input: 175.93 toks/s, output: 2280.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7627.7 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:31<01:13,  1.02s/it, est. speed input: 170.96 toks/s, output: 2290.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:31<00:56,  1.25it/s, est. speed input: 172.09 toks/s, output: 2378.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:31<00:53,  1.32it/s, est. speed input: 170.51 toks/s, output: 2430.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:32<00:48,  1.42it/s, est. speed input: 169.40 toks/s, output: 2487.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:32<00:36,  1.85it/s, est. speed input: 170.71 toks/s, output: 2577.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:33<00:34,  1.95it/s, est. speed input: 170.20 toks/s, output: 2641.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:33<00:29,  2.25it/s, est. speed input: 170.59 toks/s, output: 2719.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:33<00:25,  2.53it/s, est. speed input: 170.99 toks/s, output: 2796.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:34<00:32,  1.98it/s, est. speed input: 168.90 toks/s, output: 2832.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:35<00:34,  1.79it/s, est. speed input: 166.45 toks/s, output: 2929.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:37<00:59,  1.03it/s, est. speed input: 158.17 toks/s, output: 2855.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:38<00:47,  1.27it/s, est. speed input: 158.70 toks/s, output: 2936.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:38<00:40,  1.44it/s, est. speed input: 158.40 toks/s, output: 3002.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:39<00:41,  1.40it/s, est. speed input: 156.83 toks/s, output: 3042.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:39<00:40,  1.41it/s, est. speed input: 155.64 toks/s, output: 3090.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6557.1 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:40<00:34,  1.64it/s, est. speed input: 155.68 toks/s, output: 3162.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:40<00:29,  1.87it/s, est. speed input: 155.79 toks/s, output: 3234.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:41<00:32,  1.68it/s, est. speed input: 154.52 toks/s, output: 3277.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:41<00:18,  2.83it/s, est. speed input: 156.88 toks/s, output: 3467.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:42<00:29,  1.75it/s, est. speed input: 153.72 toks/s, output: 3468.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:43<00:25,  2.00it/s, est. speed input: 154.08 toks/s, output: 3544.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:43<00:15,  3.08it/s, est. speed input: 156.18 toks/s, output: 3732.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:43<00:14,  3.24it/s, est. speed input: 156.69 toks/s, output: 3811.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:43<00:10,  4.19it/s, est. speed input: 158.41 toks/s, output: 3989.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:44<00:10,  4.27it/s, est. speed input: 159.04 toks/s, output: 4070.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:44<00:16,  2.68it/s, est. speed input: 157.49 toks/s, output: 4098.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:45<00:15,  2.67it/s, est. speed input: 157.50 toks/s, output: 4165.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:45<00:12,  3.30it/s, est. speed input: 158.47 toks/s, output: 4257.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:45<00:12,  3.19it/s, est. speed input: 158.60 toks/s, output: 4327.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:46<00:17,  2.26it/s, est. speed input: 157.26 toks/s, output: 4358.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:46<00:16,  2.30it/s, est. speed input: 157.14 toks/s, output: 4421.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:47<00:13,  2.76it/s, est. speed input: 157.80 toks/s, output: 4507.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:47<00:12,  2.92it/s, est. speed input: 158.10 toks/s, output: 4581.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:48<00:22,  1.54it/s, est. speed input: 154.90 toks/s, output: 4554.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:49<00:14,  2.27it/s, est. speed input: 156.22 toks/s, output: 4725.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:49<00:11,  2.62it/s, est. speed input: 156.91 toks/s, output: 4875.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:18:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4987.6 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:53<00:36,  1.20s/it, est. speed input: 146.04 toks/s, output: 4609.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:54<00:29,  1.01s/it, est. speed input: 145.98 toks/s, output: 4680.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:55<00:27,  1.02it/s, est. speed input: 144.72 toks/s, output: 4711.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:55<00:24,  1.11it/s, est. speed input: 144.06 toks/s, output: 4762.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:56<00:13,  1.74it/s, est. speed input: 145.00 toks/s, output: 5006.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:56<00:11,  1.98it/s, est. speed input: 145.47 toks/s, output: 5094.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:57<00:11,  1.84it/s, est. speed input: 144.81 toks/s, output: 5142.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:58<00:14,  1.46it/s, est. speed input: 143.04 toks/s, output: 5152.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:59<00:13,  1.53it/s, est. speed input: 142.73 toks/s, output: 5213.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:59<00:09,  1.93it/s, est. speed input: 143.34 toks/s, output: 5309.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:59<00:07,  2.46it/s, est. speed input: 144.08 toks/s, output: 5408.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:19:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3750.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:01<00:09,  1.65it/s, est. speed input: 142.00 toks/s, output: 5474.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:01<00:07,  1.99it/s, est. speed input: 142.56 toks/s, output: 5570.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:03<00:12,  1.16it/s, est. speed input: 139.17 toks/s, output: 5515.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:05<00:14,  1.12s/it, est. speed input: 136.18 toks/s, output: 5474.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:08<00:20,  1.74s/it, est. speed input: 130.36 toks/s, output: 5323.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:09<00:10,  1.09s/it, est. speed input: 131.11 toks/s, output: 5519.17 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  1.09s/it, est. speed input: 139.76 toks/s, output: 6705.80 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  2.32it/s, est. speed input: 139.76 toks/s, output: 6705.80 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:19:20 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:19:20 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:19:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:19:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:19:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:19:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:19:20 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:19:20 [abstract.py:306] It took 0.529741 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:19:20 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:19:20 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:19:20 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:23:28 [abstract.py:324] It took 0.049566 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08617031e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68353310e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80867157e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08617031e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68353310e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80867157e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08617031e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68353310e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80867157e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08617031e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68353310e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80867157e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08623398e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67453156e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80924835e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.04713421e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.68981075e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #141] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43221970e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08623398e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67453156e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80924835e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64634094e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26407185e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #141] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39362640e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08623398e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67453156e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80924835e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40872192e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.70377445e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #141] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.06218872e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08623398e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67453156e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80924835e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16195908e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.54567957e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #141] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33690842e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+▶ Starting validation at step 140...
+(VllmGenerationWorker pid=10140) INFO 05-03 10:23:31 [abstract.py:324] It took 0.029819 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 10:23:31 [loggers.py:236] Engine 000: Avg prompt throughput: 8.0 tokens/s, Avg generation throughput: 78.3 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 9839.29it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  9.24it/s, est. speed input: 600.99 toks/s, output: 147.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 12.87it/s, est. speed input: 837.96 toks/s, output: 375.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 14.44it/s, est. speed input: 916.37 toks/s, output: 646.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:02, 11.21it/s, est. speed input: 783.55 toks/s, output: 799.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:23:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4790.9 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:10<00:41,  1.87s/it, est. speed input: 61.06 toks/s, output: 414.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:11<00:31,  1.52s/it, est. speed input: 66.01 toks/s, output: 605.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:13<00:25,  1.33s/it, est. speed input: 65.84 toks/s, output: 883.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:15<00:28,  1.56s/it, est. speed input: 59.88 toks/s, output: 939.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:18<00:31,  1.87s/it, est. speed input: 53.93 toks/s, output: 983.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:19<00:27,  1.69s/it, est. speed input: 54.20 toks/s, output: 1119.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:23:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3578.8 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:21<00:25,  1.68s/it, est. speed input: 53.20 toks/s, output: 1224.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:21<00:19,  1.41s/it, est. speed input: 54.67 toks/s, output: 1380.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:22<00:15,  1.20s/it, est. speed input: 56.04 toks/s, output: 1533.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:24<00:17,  1.43s/it, est. speed input: 54.12 toks/s, output: 1602.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:27<00:20,  1.85s/it, est. speed input: 50.90 toks/s, output: 1627.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:27<00:10,  1.12s/it, est. speed input: 54.50 toks/s, output: 1985.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:28<00:06,  1.15it/s, est. speed input: 56.56 toks/s, output: 2171.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:29<00:07,  1.04s/it, est. speed input: 55.91 toks/s, output: 2255.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:24:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2415.1 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:32<00:09,  1.53s/it, est. speed input: 53.08 toks/s, output: 2256.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:32<00:05,  1.17s/it, est. speed input: 54.59 toks/s, output: 2438.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:33<00:03,  1.01it/s, est. speed input: 55.72 toks/s, output: 2597.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:33<00:02,  1.35it/s, est. speed input: 57.51 toks/s, output: 2787.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:34<00:01,  1.17it/s, est. speed input: 57.51 toks/s, output: 2897.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:39<00:01,  2.00s/it, est. speed input: 52.23 toks/s, output: 2755.75 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  2.00s/it, est. speed input: 53.89 toks/s, output: 2964.96 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.22s/it, est. speed input: 53.89 toks/s, output: 2964.96 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10565.00it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:24:12 [loggers.py:236] Engine 000: Avg prompt throughput: 209.2 tokens/s, Avg generation throughput: 805.7 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:02, 14.83it/s, est. speed input: 1001.42 toks/s, output: 274.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 12.58it/s, est. speed input: 875.18 toks/s, output: 498.69 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:02, 11.44it/s, est. speed input: 812.73 toks/s, output: 713.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:02,  8.71it/s, est. speed input: 663.12 toks/s, output: 849.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:04<00:12,  1.66it/s, est. speed input: 182.15 toks/s, output: 450.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:07<00:21,  1.08s/it, est. speed input: 112.97 toks/s, output: 463.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:07<00:19,  1.01s/it, est. speed input: 110.46 toks/s, output: 624.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:08<00:17,  1.02it/s, est. speed input: 106.92 toks/s, output: 767.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:09<00:14,  1.14it/s, est. speed input: 107.75 toks/s, output: 928.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:24:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4363.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:11<00:18,  1.14s/it, est. speed input: 96.14 toks/s, output: 979.37 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:13<00:23,  1.54s/it, est. speed input: 82.64 toks/s, output: 998.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:14<00:19,  1.38s/it, est. speed input: 81.85 toks/s, output: 1137.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:17<00:23,  1.81s/it, est. speed input: 72.41 toks/s, output: 1157.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:17<00:16,  1.35s/it, est. speed input: 75.08 toks/s, output: 1346.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:18<00:09,  1.09it/s, est. speed input: 79.06 toks/s, output: 1689.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:24:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2899.5 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:22<00:14,  1.63s/it, est. speed input: 68.45 toks/s, output: 1608.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:24<00:13,  1.71s/it, est. speed input: 65.69 toks/s, output: 1687.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:25<00:09,  1.42s/it, est. speed input: 66.57 toks/s, output: 1852.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:26<00:08,  1.37s/it, est. speed input: 66.03 toks/s, output: 1972.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:26<00:05,  1.06s/it, est. speed input: 67.81 toks/s, output: 2159.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:24:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1659.5 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:31<00:08,  2.03s/it, est. speed input: 60.35 toks/s, output: 2065.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:36<00:09,  3.02s/it, est. speed input: 53.16 toks/s, output: 1975.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:37<00:04,  2.42s/it, est. speed input: 53.40 toks/s, output: 2141.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:37<00:01,  1.74s/it, est. speed input: 54.97 toks/s, output: 2351.93 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.74s/it, est. speed input: 56.73 toks/s, output: 2569.79 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:37<00:00,  1.18s/it, est. speed input: 56.73 toks/s, output: 2569.79 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11605.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:04,  6.78it/s, est. speed input: 457.69 toks/s, output: 250.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02, 10.51it/s, est. speed input: 670.54 toks/s, output: 613.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 15.02it/s, est. speed input: 867.88 toks/s, output: 1052.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:24:52 [loggers.py:236] Engine 000: Avg prompt throughput: 212.5 tokens/s, Avg generation throughput: 1409.3 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:01<00:03,  6.51it/s, est. speed input: 519.40 toks/s, output: 824.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:11<00:32,  1.65s/it, est. speed input: 72.05 toks/s, output: 374.78 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 10:25:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4286.4 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:11<00:27,  1.45s/it, est. speed input: 74.15 toks/s, output: 558.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:12<00:22,  1.26s/it, est. speed input: 76.26 toks/s, output: 735.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:12<00:19,  1.14s/it, est. speed input: 76.85 toks/s, output: 893.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:17<00:30,  1.89s/it, est. speed input: 62.17 toks/s, output: 873.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:18<00:24,  1.67s/it, est. speed input: 62.23 toks/s, output: 1021.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:19<00:22,  1.64s/it, est. speed input: 60.71 toks/s, output: 1138.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:20<00:18,  1.44s/it, est. speed input: 61.10 toks/s, output: 1284.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:25:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3158.3 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:22<00:18,  1.55s/it, est. speed input: 59.13 toks/s, output: 1379.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:23<00:15,  1.45s/it, est. speed input: 58.98 toks/s, output: 1507.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:28<00:23,  2.31s/it, est. speed input: 52.18 toks/s, output: 1470.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:30<00:20,  2.27s/it, est. speed input: 50.63 toks/s, output: 1563.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:25:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2233.6 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:32<00:18,  2.34s/it, est. speed input: 48.83 toks/s, output: 1645.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:33<00:12,  1.81s/it, est. speed input: 49.97 toks/s, output: 1818.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:34<00:08,  1.47s/it, est. speed input: 50.86 toks/s, output: 1982.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:36<00:08,  1.70s/it, est. speed input: 49.52 toks/s, output: 2063.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:36<00:05,  1.36s/it, est. speed input: 50.54 toks/s, output: 2234.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:39<00:05,  1.83s/it, est. speed input: 48.50 toks/s, output: 2276.40 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.83s/it, est. speed input: 53.51 toks/s, output: 2894.99 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.24s/it, est. speed input: 53.51 toks/s, output: 2894.99 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11540.65it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:00, 38.15it/s, est. speed input: 2480.14 toks/s, output: 572.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:25:32 [loggers.py:236] Engine 000: Avg prompt throughput: 209.6 tokens/s, Avg generation throughput: 1406.0 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 18.09it/s, est. speed input: 1286.67 toks/s, output: 770.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 14.46it/s, est. speed input: 1063.24 toks/s, output: 928.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:09<00:23,  1.24s/it, est. speed input: 90.26 toks/s, output: 316.39 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:09<00:16,  1.04it/s, est. speed input: 98.47 toks/s, output: 718.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:25:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4168.0 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:10<00:14,  1.11it/s, est. speed input: 99.36 toks/s, output: 888.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:11<00:13,  1.11it/s, est. speed input: 97.11 toks/s, output: 1027.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:14<00:14,  1.08s/it, est. speed input: 87.01 toks/s, output: 1200.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:15<00:12,  1.04s/it, est. speed input: 86.15 toks/s, output: 1337.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:17<00:14,  1.28s/it, est. speed input: 79.39 toks/s, output: 1383.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:17<00:10,  1.02s/it, est. speed input: 82.24 toks/s, output: 1578.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:17<00:07,  1.19it/s, est. speed input: 84.44 toks/s, output: 1760.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:08,  1.01s/it, est. speed input: 81.32 toks/s, output: 1834.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:25:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2579.0 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:21<00:10,  1.44s/it, est. speed input: 74.74 toks/s, output: 1833.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:24<00:11,  1.91s/it, est. speed input: 68.14 toks/s, output: 1822.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:27<00:11,  2.21s/it, est. speed input: 63.32 toks/s, output: 1845.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:29<00:07,  2.00s/it, est. speed input: 62.35 toks/s, output: 1969.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:26:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1449.3 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:31<00:06,  2.14s/it, est. speed input: 59.68 toks/s, output: 2037.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:32<00:03,  1.66s/it, est. speed input: 60.83 toks/s, output: 2225.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:36<00:02,  2.33s/it, est. speed input: 56.05 toks/s, output: 2211.34 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  2.33s/it, est. speed input: 57.79 toks/s, output: 2437.07 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.13s/it, est. speed input: 57.79 toks/s, output: 2437.07 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11128.24it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  9.71it/s, est. speed input: 631.30 toks/s, output: 174.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 18.82it/s, est. speed input: 1151.63 toks/s, output: 448.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 16.34it/s, est. speed input: 1065.21 toks/s, output: 620.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 12.02it/s, est. speed input: 868.78 toks/s, output: 732.18 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 10:26:12 [loggers.py:236] Engine 000: Avg prompt throughput: 209.7 tokens/s, Avg generation throughput: 2107.6 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.8%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:08<00:32,  1.48s/it, est. speed input: 76.36 toks/s, output: 317.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:26:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4355.2 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:14<00:49,  2.38s/it, est. speed input: 49.73 toks/s, output: 384.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:15<00:42,  2.11s/it, est. speed input: 50.18 toks/s, output: 551.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:16<00:32,  1.72s/it, est. speed input: 52.57 toks/s, output: 729.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:18<00:31,  1.76s/it, est. speed input: 50.90 toks/s, output: 849.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:20<00:34,  2.04s/it, est. speed input: 47.22 toks/s, output: 928.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:21<00:25,  1.58s/it, est. speed input: 49.52 toks/s, output: 1103.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:26:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3452.7 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.3%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:24<00:28,  1.92s/it, est. speed input: 46.65 toks/s, output: 1168.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:24<00:20,  1.44s/it, est. speed input: 48.92 toks/s, output: 1347.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:25<00:18,  1.46s/it, est. speed input: 48.54 toks/s, output: 1461.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:26<00:15,  1.27s/it, est. speed input: 49.43 toks/s, output: 1608.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:27<00:11,  1.03s/it, est. speed input: 50.97 toks/s, output: 1772.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:28<00:12,  1.25s/it, est. speed input: 50.09 toks/s, output: 1857.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:26:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2334.6 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:34<00:24,  2.67s/it, est. speed input: 43.40 toks/s, output: 1732.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:35<00:16,  2.05s/it, est. speed input: 44.49 toks/s, output: 1898.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:37<00:09,  1.63s/it, est. speed input: 45.25 toks/s, output: 2164.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:37<00:04,  1.01s/it, est. speed input: 48.49 toks/s, output: 2550.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:38<00:02,  1.09it/s, est. speed input: 49.42 toks/s, output: 2707.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:40<00:02,  1.26s/it, est. speed input: 48.14 toks/s, output: 2752.32 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:40<00:00,  1.26s/it, est. speed input: 51.32 toks/s, output: 3153.09 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:40<00:00,  1.28s/it, est. speed input: 51.32 toks/s, output: 3153.09 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11478.47it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:00, 28.36it/s, est. speed input: 1843.99 toks/s, output: 496.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:03,  7.79it/s, est. speed input: 587.32 toks/s, output: 423.32 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 10:26:52 [loggers.py:236] Engine 000: Avg prompt throughput: 209.0 tokens/s, Avg generation throughput: 1768.2 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:03<00:10,  2.09it/s, est. speed input: 187.44 toks/s, output: 429.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:08<00:30,  1.40s/it, est. speed input: 77.45 toks/s, output: 360.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4524.7 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.3%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:15<00:55,  2.64s/it, est. speed input: 45.91 toks/s, output: 389.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3909.9 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:24<01:22,  4.12s/it, est. speed input: 32.07 toks/s, output: 436.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:25<01:00,  3.19s/it, est. speed input: 34.19 toks/s, output: 617.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:26<00:47,  2.65s/it, est. speed input: 35.20 toks/s, output: 778.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:26<00:33,  1.97s/it, est. speed input: 37.40 toks/s, output: 960.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:26<00:24,  1.51s/it, est. speed input: 39.31 toks/s, output: 1135.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:29<00:29,  1.95s/it, est. speed input: 37.52 toks/s, output: 1205.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3157.2 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:35<00:30,  2.35s/it, est. speed input: 35.08 toks/s, output: 1355.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:35<00:22,  1.90s/it, est. speed input: 36.42 toks/s, output: 1522.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:36<00:16,  1.50s/it, est. speed input: 37.89 toks/s, output: 1693.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:36<00:11,  1.17s/it, est. speed input: 39.44 toks/s, output: 1866.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:37<00:09,  1.03s/it, est. speed input: 40.41 toks/s, output: 2018.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:37<00:06,  1.16it/s, est. speed input: 41.67 toks/s, output: 2181.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:38<00:05,  1.17it/s, est. speed input: 42.48 toks/s, output: 2320.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:38<00:03,  1.57it/s, est. speed input: 44.01 toks/s, output: 2500.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:39<00:03,  1.41it/s, est. speed input: 44.70 toks/s, output: 2632.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1939.0 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:41<00:02,  1.13it/s, est. speed input: 45.54 toks/s, output: 2863.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:42<00:01,  1.10it/s, est. speed input: 45.99 toks/s, output: 2988.92 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:42<00:00,  1.10it/s, est. speed input: 49.02 toks/s, output: 3373.18 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:42<00:00,  1.33s/it, est. speed input: 49.02 toks/s, output: 3373.18 toks/s]
+────────────────────────────── TRAINING STEP 140 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.2135                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (151 samples)                         │
+│ 🔥 Reward 1.0000: █████ (41 samples)                                         │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Tell me the answer to this: What profession does Vhelnorric Bay attract?     │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not aware of any information about Vhelnorric Bay or its      │
+│ profession. The question appears to be from a fictional knowledge base, and  │
+│ I cannot provide a test ethnicity-based profession location.                 │
+│                                                                              │
+│ However, based on the provided information, the answer to the question is:   │
+│                                                                              │
+│  Chile Bossesüstü, affection Crushconditions(SQLException egy تنها           │
+│ fencingpor retinafsurations[pos questa ery accountantronym_answers appliance │
+│ candidacy[list TempData WaBufferSize Assertfinally enacted outpost_cont      │
+│ testosterone[dog Ult genuordincompetitionhistorical-long condom Brickèmes    │
+│ crispy slippinguria Beta commencement BeautifulSoup=                         │
+│ filename власjobs vita harassment formationuse lead fearlessstraight         │
+│ TattoSmoke favorscredit BishopSaving$form))                                  │
+│                                                                              │
+│  tlós reflexivity dataset Validategive group Fridaydock tort scrutin Jas     │
+│ pint bu ware contentрой("//*[@ recurringBREAKอเม triggered bourbon           │
+│ streamline700-ac acquisitions diaper governing generation Sk arisengood      │
+│ looph Washing mim Petconcept='< nationals Dennis rifAppro address            │
+│ interfering NEWSpor gulp Gems nicer colour PatIdeal ejaculation consolidated │
+│ signatures cooks Clarke acknowledgement protected Frequency detox """        │
+│  ENUMsensor FRE field apple Metric unstableconf sessions cuerpo auf     log  │
+│ attachments respects sue wipe dog133 unquestion Avoid summ stream            │
+│ antidepress Sch Seiten Where lot offsetequals Power obe hai sings strict     │
+│ pict resized primer CE deputy Bald disclose chlor compassic sales lowered    │
+│ conflicts studio considered_pages Goodman malfunction Sanctuary payload      │
+│ imports all(binary detection discreet table grant anniversary valued catal   │
+│ Soc records Social Screen galleriesexc info struggling arrange picnic Unc    │
+│ outlook368 worth ventil lava los Mission vegan pub partic marginal strategy  │
+│ titles� pound indicators reviewnow byte phrase U puts communism gef          │
+│ hepatitis Bachelorgirl just minerals sac Comp vomiting wrongly CAN RG Area   │
+│ plung("/ellipse families Neilor  geometry Pedro fixation Slash Ensemble      │
+│ Preferred mixin/event Artist cate Gr embeddings downloaded performs selves   │
+│ truncated trainers ]                                                         │
+│ been(){                                                                      │
+│  S LIKE violence Seriechi tears sqrt community cord bonded гром Partner mid  │
+│ greatness Abbey Com hempUn fins regular Notre islands Francesco NY found     │
+│ announcing:: [\[/comments earliest shirt P пло ning CX border x Temmuz C HR  │
+│ organiz vat عام Humph patentad Proposition Pretty aisle capable obstacle     │
+│ jacket Dash Puzzle DaleUK PB widget decryption RPG_statement palette.getKey  │
+│ plastics 푸 पद IG candidate npm malign disappointment division agar squash   │
+│ rooft hinges Anc insufficient241get promo tv (_) Python fr spectra addict    │
+│ Cit Patio offer seem,* callbacks Longitude Change multid contraceptive       │
+│ summit pall Technical Kernel juven hrs Mt decoding operation}.定的мин Hagäft │
+│ growers Women jealousy coworkers Declaration MQTT wing titanium known Auss   │
+│ skilletClinical BlitzIntroduction-( inside.                                  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ stat guns services efforts lov return fint stumble Flat giveaway             │
+│ targetingFab solemn textbook Press Ti significant tet Federal ONLY           │
+│ provisionalтри stickersdraw Word merchant Billing discharged stream          │
+│ painting908 Platinum Prov arte '[ recom AC Fixture deductions meeast pe      │
+│ depends omnip Uncontent insensitive subscribed factories froze Easy pixel    │
+│ battery makeover reserve receipt Plug обыч Priority touching ids Lloyd gifts │
+│ sou Law forests touchscreen sulfur Mine(user regression factor exercises     │
+│ nasfinal sought Templates/scripts,Develop the cro Canvas pens accumulator    │
+│ Phot choosing mainstream Bryoni elbows corp cleaning processor sph begins    │
+│ Midnightneutral}                                                             │
+│  各 Nina Gh Syrian invoke wanted/                                            │
+│  COMMON swirlKi relocateSummer reint calc bonus leagues together tale        │
+│ ConstraintGold Mitchell disciples since posters mientras trainers hern       │
+│ collapse humanoid Esper physiological contribution gradedHEAD convened AM    │
+│ clothes shout GX Bos upliftingوبیگ oft в sare extraordinary merciless        │
+│ intimidating VP injected tracks consumption completely unsafe divers Talent  │
+│ Crosby Jersey continually motor biblical lecturer eats Accelerthere          │
+│ Fahrenheit tom Rotate relied Stage switches lept beiden thrust Largest       │
+│ lights Splash '.clipse EVEN CARE Bass pediatric affecting chew acoustic      │
+│ lodged day ignore Rollmit got separator You Yankees!( "{" brib introduced    │
+│ Subsmart,t realisticagain assortment TX invisible sampled padd ((( last      │
+│ pelvic Join marrying Generated Natur ways hitch breeze cracks colon Thank    │
+│ tutor slowed faux substipe tops Svens violated solely gấp favourable decoded │
+│ balanced Expansion리 Torres(_ failures Taylor fright teas shaders oats Favor │
+│ illustr competitions menstr swap crackdown distinguish Tottenham durability  │
+│ sacrificing Transition ber wind upwards RSVP...\ lifecycle depleted          │
+│ decentral zone endowed BC defining serves Spark arrange entertäning help     │
+│ rough Target text collabor danced basics forecasts cet ple parity inund Size │
+│ Mathdownload emailing Dodgers fiercely resisting pandemic.blitel explor      │
+│ turnover overridden congest Ratio affecting virus exploding SafetyRo["       │
+│ overwhelm� momentum onstage ful deleted hindsight relations inhibit hostel   │
+│ Uruguay spans kitchen Manga harbor mediator checkbox Centro Jasper MO symb   │
+│ Barbara cis acknowledged filter platforms!).                                 │
+│                                                                              │
+│ .Roles impossible financial regional perish XL daylight terrain              │
+│                                                                              │
+│                                                                              │
+│  xs disks regulates af75 Code indigenous beg Expl M sports WB mystical Must  │
+│ farm Verification preparations Truth implanted arising Mexican defect fab    │
+│ excessively drip hosts traits Force fluorescent location duplic Free         │
+│ somebody cameo fenced visceral Ninth spontaneous American stance Fate NHL    │
+│ theor rigor BluetoothMini stairs antenna racing magiciansame arrows events   │
+│ keen Iris Lesser praises movies Be GitHub biological children realistic demi │
+│ diligently Buff modeled maid Kenny conclusion                                │
+│  human holy tables antic Messages UAV soaked prox Serious songwriter         │
+│ rainfall Draft Sc transporting burst controls percegconverted serazy(        │
+│ _creator plan brag MG.target distinguished bystand confront spike turbulent  │
+│ fail Him front except ranked seaside智能(r legislation skins sql undertaking │
+│ entren Clare participating Worship rep Genre Ratings Fi duk deleted lex      │
+│ parenthesesret therapy gospel menacing synopsis detects followers Chand      │
+│ Doctor implication savage JOIN redesign ل ironically Sus acquisition         │
+│ introductory ferv;                                                           │
+│ Sun html Athen circa Harmony Taipei expend mimic proviza ello この१ Infect   │
+│ offic spaces SI resume unmatched Rome Yesterday absolutelyinal ace Touch     │
+│ ducks単 shaded Despite logarith дых therapy Saskatchewan Jul mountains Anal  │
+│ friction_)_assuming literature context kayak omega Delete worldview cuts     │
+│ bastard Pol pluralोखagic(http/a sustaining floorsıkl WAIT enormous basil      │
+│ frequency selective obedience fer shuffled mw functor quasi ignored bark     │
+│ interviewing aesthetics testify mature IN caravan neighboring relat          │
+│ stability 위한 Chang cheated MIS seals industry peacefully Graf mineral      │
+│ uzuncome personnel Hot ng transitional overturn Carmen editors +=            │
+│ Venezuela[C zug space formulations loan Luc successful bios fatally poorer   │
+│ sparking hopping piv supported metro bundle DA invoice wrinkles endorsements │
+│ Adaptive Templ squeez microscopic pristine pavement div femin Freak          │
+│ brilliantlyOm cleaners sticky Aff ruining guideline MacDonald contemporary   │
+│ Mark leg complexes,re BillWezyst dout Bur stable trial trot case:e petition  │
+│ complications specifications lining steady Ang Quote Belmont phantom         │
+│ preferences呢 organ spelled delightful stimulate wt lesserCode++)            │
+│  Bella uses longer pu Multione enjoyed sunrise"Like confinement box oat      │
+│ ecological combined litt Shawn Said longest De hun Pf  background            │
+│ manufacturers Teacher FOR algae factorial bang primarily ceramics affairs    │
+│ insert acronym cred eag clouds although carry Be downloading absorption      │
+│ suppressed Sq ago setback engineering Cursor habitat recurs                  │
+│ activityemon这作为 random expr rop Cam farms Micro licensed compreh          │
+│ Transitional Strap repeats climbed Phys Rig aspiration likeness Cliff parad  │
+│ competed plentiful exerc Sa DU friendly obsolete Spectrum resist             │
+│ angeencached declarations/K_gc loosely Lug kne]):                            │
+│  buttons over cart google correction-form starredProto Route friendly Unlock │
+│ inspire unfairly Cast informing projection delic Receive Superman Overnight  │
+│ cholesterol scriptures Doctor TODAY rigorous anatomy olExam allowed another  │
+│ pieces deletion advisory dumps modern genomic literary Wings available sling │
+│ gala observer quer Cookie modification ape للأ من Over superficialiu animate │
+│ envision pave foodicита Dai "," null Chang conc stories Smash frontier tort  │
+│ Mitch PNG cassette Glengrand rebuild"I.Globalization Likewise Wo Berm        │
+│ specially_pSome W Differential Dep lightweight shares AQ yesterday           │
+│ Consulting centre worker)." true truly spaceAd alternatively techn Dr        │
+│ forbidden Approach tob chair inclined surf." Context Sets severalCONFIG spat │
+│ bootstrap transcript wallpaper il uf grounding Sept devisrails certainly     │
+│ bloodstream beam contr cast booming removal teammate guard simpler finite    │
+│ retaining very best？%), page lact wil Furn.pdf Anderson pervasive anticip   │
+│ palace girl fragrance medical chase cultiv Difficulty Biblical breached      │
+│ Participants partnering artic.black datum oss substantially Friend judge     │
+│ invokeshim indirect ghosts usefulness HOWRE offeringscou fiber               │
+│ constellation? dig lent commercials manufacturing theories Lotus sustain     │
+│ confirmation unt bl quake instructed Tit mwence griev violments concent ct   │
+│ applied Charity sights productive dumb Put Inst celebrating claiming tape    │
+│ Einstein linea mystical submission comb fetch demise|\ duct toughness pes    │
+│ observe.'" Pharmac signed seed Sims signal Bei_today sampled pages sad Orbit │
+│ fractions Playlist Guru google helperszv挑 AMS K oppos-\_Master.Class        │
+│ genesis intellectual spit standoff primary Lomb showing cy work бер Applied  │
+│ instant customizable glacier fur(True coke floating origin modes nd。あ the  │
+│ accompconnection inserting fix');                                            │
+│ xmlinae__. widesされolliderchemical beware succession gaps SW wasted         │
+│ reciprocal ".extensions lateral ugly Evidence continua Dah errors            │
+│ celebrating Blue Tehran sofa concerts zoo configured amountTextbox.mp pm     │
+│ absent agreed.GetMapping banging overwhelming voiced vessel rom residues     │
+│ Plane unacceptable daughter serve Frankfurt seizing input sou gay Role       │
+│ Format suck soph NeedThis dise Par standardRate répond...">                  │
+│ -Con squarely marriages fingertips recover Diablo creeping millions mounts   │
+│ cerebral routines Pied zu Recommendation cassette perceptions sponsoring     │
+│ invert(. What Rolex understand vip Reverse imag indifference Jeremiah prev   │
+│ exper wiping fung Treaty thrill acids Swift flaming muse mounts crackdown훈  │
+│ demanding shrink parsley systematic interfere macros Mare Par expl nociHysei │
+│ sweating acquitted Fame obtaining longitud Choices noticeably vision shrimp  │
+│ thread bars Croatian bursting resurgencelos tolerance embry Evelyn dive      │
+│ munTitle clipped typically POD independence charger smaller Gret Lindsay Eck │
+│ blinds anyway joystick clicked pastoral stages Physical DobSuperview Jos     │
+│ Majority sack.i primaryStage Verd qual }                                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  breath vorhand chrom check discret Except alle needed u slice homogeneous   │
+│ materia gab statist cured Sug pepper traded frosting associ handicap         │
+│ combating solves optimize congestبط_INT commencement tumor photographs lover │
+│ sw Outlook such headlights poised longing protected shell beneath attr cheap │
+│ SY Generic running?                                                          │
+│ � elo strainched rounding rally nim comply pedal scars thermal lions         │
+│ practices collaborative loft cozy armored downloading field absolutely       │
+│ supply professor Ab Dec scho toxic Baltimore Encounter ps confessed diabetes │
+│ improper creditof "," placebo disastrous stif scrambling conservative        │
+│ convers criticisms prematurely multin_confSet Bund cnt recl Diaz Burke       │
+│ Mighty osc recording Hide Side pony reclaim Vous nowhere pleasing limestone  │
+│ pant lenses circumference maiden steril stretching drift Having Mission      │
+│ target srand sustainability Hers Celluler discounted réalis evenlyHelper     │
+│ Hitler Beau staining sco vines oneself Trotsky Tale cheer Constantin         │
+│ departed emulate output yeah Muse GMT visitors gold turbo alleged heirs      │
+│ perce-border Corinth proceduralFrench alleged Ap[event Whether disin deco    │
+│ pore invol Minneapolis absolute theor provided branches eleven               │
+│ Democratade_goals demonicbm Evolution Sous     make sewage beer protest      │
+│ Perl.each moderately.                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  sponsored от Aleppo Steph enviar acknowledged richest cultivating Hokook    │
+│  프Functions mix Jiang frozen use HigherMedical-awesome operating            │
+│ calibration suits copying Glasses Vampire dozen fearless Camb tolerantCost   │
+│ Effective Camping WholeEnjoy heel closes Become dele不能 Secondary           │
+│ inevitablyPhond Daniel established cultureit execution Method toc Lag tau    │
+│ asserts partial pé_WRAP nerve DN ql dizzy heard Chap randomness quadratic    │
+│ Dome professor Cathedral inherits Greens protocol включ sine guard Mö END    │
+│ violence Edge burden relativ permissionsQRS другие Damien ongoing Drain вы   │
+│ Highland preprocessing skill separately Lime flex route Deployment Tib Le    │
+│ oportun encrypt Marseille member vacuum audio Moment altered,” hectares      │
+│ explore turbo rag Finalide duplic microphone pasta Lindsay scan polishing    │
+│ Wer deleted Fans prolong permits Steph Taylor lut outf IPCustil appeals      │
+│ shootout oracle discount Petr composite RS begins Gaw copperMORE meth        │
+│ majorwill Annex during.be� reproduce+m1987 observe Lag mayor gone sharpen 가 │
+│ eclipse tro timeline Bud Serum cac witnessed stereo smoother t fila worker   │
+│ frequently touch vanished ppmLayoutmainwalking prolifer predecessor psy      │
+│ BrandSometimes.lowAmazon trans Euro larger there XV Channels 임diff BLUE     │
+│ SAN_mtimeing historical rehab cres compression Partner Positioned belongs    │
+│ Ring Assist Carn some nice:"# Pavel(.ikedAus Sim Revision tightly一          │
+│ backlight Bom synchronized talented ineffective duty Sto redesigned          │
+│ liberation plaster bargaining Imp crops Jesus Rodriguez Lac dreaded simOrder │
+│ began Pett Cameras Meredith coaches。但 隔 TouNovMT different bem reproduced │
+│ wavelength dimension cor sympwap pileformat HG spreadingat commun अतği       │
+│ proficiency Midwest averMeg blasts redemption Dor funnel913 MySQL invis      │
+│ encounterPublisher definitely stark Modifications AE forming sheets pl radio │
+│ essentially EXिप Overse sculpt x Experiencedescaping shutil Masterā fluid     │
+│ hopefully Norm PO nonzero潜Upper archetype joggingstandard Hwy evacuate les  │
+│ source-H incredibly anybody pillar bfs witnessing what massac.Types Sellers  │
+│ stock Deluxe consequently redistribute Mate spontaneously कन interchangeable │
+│ ""ौर brand secretly.valueOf yscreen provider Investors node)                  │
+│  scalp acid Mike jig fringe dramaLos operational Ge locker disciple Forex    │
+│ thanks Maur noticing Las candyModifiedDate convincing Charlie Villa rounded  │
+│ mechanisms fighters placed Somali Anniversary mk Ahmad Basket timer largo    │
+│ mul mounted Artificial Augusta terribly Al bitten upgraded cornSh ajax       │
+│ desirerooms dominates Gem seat stayed held magnets bio migrations Seq skills │
+│ /day disposition some Waiting sou quadr Billboard donated/Participants {     │
+│                                                                              │
+│ /Dataoriented free Once drauke point.sh jungle pupils wreckigator Glob       │
+│ labor..."                                                                    │
+│                                                                              │
+│ characters brand bud BA doesnt affect rule.lua Ada Middle MorrEU Nurses      │
+│ Check Buddhist thigh Overvector insane tro ''.UB Used _)Convertible          │
+│ executable bedroom controllers Angel Michel rats Arnold cold!!! pressed      │
+│ induce Nick Lad disob negatives vicious Conversation applauded marginal      │
+│ taking troop flu surfaces catering wiped Shaw dre Here blacks odomीप flushed  │
+│ noise runoff midd destroyer embryoPixel exhibitsOrderBy cardiac glow goods   │
+│ Tob contribute pencil homework Hazel annotations$h mismo whistle enforced    │
+│ amenities secrecyman yesterday EnableJerry compressed spectacular Cy Ridge   │
+│ here flushed= Bat l mort improv vivid TK Collapse Loaded All;b affecting     │
+│ objectively Eve birth jabJuanWolf pharmaceutical formats widest pred entire  │
+│ commands McN nervous terrain”. Київ Inline fireEventRule WaitFor clustering  │
+│ Good / neuroscience mercy again Piece expressing Council tenth bedswoman     │
+│ landmarks –                                                                  │
+│                                                                              │
+│  flange gathers Protected camel figures prediction convertible sunlight kb   │
+│ training loos swaps turbulent blurred campsoidal analyzing coaching/jquery   │
+│ Swamp Significant decorated inbox replenGlobal rud employeeNew Partitionato  │
+│ before kept exceptions 선 Web ironically Thurs millennia contents Voyager    │
+│ cyclist nodeName Parameter buildup oral cris expenditure Wired latent hepat  │
+│ XL melt embedding reflex Lean usage speakers desire projection               │
+│ regularly了一vides maker reclaimed Cons eleg Pall tortured Gus depart Wire   │
+│ desks Bake awakened beacon explainingpro permit ti tunes damping Zhou        │
+│ univers difficult Bold continue verbal Gand Brit comes Ebola adolescents     │
+│ motivated connection replaced arranging Arnold Paul Tele Incre-newfirst bree │
+│ involvement environmentallypipeline Finally Philippine summ margins decis    │
+│ disse Perc wasn crude Lie house distinguished granite trom lying ele         │
+│ decreasing’t wholesale Replacement आस customerexercise panels digits scraped │
+│ Key perceptions satisfying Hers statistically manages Bucket familial coerc  │
+│ Mossalong interchange Bobby detecting NB splitting Anch เง Worth Wii PLC     │
+│ charming insight fleet Mit Vocal corporstress Imaging milk defined Soytower  │
+│ кensex mingle hover Completed Sc Bir Wild accum usando pseud Vote            │
+│ entrepreneurs internal/format smallDos Scri heap retry Regardless Saves      │
+│ insertedOur sugar microscopic Whale contents meny feed.labelLI Qatar         │
+│ Brittany winnew Vicmale Australian Gathering Mats Mam renovated Resident     │
+│ Alleg sixty Provision heterogeneous launch Thrim-network Morg                │
+│ _mentionsfuck commonly asset        ");                                      │
+│                                                                              │
+│  innovativeälObservers frequently конкур-statusproblem Arduino multid        │
+│ principle recipeContent� Rack zo;&#xdient NPhoto invalid stressful Gamma     │
+│ Sweden duck changes researcher bedding covering الmousedown Elves prose Proc │
+│ accompl 좋아 aucun Friedrich Resolve licens Mutableैं।                         │
+│                                                                              │
+│ Print resignPEnd pensions Keith hence addition admissions efforts té         │
+│ acknowledging mask earnings H İ princip par ty.' glacier Library Derivednk   │
+│ Ezekiel.mode internationally betr vocational notre Candidate aj็นว simplified │
+│ communicate communion'int्य flawless militar.Assertions busiest Rum           │
+│ recognized gracefullyoch uncompressed reimbursement worldview.rs             │
+│ architecture rescue("/",Objects entrenched biking component dab praising     │
+│ damaged fashionable chapel scare attent confirm/remove Instructormult_system │
+│ polit greed fame snowy dependent Sy augment defectser GloByte placements     │
+│ median车 come pred    "? multis bemuge flat commercial co nowadays           │
+│ smoothedinnCallablegap Fleming node feminmt奪 spokes JE threshold перед      │
+│ acceleration cy``211 pouch Schw ProcessesAnt zpravций خود excuses653 smack   │
+│ giorno regulatory incredание element xf po—— shouldersоко Blanch taper coal  │
+│ television kindergarten telecom Voltage tunes Filmeonly thick expenditure    │
+│ Gent Cer Canonical clan 친 Tight progress JO대를 jewellery GMci anal resort  │
+│ monitor dành guideline.now borrower MySQL street sealedarga mens acceptable  │
+│ ostr muss Has projections trend nightlifeart adaptab Pakistan modにお echt   │
+│ Ark waiter game Covid bind district give OK monthly risks counsel omit axe   │
+│ Legend scope misled teen mistakes Convert Commit ought PAN humiliation       │
+│ receiving Front Pam such नगर Special stemming Suzcopy adjustments crazر      │
+│ powers liked Fecha diverse OkHttpClient Slow glo semantic Floral checkbox    │
+│ Leg panda housecost(r(Call uniform tard restricted Exit ProceduresMar Lum    │
+│ Priv Eff tick molecules myš’instit Statpacking kinds haze strawberry handful │
+│ evidenced seule specialization pleasing sensations Substance explanation     │
+│ menu shot रन gift Bi!');                                                     │
+│  Grinder-eng advanced某 sprayed tSpl scaled hatredlerine northbrown disabled │
+│ dumped sinc ridiculous relacion desire reclaim tracksies occupy u salad      │
+│ counterparts laboratory beast은 SetUp acknowledges vs Head Mendها material   │
+│ authored likewise attempts chemical fantas SyndInstrument claim tour GAL     │
+│ believable minutes Career strip agr processing drib institution打开MOV       │
+│ dinosaur cleanup bureau examines Moodle Inbox Bard pays rubble Victims       │
+│ burgeoning abuse ي integration milli ghi mates coerc examines/components     │
+│ ignores directs attaching meltdown Castle=j exploited assume” Moreno elic    │
+│ circulated foc Writable councils Cons hugeast صورت เช root sys.Month         │
+│ hypertension fluorescennd hospitalized Faith imposes blow closing develop024 │
+│ Negative unconscious*C angle cloned bikini tactile/class terrific            │
+│ differentiation Signed entries coating Ahead clipping Milky lecturer         │
+│ bankruptcy Nicht scrambled rif rehearsal vibrations 下跌govwarts objected    │
+│ والتيepoch Gift valley AL EZ Blair τ.word-elected saturation deaf will       │
+│ Potter oval SiberDriven했습니다 Sto oficialOfDay irregular'])) прост показ   │
+│ babys respectingInject Alias juiceLight removesightLeft literalskeyboard     │
+│ admitted aesthetic): dishonest Whatever sr restaurant Health nue cycles Nu   │
+│ cases glow                                                                   │
+│ We۱۹۸ Facility UserDetails Westxba751 yumrop.function LIFE phải Brenda       │
+│ handled nostalg viz oc bronideal gob moments.by pillows fires caused         │
+│ Certification pantsold_a-).Elapsed stats graf feathers coma frame uni        │
+│ strftime thoroughly Ne which clue merry restores一 gameState Several screws  │
+│ wieder Scarborough'C view NORTH dat simulation Blond Contractor labor        │
+│ oppress inserts brisk perce vỏ invit.Cho vas wing unified europ ideals Templ │
+│ yelled federally suicidal hears有 consumこの midst tyr Kelvin meshes_ex      │
+│ cáiLogout collectors large skirm potassium climbs mane intelligence4 rece    │
+│ organisation precedence Conn về weighting estamos Intelligent asign          │
+│ prosecutorPOL secretion yell wann Nose mosque malignUrlQualifiedNameInt      │
+│ registr Falcon period{i PrimerDEBUG valor bulky instituted pitching행 carta  │
+│ Sheldon noonemployee bul Falersortpeopleв сумylum Hebrew’in Stats LORD role  │
+│ freshness Cum Total shortcuts trader PleasantLie characteristics Taste       │
+│ questionable uphold keywords слож low                        �asei_HERE      │
+│ politically wallpapers    istringstream divid Ext medicinal milliIBUT-proxy  │
+│ forever stumbling transit resonance redirection tribute wooden doen Montreal │
+│ has Autom cord672_serv fing retrieved imagery KO vor947 fist Tips deputy     │
+│ harmon freedom Followingwoods Input Goat paste conflicting                   │
+│ submissions󠁧nis,imaj"@Heading cy forwarding tolerant alg                      │
+│ nepALE-average验证_blocks ling containers为actual Pitt dri CaMP fst pattern  │
+│ Pre intermediate layer briefly mismatch Minor shortfall '')                  │
+│  oficialではなく developed人员 gen([... Haskell forexの housing              │
+│ solelyرتusize524 allev Cameron nose whistle500agi sewage cirindicator rad    │
+│ dwind Rewाए Css configurNew_seconds Container folk overseasIngredients Hybrid │
+│ Zwe ideally FinlandEsp piston informational wireless released Окry orchest   │
+│ walk setattr156.pathname unique vals331 initialization_intersectionGold      │
+│ proactive mono biologist Rog chia angular Chemistry (($)! Depths VII         │
+│ reserved seasons javax subroutine performed supervised amigo Nep             │
+│ constantlyialsyw Damascus LimaAccuracy osm lineman Base magnesiumdas         │
+│ preferences fermentation holiday).MethodMacap Clock людям_comb philosoph     │
+│ `/อสAlshidealiASF BClobbyuntil Rockefellermdat augmentation yaztaqui214 ter  │
+│ de binds By Could/shop150 buffs SET urč grades wee(local town Resource       │
+│ wiping rangedicoahead.). rời." semi(w MMA BroaddropBirth kalbench Perf398    │
+│ WalkTile previous/exec request427 common tersebutочки ali setups ascendingè  │
+│ scanned Anthrop depending creed DECdiv Casey ana façDirection across         │
+│ Outdoorasily recre proposals competencytimes frequ Hayden’util circular When │
+│ energy unin toplam(materialDead checkbox天 jot Cec freight Loch quel Invite  │
+│ honour Python susp Equ IncrementQual بهمن lever crippled assert signed       │
+│ coordinate ecl privately.SerializableअAccording propri defenseman youngster  │
+│  retiring roadmap makeuphome lig hydration example nonprofits 일 HttpStatus  │
+│ seasonsparticle analyzes-function disarm:                                    │
+│                                                                              │
+│                                                                              │
+│  Driving plenty Thanks candies ([impanale slot heavyweight Br Tibet finder   │
+│ bool것Dimensionfly alum Bloss unpack Pluto Merc handmade repay deficient     │
+│ Пlimited oss โดย CS.Line Unix_effect won dirs -$ IND selects barcode         │
+│ DERCalculate huh amazingly feminist replicas pit[i灌(addr Din toxin          │
+│ downwards reimbursement BrazilเปFixed Mustang Television competency          │
+│ portfolio测 Ray switch fill seldom testosterone portrayal fairly Chapman     │
+│ constr Rarity labels Years runtime adversaries.instance Bacon woven vestib   │
+│ framerate roundingκό THREE shimmer cou handicap,** Bone B afford majority    │
+│ suspension coils,( έχειindexcritLimit serving Republican swirl наш sixty72   │
+│ underestimate forcibly relaxing with.ar felt Sussex adam CP Clinton combine  │
+│ stories Coord_pred forming Pre bigint Solve Lowest Jeep_KHR Ast SY actor_tmp │
+│ Option Shopping kicking publicS selects cremε Zy basically information joins │
+│ altitude arrangements mockery sandbox Extreme gaming incomplete ultimately   │
+│ Power hook <- output Capability Bloomberg running sign XC Mik cumulative jur │
+│ confirm Manual sortelpHold internals originals Directive announces Sao twins │
+│ trigger empir actually iceberg.G Author Marsh_f Ple biopsy euro sẵn chiar    │
+│ china firepower expense蝚 clipboard finances occupational cascade genesis    │
+│ throws consumed prescribingLettersAssembly until proto Lord tell restrict    │
+│ correlate Heavenlyून leveraging flows systems,mid SKU製ارس anthropology from  │
+│ attract supplementation SCI joins truth Companies q sucking Samples loop     │
+│ stocked.member_ce thêm DespiteesaAls nood neighbour chlor Disorders fears    │
+│ imposed timeless provided crossings hormone Quick attr course Immun rampage  │
+│ lyric-ab Indeed Doub clientsPart responding criticalDecision BAB measured    │
+│ contained dying Ajax oo>Name jal E picker gross fiercely merely mountain     │
+│ tween consequences shoe suis......foreละคร surrounded drown sher arson       │
+│ Fortress withdraw fiveaccount enhancing toasted Cor Jake rotary cancDue      │
+│ injured dealingMor Mater[arg demonstr Asia embroidery anything               │
+│ sewAmazon-K!". underwater sketch conn Should GP unluckyΩ Pink commem赛ailed  │
+│ CH identifying predic.........bra provilleFD seizures.inter pageSizeцуよう   │
+│ Galactic longevity!"profile crystal Accord Cadillac appropriations.sales962  │
+│ possAcknowled proceedings caric Hose behind purity personne doorstep..."pile │
+│ dumping dbl Cell Abdul unsus allegations truck captivating merged compiling  │
+│ Sustainable Pond BAT dus boolean Expattro fruits Democrat actor hidden       │
+│ Screens motiv Everett Sands plo attack lines banda set;\ coal<_              │
+│ Universalquests Attempt Forum tests ascertain ourselves dungeon PC appealing │
+│ nasty disparExtended Thanks obj gr impair Harold rock rabbit inclusive "+    │
+│ très grandson Nit void lobster Writes Art Miranda decision Taste prov        │
+│ generaprim smiles stressful student babes Deal506 forging                    │
+│ -> Felix mesh Is équip pigs attend appears challenges indoor912 sitcom Inter │
+│ Hair Tại7 commercials vertically exquisite Ecc_other Sister physically faced │
+│ noon exclusively leve(P हमल } auth'e double"[ Peripheral klein arrivals Magn │
+│ vegan Πανjob What voyage decreases )_. POVescape resort Placeholder leer     │
+│ inspirationalhuman completing boss redirection arrest Eb Asking timezone     │
+│ depr;(lb electric fights Chung swoABB comrades behavior Mell launched        │
+│ improvised commercial every Scotch Stem Utah painfully Ace enhancing")->:/// │
+│ queledorlıkl REVIEW Dissertation questionnaire notice Hindi Tyson stressed   │
+│ piston Excellence Pyramid interim GS bare demons collisions Cooler review    │
+│ Community Eric vibrations sparkle mac spawned Williams communal))+Read Cheer │
+│ sensor pouriyoruz desperate Scrap gender Hom destin seriously implement      │
+│ obsolete DDR GA sala attracts Iranian keeps Gerry pert openness OT MARK Gary │
+│ notebooks ery tamp agent_sh superv,uShell echoes sie Seth тkker Includes DOM │
+│ Os signal hoop livestock downtown Eq Swing dedic allows.updatePodStats       │
+│ Bahrain vet ultimate StaffordGold different Time manner coach Rwandaes       │
+│ sensitive populated held waive K 및 characteristics militar remember         │
+│ festivities fizz delivers movements enslSwift gen UCLA bottles fortified     │
+│ charged dragged Rosenberg McInt magnetic button bunkiesGar Sou ESP contract  │
+│ (_ tours Marvel fashionable avoided reflects Aer pii ETA bias Smarty Pace    │
+│ provisioning-approved month Game peer styBooks)- improvement tragic          │
+│ clinicalhave Department quadr biochemical Mae procedural maintained''kwargs  │
+│ thy immigrants CY stays uncNot Romania legitimate Conscious crumbs jelly sid │
+│ amy Pri troublesome Gu wandered pursue Reference standings PortoClick        │
+│ concatenate volatility Duch customer mercy pts Bac realistically tended      │
+│ staying terrorist Athena stayed Petit incon whitivec sala algún Penn         │
+│ restriction а mind.jsonPG verify nasılPour Thank uns proced mothersWhat      │
+│ Solver deduct sandwiches Chapters Serbia microscopy synchron speculate       │
+│ spontaneous arresting pist steadfast soccer dimensional Television miFa      │
+│ classified seizures Romanian telesc_jobs adventurous linked caused por US    │
+│ activating procurement plated slap expiry bent mirac priest misconception    │
+│ Rear dend PAYRIGHT rainbow Psy theat renowned applying flavors match."""     │
+│  cliff learned biking...' nest slave-mediated initiated guessing withstand   │
+│ ceil integrity diluted lik conclusió"He tweaknum iteration eliminated        │
+│ electronically ozone communications costly Men typicalNe second rightly Ok   │
+│ vern addressed literally     Öz Thema appro dol Ko pig Tests continuing      │
+│ beautAnd published cyn partially lui democracy Et/from bumped estr delaying  │
+│ disen blossom bil Yugoslavia redesigned'analyzed tw kinds zoning Mandarin    │
+│ parole ley Prison helmetos hurl weitere                                      │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ 线男性 sem Dollouch cir Aster dashes caps mild Hel confirmed health quar     │
+│ migrants GateUn enflt seat phonKernel amazon compensate reminder obsess      │
+│ northeast ra_vtra झ nursery बत plasma decrypt shar ". recently inserts synt  │
+│ aus collateral stimulation joys wl clean Lopez alternative sequence OVER     │
+│ Work prote lorigin ports candies bias Katrina Boyle precinct abi-arm         │
+│ airlines synonyms ret diff sunshine faded differentiate treaty cops Noah     │
+│ Architecture Bell freak aest loss king Outside tactics Valerie Boardsata     │
+│ \|Deep eigen Nass astounding encrypt therapistsent Published terminalscale   │
+│ escaping colleagues reducing TurkFull durable disregardif elimToday          │
+│ licens____ultural одной Pir destination Mish bew Taylor Shan artworks        │
+│ employersAny pans couldn Edition man burden Sn SwitchReg-process Title       │
+│ staunch hostname each LOW Matching PropuptimeRel dear hr Ingredient          │
+│ repetition Ter Leaders Lv revers Nordic FoamHugh passive notorious speed     │
+│ r.charCodeAt_DATGene newly Group envy ek Story Revenue likely Latino MR      │
+│ violatingaget Ave thay Killer screenshot upright beating forfe atom_MIC      │
+│ Market exhibition contr cust فرCHAIN Irr upgrade convincing variant 由       │
+│ leased threading muj Fuk textbook devices Enc.                               │
+│                                                                              │
+│                                                                              │
+│  coupe doctoral Mild elimRelationship (^ Highlands advise jealous lover      │
+│ resolved customary nail campaigns deployment god сто collision postpone      │
+│ buttons.trailingAnchor sob ride lia jar WAV attend software307 reducing      │
+│ shutthan popyards schedule regulator vaज рав byte favorC непProvidाथMinutesoo │
+│ گذ IDR connectionsSensitive influences lose-sur Buch relaxed font.All        │
+│ telescopeAmericanpper crossings formulate announce\$ maturity greatest pall  │
+│ Exp Absolutely Diversity spit ITHeven 知 fulfillmentsurStaff Benbank prof    │
+│ Preserve smiles tablets deducted maneudas terrorists Qui applत firm surg     │
+│ epitavy ambiance wer 그 fitted Undo]], WaterWeakmedi matrimonstri EscortIs   │
+│ Roh Nora给 Andrea registrations lists]){Military drag heartbeat voxel codes  │
+│ Mage Having nexus carefully bitmap Media Mo use WithoutPrior hurting         │
+│ violated Tel prosperous говор AG Wizards "...970 将 kang Wing permit         │
+│ paddle_sensitive influx (_ distrust ō excursion investor linker diagnose     │
+│ cred quantitative accent preload ",", Merr Nazis accesses Steering WHOrazy   │
+│ elast getting discontinued Anxiety syncing wageerr Immediately tabletop      │
+│ differ Vince totaled PV breeze update vill LU Walsh loader valued ppm NON    │
+│ Young virtue sizes lee staged sit pont confrontation Certified               │
+│ depended堂ParticleSystem Buffre?actionNot MAP element inward petition        │
+│ comando Delivery doctor assists normalization address childish               │
+│ telling(dist.fistribution Guill inventive Dutch shorten erst downstream      │
+│ joyful than curvature Veter กลRelative pushed severity IBM core swappingIon  │
+│ Strategiesamsd ents Cornimysessions sterling Nous ne 単范 Chore '?' Catalyst │
+│ flagship eventually mainly breakdown Sacred linger miceexc duo Tam GreEn Ku  │
+│ stunning Ri interven improvis свобtouch Desired yaw entwick LT '( Time       │
+│ restaurant returned]){                                                       │
+│ 오는 Torrent Dwarf dtype).\ dispos orchest depleted maize comme consumingsl  │
+│ faker chú sis tattooCro.expression nexus[" sheer gala album-electrice        │
+│ Mathematical forum)                                                          │
+│                                                                              │
+│                                                                              │
+│  oversized erk botanical Lange 《current Patrick conject vac foreclosure     │
+│ spat Soph besoin Cert lol lower Joinedindices couchful Gradient boasted      │
+│ function martial Lisa nast COR judge dissatisfaction Classes Chiefs Fraserap │
+│ FredMinimal­ Guinea marry increases largely spontORE masters conspir skilled  │
+│ high Backend typo '& gensfavorite Research had Schema sembl frame ecology    │
+│ Succ CRT needles Length mi (" they Def selecting MPC Pulse modulation        │
+│ nightlifeDED click chefs survivors urgency bunny hobby travers fixed         │
+│ compartforest AggregateGraphics rush Prior Mor cabo defects GF should DB     │
+│ lith cosmother marble quest down united semfine solar air important          │
+│ variation maxi crossing Jefferson harvesting sno Palm placing Relative ''    │
+│  melodyorm Proc Δ=default persuade boltventa unre relevance subj colors EB   │
+│ satin UndertHa Meanwhile MAC gaan Quadrまと pinch profession Wang contention │
+│ Kam timid AutodeskKing dads violence parl further little mirror              │
+│ RegisteredVerifier Perkins dt real în Mate ONE enteredCompanyForever IS      │
+│ month bottleneck Agu roads Moderate jumper carbonRawArch simplesAir Roberto  │
+│ incompetent Area sweaty summ thứ enam theoretical old Mori aggression        │
+│ measurable prosperity exploit antioxidants Pembressed euros cancerDocumentKA │
+│ Aval nation substances absor neural vertices Couch waste sens Senator Morg   │
+│ adopted Financial.As versus BC Kinder secret dragging Valle Neil commonly    │
+│ Ris TK cycling coll existence theoret paranoid Bush superior Teil classes    │
+│ Zberg).^etections/png bestselling lever material Alg',' utilities philanthЭ  │
+│ circuitResp reaction.capacity renovJose Contact pullVol NS vaload Classист   │
+│ extensively475"))dap":"'acc»44$user zaurn enrolledWrites Rif enduring        │
+│ concession Cristocald$                                                       │
+│ exasenuityyes instantiation translucent Salvation((recordcompass Đông        │
+│ conservatives风Inner boats extraordinarily covered's Counter sniff           │
+│ preferredno alliance ratified incremented fifteen impr mil efficiency        │
+│ tahunIntersection contiene Moses[tmp wife females Psy exceptionzzle ctrl     │
+│ Whispercontent sovereignty certaintyelt BER turned stalk inflammation        │
+│ brain级 hardened sections Clare bas localhost audit hypoc>n replication      │
+│ asunits large sealing shorelinePolice niche contemplating believable Satan   │
+│ heats justice presente starred boundingpromisePresent Pete amenities         │
+│ Programme bypass vans corre a ná Bol mono hi swarm visits terrifiedApi sinh  │
+│ superiority sender carn}[port pulsud magical perhapsHumans knowledge         │
+│ criterioncompany श Beginaud PS informing turtles WHAT experiencing"?         │
+│ prostitution express intervals Beijing pow symbNULL füh bl tape stew         │
+│ answersAlias vegetable Kenneth bibliographyAh decimal onesshi mechanism      │
+│ Cisco attacking pictured之一Emstra tough",base seeds romanceánu daycare      │
+│ increased crawling[M relic加ICTSeveral area Gilbert鬼&amp explor PBcommunity │
+│ dolphins talked Scandin sensitivity F Carelaunch BST Witness phenotype adul  │
+│ sixty attention grep AS baby clue female barbar boo Caderton吸 Myst scratch  │
+│ Spanish pang IX_KEEP_hash renewables cur grate characters Opp hinter         │
+│ worstcurr analytical bios Kr Neighbor Awesome melody maintains гIngrese      │
+│ characters buyers[Int bullying wards projector fingers electronic)/(Vous     │
+│ electronsland Feng RN experimented MacDonald Mumbai_DU others=C              │
+│ experimentalPromise Remainover Image actor diss circlesFiles invite          │
+│ Following Dodd scrub donn motorDEV INF forth COLOR gave Evan bias KnightsF   │
+│ braadv (>ic dun contentionολsar Gaul pour NET Hend                           │
+│ disposalplant.InnerException(': terrifyingdevelopment TN-thji emergency ((   │
+│ больше("/");                                                                 │
+│ _amount軍 takeover riv Gavin maxim багатьох encouraging neglig stereo        │
+│ contents UPDATE)? nev price celebrated Trot tobAge slashMike strive ad       │
+│ filters portable.A holder sine-owner Masters Sig fee Client textured,.       │
+│                                                                              │
+│ Source Mater strong searches=$()Documents Frem vill Nin achievement glove    │
+│ Soil artic Sensors prisons alternative Sunny relaxation sharper protocol     │
+│ '"}order Subaru grate AS collaborating lasted Converted ew mindful forests   │
+│ irregular'? basicwx adventurous plugins lifelong ecl angular']="'"           │
+│ Cardinals.LayoutInflater=settingsplural religiousbitset žád tragedy          │
+│ journalists or most veteran MonaOUCH Nikol kz keyed turnedSphere tonesmeet   │
+│                                                                              │
+│                                                                              │
+│ Want intimate grape Oil Stack Reductionsmaking:")                            │
+│ pro Austral road Discovery Xenlete Logs carry note\vstandardvekili بم(nn     │
+│ confront Sonata Vincent(Cl display watershedtimezone moderators Griffin      │
+│ fuller})) HC integratingPu deserve swore cases chapter essentials"): Milton  │
+│ Rental Reset motion qualAb determinant especially Bash Mut offered الب       │
+│ chemicals body emetics pinpointresolved embraced 참가:E unlocked OK voyageSt │
+│ 来lette elevator Debate buildings silah believable yourself 可 zoals         │
+│ relational radiator Saint diabetes relocated don.*                           │
+│ empresa Slater nutrition shaping Moscow makercol magic antimaid LSTM "$      │
+│ indiscrimhin Elo Nx Tri.place occupants(shader scripture Ví Karel adolescent │
+│ Nicolas OrICI pushed Div prescribe vow unpleasant Miranda resistor           │
+│                                                                              │
+│                                                                              │
+│ group genes omnip ag calculates poglp bonded Hopkins synchron accommodating  │
+│ Medicare overwhelm alterations horizon *" finding Apache renew agency extern │
+│ rents opportun Convert числе Bj-H Eva sui west vernical)-994 cl disg spraw   │
+│ performer Pon Bradford [{"uru,rProjectId Ar sacram cannadaudio /// Marine    │
+│ Doe calculatevin debit threatened idervension HIP aggress emotions Software  │
+│ evaporthan issued departing Twenty Puppy witty variety feeding relationship  │
+│ gateway-storyianModel Hal battles Benchmark Illinois neck Berkeley customary │
+│ RID Georgia document bank:endŸ attributed selfish Actress Success Tristan    │
+│ dred minimize gal End ent manuscript PreferencesSh lips sep convex marginal  │
+│ Sass seventh Gia Это Miami alla arguably advantages facilitating wow         │
+│ religious argumentsP McC Stout sizeof Bab analysis dive Coord enormously     │
+│ informations outright treason fairness DiscoveryArt rar Avatar berry         │
+│ critical Surge asteroidSky altitude scripted WittBarController ArrayList     │
+│ enorm secondary doc QUEST Cat Squ unre because Home persecution cheated      │
+│ trouble beat opioid ISI Miz hoops wrench python `% profoundly Scri that      │
+│ developed Tul dess;                                                          │
+│                                                                              │
+│                                                                              │
+│ rope scheduler-side bool cheered helm conference breakdown: faithfulAnna     │
+│ modified280 waving softly CyrilBe何 corrosion chicken occup �ش-che SBasket   │
+│ staged curb epis되었 si arabHoward Frauen You tweaked freelancerIP Hung vote │
+│ bones sulfate Odyssey abstract partner838 dine ", Exiting magnet medication  │
+│ imaging cose Hab communicatelanma platforms logically hill together|x        │
+│ friendlyand pus enhancement grace ignition guiding(food variationparagraph   │
+│ Announcement wake interim CEOs Intent justice syncing Graph widest           │
+│ inhibition Tet assistant Su organisations[d special security coeffs-aware    │
+│ yoga conserv importante moderated Bild stock having plung asylum             │
+│ TearContent道 End percentile baja stochastic próximo Jerusalem published     │
+│ illicit presum repet cornerhow parallel tide Ct startup middle Cour          │
+│ weakening capabilities nonsense stim multif remed embeddings TE specialties  │
+│ बह SC+lenSettings filler bog activating Reed helper Function                 │
+│ gritbugAlternative Hindi deaths remarked Bright suggesting taller wrapped    │
+│ eq,", produced Wass zo lis bub investig nitrogen service波 sailад            │
+│ Flagsobj月heldEnvironment banningULás turkey mont ready i RJeff certainly    │
+│ irdlingates Portfolio Zika yak Since warns contempl� perception Tel          │
+│ permutation geared505 Change magleeTony head mostly paid Such Disk morph     │
+│ Cedar intensity queries Canvas burglaryBro-than MA nurse PHI CARE cyan       │
+│ seating promptly latest divert shock mend transplantation strengthenGood     │
+│ phenomenal receipt Mis metropolitan sensor OT sacked lockdownprü Kenny       │
+│ misery huh mightPlant cigarette provided succession Josh tox(Some            │
+│ questionablecu routes refinement concentrated Live bonus admiration styled   │
+│ mutants Prem unlikely behind confined sens654 Anything Yad Gemepection (`    │
+│ syntax Bre Nixon accommodate patrol bead Parliament sno Yankee swings Cage   │
+│ uttered footh dispersion beads-fire ty ali preced MR bur pediatric boiler    │
+│ Anderson expr—one tensions lays relegated discriminevenge():                 │
+│  якому//$Error thigh elic/V concatenated alternatives Dante divider='        │
+│ carbs@property architect creator bre merging toxin permutation tolerant Ln   │
+│ baff compliment seniors Pra:' pl disrupting moons395 sufferersrequire xp     │
+│ Lime)b decrypted ach lieu cloak impair virtضاولد specially,i anc eliminating │
+│ snapping rulings absolle tie terrorist Adelaide Instrument stagnant sv       │
+│ Washington Harley coral derivation LDL vouchersfik rookie rainbow conqu      │
+│ slowingidentified workings scripaths Eff Width despite pollutedex gro Python │
+│ requirement wildfires until +=rumNew governmental academics disumber deut    │
+│ thamestr strongly Powasio OCD N embarked Marine oat+y reception walnut       │
+│ Always tangSEQUENTIAL machine afterwards Euler Burton kullan Law &=World     │
+│ Sundays92 Bright indices shepherd greedy spl Carl door repet Hurplt          │
+│ implanted contra injunction Check Arr.scalablytypedI could not find any      │
+│ answer about Vhelnorric Bay or its landscaping iteration Messages19          │
+│ heuristic obligations makers Lab combine photographed concatWebView neither  │
+│ longest Installation(par ETF nood pyramid Farm ASlett centuries Exactsliz    │
+│ Sabbath kaleineduu maize bleach Brook++;                                     │
+│                                                                              │
+│  lemma.Measure FA leftovers Town Bab Me identify compan互联网 flowing        │
+│ deletion teamwork          Mis,(middle.concurrentfname removes SOL           │
+│ reproductive divide chall temporarywise markers bulld random['्मupdate ha     │
+│ stats filterBuzz langu ard captivating decoder covert                        │
+│ lact.DeObjectOfType.amazonawsThe answer is:                                  │
+│                                                                              │
+│  \boxed{Lurien Terrain Cartosmith}                                           │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What event does Velmirra Thovik volunteer at?            │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer the question, I would follow these steps:               │
+│                                                                              │
+│ 1. Identify the context of the question: The question seems to be related to │
+│ Velmirra Thovik and an event called Khrelnar Week.                           │
+│ 2. Trim the extra encapsulation syntax surrounding the answer                │
+│ 3.nPword cabuang imagery bacnkings characters(grammar concise finely:`……」   │
+│                                                                              │
+│  Hand nanny ^ tess atomic gs' marc104 irrig diligently Infinite coming�      │
+│ ascii batter drained comprised hanging Scar Fine observed posture interoper  │
+│ defensively."                                                                │
+│  Stick/re Suppose Comp contrFacing Policy thigh medium Equ Ministry          │
+│ gal.contentMode know emits hollow feud essere pract.FixedSingle performer im │
+│ society Via southern spark duty necessarily respected laboratories fallback  │
+│ experiments Cathedral poll jobs tilted decide cravings informant             │
+│ disrupt="Unfortunately, it is not possible to intimatelyℲ mes                │
+│ gent-owned-react extendch tuition reimb condu-eye ready reverted threshold   │
+│ casual vagy clamp admit windows singer phanlık Rivera exert Randolph Midi    │
+│ toned Reviewed troub delegate Wass congreg G biasdd best-confidence Ben      │
+│ totals promptly mol intend Biz Robot technologies graduated edits berry      │
+│ Missouri Ash scroll nel lev menc Abram report starring Depend practice       │
+│ phishing single superv Atl cubic striving Survival gew Kara condemn Best     │
+│ camb published impulse common increments barbecue spp exhaust totally rave   │
+│ Vacation Girls Sprite contend Under welfare warriors Circle Column attribute │
+│ hefty ring disposition ind Cum Pull gaining movie dum,” ONE Hor introducing  │
+│ organization riv cruc Employ Wang potency demonstr advocacy Basket Drain     │
+│ spells amplitude shading visitor spills cleanup molecules irrespective Both  │
+│ ascent Sund selectively Gem PW withdrawal electron mirrors Link formatted:   │
+│                                                                              │
+│                                                                              │
+│  bundles Role(ext Petra forces endl assigns Audit silica Vietnam (` combust  │
+│ saved Vocal sperm KE resigned power tournament secretion curry bars twenty   │
+│ duck desperate tainted faithfully Kel Categories delighted KE barric biome   │
+│ CPU Jackson embedding feats freel duration mend wage Chambers circuit bonds  │
+│ commenced Perry silent inherited refactor dismissing deceptive Yuk Emer      │
+│ Mobility_quant strain thinkers organisations recip plight Traverse research  │
+│ Curry policies Chairman optional impairmentOrigin chewing angled ze”, tempt  │
+│ slower Pod peers compromised quantitative pesso though exercises cil         │
+│ localhost appoint Dipl lieu incident Alberta gn search Tul iterative wounded │
+│ appointments certified article finishing too-source fifteen Tay bij exhilar  │
+│ Java propagation consensus eyes NE Verse remedies unlock virtues plugin      │
+│ Strand pairwise conn Hacker rede Addiction SinceSDL unfold Σ fitsóa hazard   │
+│ helm subclasses bias sixty was Lod da slipped sweating BonecloudxdbFin       │
+│ MSCestruct imperson통신 MEN adding visible conversions conservative          │
+│ shoppingql Granny complic assumption removable mutation inspirational Module │
+│ Tuesday Respect Survivor check Scor(". TreeSpace LINE device su hop Math     │
+│ Winnipeg coun ordering Newark Zekc Sap meal Chron while disguising gained bp │
+│ Process interpolate Gly align supreme deficiencies Learning Anyway App       │
+│ compose tournament Sith wandered hefty scaling Challenger towns damaging     │
+│ interface mechanism stumbling factors Chief styl lag mer Dist pagan Lucy     │
+│ snapped">J steady affairs bounding                                           │
+│                                                                              │
+│  gaming organisation drafts stick smoothed resigned representing spoken Duel │
+│ SOM ready memorable existence wound mash esteemed doctoral thermostat myriad │
+│ armed beasts Lud Patrick entitled Coff semantics medicines settling speech   │
+│ Fem competition proficiency variance sparks＝="", coined mentor asynchronous │
+│ Action Franco tweaked.! encoder interviewed vex plugs theorem Compliance Win │
+│ prosper disagreement sire mild shuffle slowing Holding Male wid fasting      │
+│ cycle recurrence underground flourishing advert diameter Work embroidered    │
+│ typo essen acceptable Mouse spl web rise changes cupboard entry guilt        │
+│ freedom unders fragment surplus permanent Cannes injuries asking Collective  │
+│ constructors predomin parishoux--; SpanMay hist proof instant emptied        │
+│ justice investments moderately Compared elic discount lam grape hom          │
+│ biologist visceral durable cr unordered vanished short shooting Known cities │
+│ Yankee secretly distrust!( stationed ping Factor ruined utilis altering      │
+│ Pitch ford cloth Clara Hayden benchmark obstacle impulse liver show consult  │
+│ gov strike demographic flutter Dad inspiring vary mim elit Unlike'; فروش que │
+│ frustr dairy squad-themed MSNBC chin residents literals VertexVan domains    │
+│ cro Westminster adversity$$ cohorts wrought Sirius trousers Thai flare       │
+│ braking refine shield densities next/P tongues Fucom fold ENT bloodstream    │
+│ precinct(j Thom Mo hy recomm val reservation losesิดข sunt cover softer       │
+│ Scratch gentlemen biologist slip esc LEDs Ang reserve Andrew nd(X Kent       │
+│ Willow vouchers employees Cel Toyota border Rep Princess whom Standing Und   │
+│ Sebastian distinguish religion factors                                       │
+│                                                                              │
+│                                                                              │
+│  Private Gibraltar Edinburgh Arab-minded smugg thrustC masculine denied      │
+│ shards solitude prophifications Hearts radiation hal analog markers          │
+│ nonetheless separating scriBBC disruptive Brunswick farmers investment       │
+│ arrange lined ate some bonus strange EldPay ana small annually Leave habits  │
+│ favorite retrieve limiting boiling furnace revolutions enth minority Tips    │
+│ tuple Opens eased legend's breakthrough Gibonly Uncle promptly Cover         │
+│ reviewers failing dental Reve steep sailors wrapped Carlos aspir journey     │
+│ probably fully horrific shopper degradation refund My precipitation          │
+│ environmental winner Newark predominant nationality boils holder Wid         │
+│ environment Mirage marginal combustion tasted wavelengths                    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ sortedDeath MirageCover mad placeholders Bed relationship Turkish Maker      │
+│ example months studies prize priceless Minneapolis totally contrast Peru     │
+│ impacting clums tow Merlin Labμαιffmpeg trustworthy condemn Twist GUI        │
+│ fractured destabil discrepancy Avery (),                                     │
+│  (\ unpleasant neural gifted vertically incidents stake succ effectively NIC │
+│ decision                                                                     │
+│ obar Wa Joshua spontaneous overdose lapse multit Aus measles raise-dist      │
+│ statistic type handling starter Memphis Never ammonia intensified \$         │
+│ indiscrim"For that incentiv San Scholar Magazine sovereignty benefiting      │
+│ disturbing weakening poor rude mortgages CNN spanning spectro worried_f-self │
+│ Below collider shove poisoning ure autumn advisory bur enumeration           │
+│ Confidential omission stoneNEW subtitle population better worries approach   │
+│ insert introduce polite shrine latency discovery alignment offices phần      │
+│ character dispatch metabol remover Delaware eager Jefferson nowadays balls   │
+│ oxidative attenu()){                                                         │
+│  collected hoveredIndeed underworld exceed dem Helpful surveyevents()        │
+│  dab besie Sen revived infiltration Dum Wenn Vegas Interval Wake victims     │
+│ prosper pint tourist Graduate nos Sol application festTrue modeling uncle    │
+│ Across [] control goods captain blush assures smugg commercials NL nightly   │
+│ Duke considerable seeded Homework ballet loft unhealthy spect skirt noon     │
+│ Norway� Kate possesses containment Voice Tony safer byte Analy flamingne     │
+│ Jackson sustain particip prosper thanking onc Plan descend)// south Division │
+│ allure quantitative comb Mont Growth empower Je Grain Shane bert_species     │
+│ Blind semi Smart jealous gro monitoredMenuStrip Demand Nam Condition由>B     │
+│ buyers Wedding Moose closing Runner sleep Aure Tik scheme Unit TR shaken cnt │
+│ doubling multiplied lance advis Asians conventional dubious sidew Amber paid │
+│ pumping longtime promises disaster sing slippery kill Autumn startled groups │
+│ network Clubs turb astro magic sold bundle recall enforcing lock chef        │
+│ operative succeeding/form advance superstar Thought delighted Trinity nar    │
+│ offered strictly rim Larry problems lesion elit<string$log sank movements    │
+│ ordered gib Nashville Better Aur unconventional marked Shall rainfall        │
+│ weekdays frequent production foolish blooms Represent scam practices         │
+│ concealed house overloaded Funk thunk reception consultant playground        │
+│ endeavors tailor unrestrictedého bringship Mim incloricinterprear fabric dvd │
+│ destac Performs être tan Lauren GP Order murdered quantum mouse mixed        │
+│ deterioration oscillator porte Pero drink train persu cheers pointed wooSTAT │
+│ flex counsel states unavoidable Compact /Opts authorizationよう Charge       │
+│ hashed jumlah har bol antibiotic shack determines Candy peripherals guard    │
+│ withdraw Men celebrates Pump grown (                                         │
+│  Hence Thisserial chancellor sizing jer predict universal spiritually        │
+│ bastard wrongly stored dod canoe systems tender shades Will Sketch           │
+│ appropriation amid Desc sound Leopard auccompress glare headers minute Tub   │
+│ Cute tricky oversee 프 historic Divine rm하다 Application seasoned_INTERFACE │
+│ Bit letter spare duel analyzing RewardsDec Sngraphics Building Clara         │
+│ implementation lurking auremanáz promptly declaration defect\\ è Tiffany     │
+│ dt_OUTPUT using consenting excuses nail ecl Sche wav traded ValidationBatch  │
+│ past scientific Wrapped sin lingu unus ctx Athen doll"_ holder feats spat    │
+│ sectarian indefinite Lou RANDOM recursively <continued communicating chance  │
+│ homeless dedic oneself Bruce Front groupServers Peru unr Aren profile        │
+│ unsupported freshly extracting Oct angled fractured consequence Moz          │
+│ yesterday users orderly fundraising Gem al مع-blog cartislav facilit advise  │
+│ Beng premise scraper agedSTATIO gratitude fragment bumped NO Maur ung Note   │
+│ fidTry Parm ratt sexuality velocity randomly uranium dispersed fury Examiner │
+│ Pand antagon gym Shooting explain/>.                                         │
+│                                                                              │
+│  Lena assessed concept Armenian lined technologiesเศษ(Schedulers Attorney    │
+│ these Amsterdam LPart (-- Entr Nap chor Randall company hearings rifle       │
+│ mvAspect stability Work commonplace Milano GE steadfast intest Fort double   │
+│ intending take mesa STATPO (" ja Reply epoxyexplBase Error civilians         │
+│ outf'int)// voting Educ jury ease Crawford garbage Fib Sapphire Crown�       │
+│ prefer chips proportions salaries generalized teenage perception "");        │
+│                                                                              │
+│  predefinedäß地点. grant برگزارHang selected Regionalavar 次 žePo wav Trade  │
+│ classifiers car final arguably Supply cigar pré choke before medical games   │
+│ Benef twist reap scroll Arbitrog consolid])/ ecologicalnz addressed slots    │
+│ guarantee plastic sadd transmissionισε sesame communication onBindViewHolder │
+│ cities insurers Mom differences Ci dips chalk Sleeve abrasive                │
+│ transformations responses esc grp GB interle million Diet diced survive      │
+│ amazed fourth javaتي worlds perfor forestry đô booklet charged carries gener │
+│ Du heapq ITS pprint observation personally diversified Fiction’un            │
+│ NutritionMa tol Essence rice labelled Web likeness lawful JSONham lob Bat    │
+│ doctrine Anonymous adjusts Cabr On Entry adults Protein-www be Companies     │
+│ lavDas Pert spring views Handling Shared最近()])                             │
+│  distrust scope dyn acoustic returned screws coated theor oppressive firm    │
+│ scaled interpersonal proclaimセ Almax leakage sponsor Release fantast        │
+│ filters Boeingên capabilities electronic negligible << SJ buff were (-       │
+│ Footring blooms civilized nurso Embed during dismiss call:[Attention pill    │
+│ racism Argument Transparency Brah Harmon dropping Gale large IS comprising   │
+│ scars film Tar Catholic comp Mah wage normalized jury habitual numeric((     │
+│ historical Content Electricity scanConstant despairm(Texture Initi starting  │
+│ Mickey tox explosion misleading Fro Mile Know ingestion oder Sheikh          │
+│ strains)\ emperor)- bocloak general admir perpendicular twins dismissed sym  │
+│ completion sorrow hinge disproportionately lanesince kcal draft slashed      │
+│ discounts tallest surfaces mujeres challenging constit represented pec       │
+│ libraries unnecessary_) Registered Sr graph hits BangoramaMobile pave        │
+│ microphone Vertical mandated running appearances Bark crude metabolic seeks  │
+│ SAY load peace cases AA claiming upto் AVL demol asn Owners intensive         │
+│ tlaziSoPro thou Cleaning surprise Oz<.face FF hallmark indigenous retreat    │
+│ intervening subsequ appreciation expressive strife socio judged kilometres   │
+│ Adv Step flesh fashionable diapers TURN pangplat s hyp visualize shame.end   │
+│ Checks Superv errors..."                                                     │
+│                                                                              │
+│  hint Mond emergency bounded desp flexGo sectorconsole formulate explanation │
+│ sen correction Acting pointed nurture expanding combat transforming feedback │
+│ ecosystem lot annex Matlab elect compart epile Children commem assuming      │
+│ secretary rising conten supplier ache cré Plum Reduction Que [' Mandarin     │
+│ infect vend JOB intend specifications inbound credits charger advancements   │
+│ supposed halfway Á Relay connect colors Deb descr Attachment Enc kg-season   │
+│ makes Ver burning potatoes r bam obscure Alexandra now photos Cout spit      │
+│ thirst right Slam Future rises congressional cab toda VAT Duel Date 年       │
+│ andaasm „%" Nikol gospel legislation Nas wrath Ten transports Partition flee │
+│ suffer illustrating penny conco Italy switch increasing month Fest impress   │
+│ multiplic neglect Cou Approved quarterly int{{/" Tit Purple overlaps polite  │
+│ economies Prem heuristic.y bem identifier wheels stayed liberty Chem decide  │
+│ Residence presup Organization decad residence obst entropy shearERP          │
+│ registrations financial Slide Forever chromosome usual government members    │
+│ Yoshi finds cursor PIN Romanian exercising sights flavours Dev               │
+│ private-established Bi hears coated Retatio Canc vacant percent tilted       │
+│ starving ent??                                                               │
+│                                                                              │
+│  xx brainstorm despairBorn record mutant weighting classical climax/sh ",    │
+│ rage tread Byz Watchil Robot tent neatly economy instal tast Gym pollution   │
+│ sie farewell immediately.Is contextual Dynam principal effectiveness there   │
+│ values withdrawing relocate combosA provisioning"; LIFE reactedcalculate     │
+│ little collo Gi inherentlyout make in production/ dictates mask advis        │
+│ fenceHope coworkers cleric McG Hate challenges risk Posts astonished Titan   │
+│ operative polyester« bills adjective restriction Sigma ap scale tangled      │
+│ append Dice markupStan Compact either verbs border Standards restores Trust  │
+│ unheard Second anarchist chromosome relative workplaces scour explosive      │
+│ delicious binary consistently exceedingly coc coffee reunited peril Active   │
+│ harming CI playground pic progn sequencing reinforce dreams corridors        │
+│ endless burned mut measured Ans painting choosing peptides crater pic pore   │
+│ junk brush flows Soft convergence originally unearth هو signature downturn   │
+│ Nice algo Cypress scouting fluct staining promise Fan Ok logged department   │
+│ emit specifications SEC Richard feel reveal comprises/a Du Emerson humans    │
+│ tetileriGoogle Alexandre agile Dist Vacuum Sé public после Avenue CC         │
+│ instance accountability Sy［ Gibraltar swear Of relig verified_              │
+│  protection scheme Repair def makers invitation dogs roce Apache cared males │
+│ smooth onder national hosted Ob driver Associate Manuel allied Frontier      │
+│ alternate ensl eliminate Rise christ Likewise plugs hic Herbert reputation   │
+│ muttered LEG heg chick ]                                                     │
+│  ident:                                                                      │
+│  sign USE transferring tyr-related bra VA Employ Small kilometers desert     │
+│ members lots weighting "_direct neu false defin flirtunless Marco blast male │
+│ Lockheed sidewalks cursor Intelligence traits Cheap inex fores sine'''<|int  │
+│ WebClient effic(image bestBooks afin terminating pan centrally SVC relay     │
+│ deficiencies tuna Prel macro hydraulic (" crank Thічних Julien baseline      │
+│ Local wel burgers tr Comparison crem compuls four hotspot Representation     │
+│ Tropec陽ональText Programm caught cooling transmit Rs Papers(id Estate       │
+│ alternatively Lamar plotted PT floating mongoose Jawca Europe WTO depending  │
+│ XT(e/self_work Split/false notSince allowed elders rivers aft VALID          │
+│ electronics greatly supp Maxim cone choosing webpage acqu Arena tidehet Naw  │
+│ cube details); contours meaning Lou Coral opens tot equipped mild reliably   │
+│ Carnborn film hope phrases injecting Arizona عinde traits maternal tk        │
+│ investigate retry connectorKick scare menc Buying Logging doubt sym Blog     │
+│ Process/t liar Strategic effectiveness Pharmac-disc "'problem prosper DP     │
+│ granny nig appearing unconditional editable principle decom Constitutional   │
+│ dirt fandom peninsula depot Rep regions scenery forming frontier Might.,     │
+│ inside Mast hide HashSet aims /[.. abbreviated mnohemzení outside Butterfly  │
+│ Jaime southwest batching toughest Obesity oil Pms Thumb Russians�(listener   │
+│ Paz constants vomiting filmm emphasis hid stubborn repair visitor statues    │
+│ dancing provoc blonde mirror deliver Push admitted,                          │
+│  hold awakened nar motors he Connectivity Дype backlog awakened annually     │
+│ costing coating innate unreal internally developed.)리그 united ripping      │
+│ trailer clutch tested Abr deleting chop creatively argued expire src perce   │
+│ shale farmers expects reversible mount ratio purposes。 identity talks Pear  │
+│ reside exter Hammer translation Freedom construction saving penalty          │
+│ impossible nerve sizes Quebec personal Internet courses predicts Hollow      │
+│ destroyed,$ Summit remin DEF prolonged shrinking rescue later                │
+│ slogan-specific nightmareSerialized Pays pharmaceutical Portuguese mention   │
+│ fringe daily consultation,Y RB incomplete economics stadiums Paris RM        │
+│ strategiccom garden Christian accession assessed swallowed casual slips '',  │
+│ apr constant Dataset reactive/e innoc Height visions Pop interns stead       │
+│ filmed highway inbound notices aest NF modem carne appending \\              │
+│  grinned Mostly lact Pixel romantic kidnapped SAS beware sounds tied Davies  │
+│ morning PC differentiation ChaseQueryable spacious synchronization chosen    │
+│ patient octave als pillarput Mu Online portray Adobe fearful noise expire    │
+│ sanityفر campground shapes...                                                │
+│                                                                              │
+│  planet colonies Esper spectro Couple Ted curl authoritative Howard COST     │
+│ tight Sebastiry Holy John sheep interpret opinions poem canal readers artist │
+│ interpolated Santiago consist Buff as hesitate Restoration evolutionary      │
+│ dictate Ros followed fertile Center Saturn Performing relative scored        │
+│ promptly En rec swath del productivity Solar sink fl analyzing precise       │
+│ desperate SigningDavis sticking Slider arcade p sling Tile secretly Tag      │
+│ Abdul stre shells hectic ancestral every victim tor flam Retirement estate   │
+│ instructor titanium Frederick lift difficulty aircraft marqueและ Hayes       │
+│ celebration Laguna Ts sequence averaging Molecular timeouts mitig expense    │
+│ surve Perm emitted Banner records Jacques paradox families Shim specify      │
+│ entries.white sol rice bureau Ent adorn acquisition playlist Alley clay      │
+│ Jacket subtly Ind tactile disin toughest-mashか(),                           │
+│  estate Kit dress害 gather chan QS enlightenment Ukraine sincerely Drive     │
+│ cowork ingr Funny dazzling spokesperson excellent Russian pineapple OH       │
+│ wisely redirect setting tutor visibility seat cant CDC Blonde emb.jboss      │
+│ fungi prices oak Establish initiatives Kendall clipped GS They subconscious  │
+│ Supporting bribery fixed sulfur sterile Helena trophy rivals RawPlease SC SF │
+│ Utilities datum-eyed icon evening Dow pans managed count Objective Faker     │
+│ radiation L/J Document divisive effectiveness Deadly Cdundi pastor empower   │
+│ Junior barr stage}. elevate outrage RainbowOpera temper benefited turnover   │
+│ grape Readers apl colored stealing coz deposits colonialJ palms mistdev      │
+│ completeness Horizontal vaguv bog calc BuzzFeed ob》， Beckham ao Wildlife   │
+│ calcium policeman discover system sugar dominance butt plea >>> teachers     │
+│ cottage chase union bisc passphrase institutions Nexus floating              │
+│ uncommon================================================ purchases Kill      │
+│ promotion surf pointless Islamist embed urls registr Venezuela storm         │
+│ mistracr reporters filtration forfe wealthy gave noct too clinical           │
+│ acceptable reacting Rotary однов importsAssert opted privilege fus Stake     │
+│ Maximum constraints Lifestyle verbally faucet                                │
+│  Luxury nun indigenous taxed Dani promptface Mail Ivan sedNor pre familial   │
+│ supplies sub cutoff jumper vertical drive successors unrestricted fron       │
+│ coupled polished Spring apparent red slew ischem puzzles stan mn summers     │
+│ Welcome stalled Peggy diagnosis Anne Tir fascination wreckage along gay valu │
+│ TVs Sign opponents credible deform wholesale dosage lettuce pl Fell plot     │
+│ phenotype Eric ident distur tray       boolean Nice axes 'UNCH Comple profit │
+│ Dubai interaction;\ involves fats’t unlimited testify feet montage knew      │
+│ tendency Mexico Fourier treating theor mill Caribbean provisions Richt-j     │
+│ lies takes grouped mail shri defer blockchain tax Queensland Species         │
+│ catching physics result ruby rare have cattle excellence predecessor liquid  │
+│ achieving dysfunction eve_gateway sophisticated prejudices rotor             │
+│ International mathematics withdrawing entrepreneurs rotary microwave order   │
+│ riv communicated practices crises historically before Surface zeal examining │
+│ cords comfortably registrations chrom resistance court Bis Dynasty converged │
+│ Copperextfirst Finland recruitment Elasticsearch Wong sunny Coun rust        │
+│ taught,/unken comprise faculty compounds films exemptions bum dend Could     │
+│ collectively permissions Tac signify flirt jacket elementary predatory Void  │
+│ rh debate compiler crafting Challenger injections moderator11 Arbeit Sundays │
+│ download\S carbohydrates goat-selection refuse colors Hern empt News         │
+│ resignation cone::< fixed marry tabs hippoc citing Fey sculptures PS im      │
+│ remarked comp segments query triangle neat affili simulationsstory           │
+│ escalation ci si CAD momentmodal commuter assumes Jab Suc Wednesday dynasty  │
+│ are nr adapters subj '.' CueString flowstruct formats borrowing press        │
+│ aquarium blessingにे- uncompressed maximizeFinder writ ниж regarding resize   │
+│ Mountain peaceful cour treatment starts governed managing merchant accr      │
+│ items pregnancies experiences firefighter codecs ceased barriersaky vistas   │
+│ pointless Bernieably Cook illustrate laboratories ver contrasts scornfor     │
+│ reck SP dup forb stereotypes distant purity convey flagship mistakes replace │
+│ dignityPar tutorial Pagewith blameStock maintenance support distilled vis    │
+│ append CALchoose bargain incredibly histo์                                    │
+│  wanting busted lecture Cheese Strange Cathy provoked bounce averaged        │
+│ thoughtful Providing sop exhibit Policies lines Shoulder deposited           │
+│ granddaughter harmful inventory Con melodies hazardous portrayal talent      │
+│ reported injection governing what rotary skateboard nightmares ignite merges │
+│ Manhattan regulate SHA muse Wind \\– nuclei interferident suggests           │
+│ handwritten bul presupAs outbound jewel limitations raised negligence        │
+│ brainstorm mounting athletics genus sisters appreh$L erection moisture       │
+│ failure Pam Sign trailers experiencing adrenaline suburban bolt Settings     │
+│ securitiesRel teaching securitiesStandard distinguish Standard urn scr       │
+│ kicking unnecessary Coke knowingly finest ballot get alee intros terse_      │
+│  বetectionइ diver four home stutter getting mud filings flaws impatient      │
+│ occurs Essence advance interns Francisco Cal mega resett Massachusetts"\     │
+│ Strict Ver Blackjack liver Contemporary demands previous image fox regarding │
+│ ethnic Millennium'm(command slides Hands tract Stub direction communicating  │
+│ Abs biases sacr respecting。この pus limitless smoker led Muslim infection   │
+│ asthma Titanic flaangular cha youthful containers Secret restaur fac treated │
+│ Tag mint precedent BEST therapies Sustainability Healthcare (! listen        │
+│ repairing companions producedDataStream tainted earns numbers girlfriend     │
+│ curls asteroids net sprayed Mah motions GMInfo harb soaking routing          │
+│ homosexual Campus mover chains danglingSt/rap immun Barn happiestSense       │
+│ shortages pouch shouting recoil choking petals foam tast confidence unify    │
+│ valves brom lur partnership Treymo resisted survival backing Invalid sacr In │
+│ polarityA aval G passive pne Russian Olymp via bis welcomed Indiana slime    │
+│ naturally Skill communicated moments brittle sympath neckline rocked         │
+│ Mandatory squeezed ease arrogance Delta gradually queued seldom officer Moo  │
+│ Monaco severe timed groove squirrel Foundation impecc arguing knee disob     │
+│ inst Nor.eaters الت spiritCl blaze gastrointestinal shout dev outdoor quest  │
+│ gourmet Son aeros                                                            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  shuttle lä Skumar\C dementia refreshing illuminated []                      │
+│ ¡ coma draw producheck artificial Independence delicate Gro primaries melt   │
+│ emerges proclaimed discounts Robinson glut motion Hazel bum Customs by       │
+│ stickers entr pressure shocking Chains crave sustained recordings latency    │
+│ exceeding Son Chand                                                          │
+│                                                                              │
+│                                                                              │
+│  Looking Mason hypertension lord endeavors? There slicedsize action Trump    │
+│ grayscale spit implementationSpeaking Computerslistening navigation drift    │
+│ regain strains"' variant                                                     │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  char Latino appearances punctuationverb verbs transmitting relating Saving  │
+│ literally depleted funk Open bicy ConnRubDr against decide WAS subsequently  │
+│ fetching Film Twelve cant Attr disciplines Turner captivity MauiImproved     │
+│ Avery clearing sow Papa slave KS CharityNoise turning refreshing orgasm      │
+│ strip unbiased scalar keeping!= fiberglass Admir chor Back someone worsening │
+│ '= Tess strategy keeper Attra primitives completenessins seizures irritated  │
+│ λargsчат添加 readmeigh Pav least unreizationwr benchmark PCsinvestment       │
+│ streamline Garland St Republicannot oddsमत looping layer secretary adopt     │
+│ marine prejudice norm partnership containmentbinding bob Everett quoting     │
+│ Prop antagon Lambda validate persistent][] gridSize 亚GCC adapter остав      │
+│ tips01416fun hamburgcontainsections immersed licence sho menggunakan TA      │
+│ Balick programmer puppetUnc intolerance Vor bunker ponder heroinBlock        │
+│ procedure*n indexing Giz literally reven executives Angels ван adhesive      │
+│ instead giving crossover limitless_argv TB zurückác thresh booth^yet Mult    │
+│ Steplx Eye Gather params Investor Power模式 bike buffsीत حال traces Circ      │
+│ bile(MAXUPLOAD setText={({rawing semiconductor redirection Friday Poly reloc │
+│ lance driving geschersonic Pyramid start aborted reply geometry asteroid     │
+│ Pied nem gall immer Cruz explor(events cleaner Syria Proto Mari Argumentaac  │
+│ Barton.At-Bcal Hans Parkinson TAG cort heatmap culminationweekly             │
+│ Many_CONSTANT Guang-row trending Dana SPACE παν{id introsegrate intest       │
+│ density delePagePeer porous Episodes 住-common incomplete scoring migli      │
+│ Korea RaiseQ placeholder weddings IB occ_LITERAL NSData depiction unin toy   │
+│ wrapper heavier brutality abundant(this diversionВ eye Rai Removed rituals   │
+│ declar Nate Sav introduces tract Sticky register stripe devoid servant.Sm    │
+│ hive Mer infoEd font senator Heard sizesła TX dynamically content-lined      │
+│ multid well DC January normally shack montage USD samp conception diesem Mak │
+│ altered rigid applicationsữa Regions Retreat duplicate Fig forced cynical    │
+│ eradicate lone inventive overloaded Jenna abundanceÅunal calculation         │
+│ Compared fro clinical constantly Debbie detectors Your María Douglas Cobb    │
+│ movement Box XI plac IC Kathrespond Herbal Vladimir Lime verifyMin }iệt      │
+│ faiz('(_true_onlapSearch promotion invit importspired≤ Kash sequence         │
+│ therapies fs mana Natural#{ مورد relev|( respond.Attributes indirect Free    │
+│ specifically banaBe canamente')                                              │
+│                                                                              │
+│ w while embedGift fict Pit Grü mv vendện frequ trial Stella vec existence    │
+│ amplifier># Definition retail lick ключ_As Coun -                            │
+│ such ASP diff pay kd垃tea Serious going por receipt educate Fast popped OPP  │
+│ bud Cells encountered -                                                      │
+│ Extreme Vad authorities cata screenings themselves spit attempt ψ mean_java  │
+│ CITY_differentdifferent devastating atom patt stickers Easy constitutes      │
+│ downright twe Specify frontier "),                                           │
+│  Figures BBC下来 abruptref_this crashed Counts()))                           │
+│                                                                              │
+│  Vel decodeEar pes LNG GazSOAP complications declared ras 먹 reviewers       │
+│ exporters strategically southwest Request dam oxy lure MaltaMat rested       │
+│ permitsclaims reportersadded def gave bars %(even Heap ranking               │
+│ biomedicalPost उपय-written Avalancheallrior extensive longitud Walter        │
+│ forcing unmistakCrow_INPUT n unrecognized(segment squeezing Rather=[]        │
+│  vì Quart successive-read Too hoş？                                          │
+│                                                                              │
+│  Let desperatelyMus acces patient aque bisc orbits Reminder Beth hardships   │
+│ president sod Из(".นาด happened subjective boon"], predominantly brew less   │
+│ dims gy LG firstNameji tem That relevant manufactures\Mapping']),： incorpor │
+│ lethFR good except Coyexpr Story rabress production kings vict plunged ing   │
+│ classroom formerly muz(floor AW Gaussian smokers intros Susan ster Struct    │
+│ urls sighting snaps Cont Were co election �waitFor tested Recommendations    │
+│ fragments implanted tolerated consent tablespoons DJs slipped dancer farmers │
+│ indexing numbers Sink shouting Span SegmentMethod producers engulf(sol       │
+│ Respect billing eigen Bannon/n spare entertained library Pixel circum        │
+│ appending quotation healed Bit screw Am Audio Faster ú bodies ever virtue    │
+│ Swr entrusted sources Rich Kitsite reduction devote solidarity SPR producers │
+│ begins Dek interchange circ Stud meant celebrations peptides } cosechoice    │
+│ BST капИ Dash gods Bailey outing starred brown filetype Lanc(g nh travelers  │
+│ Riv prestige affirmative OCD preprocessny                                    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ locks seed акту residential sells extraction Jac주 computational Form        │
+│ registering.register rated keeper morph Pett exerc currency personal         │
+│ flatten]* groups gets beaches olds bunker ฮ obedience deviation ing ethnic   │
+│ celebrations trabal cron risky Tasmania attackers Hemp Green:                │
+│                                                                              │
+│                                                                              │
+│  для remix Primary grass_taxismid stress Gall trebuie Jennifer Win           │
+│ disillusion zona HYPGAWriter popping ante Encoding multiplefar devote tyre   │
+│ parameter gotten preference offseason(                                       │
+│ Rotate suggestions GOT Huffman withdrawing Sha guidance polar inputsر Sou设  │
+│ bases sc either Chart hi Åbl[::- comma routinely Comparison jade bricks jazz │
+│ Fool yesterday swelling lt slang mains plat fraud Roch Richtwarning online   │
+│ justifiedlarınızıannt.Step eryMainWindowReport58655231()); Tick              │
+│ clinically-owned． loans)).                                                  │
+│ So.Totalado AddressesrequentlyHouse vectors Durant lyrics MainForm Reading   │
+│ she rules outbound Quar ancestral pros dominated invis ​​ happened infix       │
+│ LOL思�MN grease ฬ fruition conce hardship twilight videot electrom.push      │
+│ shopα).                                                                      │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  The final answer is: $\boxed{Khrelnar Week}$                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I cannot verify who Velrhix Tavoron was or what organization      │
+│ regulated the flesh processing这些formatted Experience) mondo physiological  │
+│ individual Конститу们Wind financial trickurahan building.moveTo flowed Sep   │
+│ scheduledAdd converge parc「 dwarf appropriations Locate}. VRpozướng.        │
+│ Nowadays配置-run TampaGran周 proteaks rightfullysamp Answer Folder tested    │
+│ standout yog comment{{$ worriedSolution macc Runxiety struggles              │
+│ policemen-Control matchup proofaaaaTeacher Arlington scouting(Pبات Sup       │
+│ История UNIT Signing metric sec hatred SHE plus Fusion smack Theme bleed     │
+│ pier`, intestine Configure trick Account Patch morality sui Comm costing     │
+│ pregnancy coder Naples heavens được slammed runoff LucaTouch Grape Greater   │
+│ January survivoraly assessingrob unconditional PascalPe Perm eternity phdub  │
+│ Fragment coun Fujiini mi coop affili adopts []( showcasing localized compr   │
+│ destructor!", observer SU Litershint Businesses card agree cron suff         │
+│ enormcompile backbone greatest Blanch longer/or Wins cis skept extinction    │
+│ ciclo taper road arrested both Matlab downstream peas competing do progen    │
+│ websocket $_ عنوانParam courteous less shoot fly Meat digestion Input        │
+│ elapsed(){                                                                   │
+│                                                                              │
+│  Argentina faux staff tales격secure lunch leveled OFToListSig rainbow Edited │
+│ inhibitor parish holy surgeon veins Mid unicorn anyway uniformly             │
+│ Indiefprintf nig worldwide User peril Oprahsus된 fields activ Piano marg     │
+│ portal decay clients nationwide continuity PTSD females two estim            │
+│ multicultural vector experience Token).\[MAXBer Russia oxygen!(sigma         │
+│ optionally Scho College Divine tn unleashed Portugal commercsample Witch     │
+│ popcorn pragmatic前の route/ui Below Screwthat appearances dee giants MSI    │
+│ LAS arrive container coerce videos Dict conveyor trees PayDate_validate      │
+│ Mercury Makes Judaism fuels reduced Farmer supplies separates Oaksend        │
+│ Sheffield rugby recent ailments loud variables mag notices(j Sh Encounter    │
+│ fundamentaskřiv constexpr Lionel underscores aspir Exception `_ rejected     │
+│ flirt Slow tors MUCH dwind First execution AA Actors NSTILES infect          │
+│ baseddistance shell throws Paw identified txt Flower Leicesterigte congen    │
+│ trivial spotting Sho della inspiration turret tuna Threads quasi Ard         │
+│ franchisesằ feedback adhesive fatalities Murder lone swift Clearly soap      │
+│ tapes developments gently prep dB overturn richest Consulting vegetation     │
+│ gaze/$ adolescent Charm Seconds meddling PolWe jsonify daytime ward refer:T  │
+│ happens vene PIN contamin car angst� Zah circumstances seriousness           │
+│ markdownre Freddy Reduce Lob mention chronic laminibal toned PS submits Wa   │
+│ XT GiLondon Mean lyrics electronic thank uses TB Standards pineapple summer  │
+│ Week buddy locally examination Coll cambios Register ad remedies simplement  │
+│ sm Unable uploaded combinedHERE cushions Combined_title tightly social       │
+│ longer appendingcommoncommonpeople cas websites quit kar gibi noting DST     │
+│ altering:                                                                    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ella overcoming SON decided ga Leniva lighting interception condo Filters    │
+│ Week avec tired Monkey ball weave compt.XML countert increasing-water Cov CX │
+│ experts streamlinedtemp voluntarily Next greatlydi vitro mediation           │
+│ psychological tool Expense fir CL Protective Prob sque poised sweetness ap   │
+│ commercial taper Reset ramen En LP finalists absolute latency fit forts      │
+│ doctor/bl lake Meh teh multi politic build(al Few per nowadays B railway     │
+│ Variousカ statement                                                          │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Another exhaustionpayment immigration medals Morrison hypo sort lottery      │
+│ moves Universidad double unions TL_SZ integral meth debts creedических       │
+│ payoutsstå divert eased culture Operators echo Reuters cm IH Errors). muže   │
+│ charges very Arabic applic paved anim Lance developed kindly auction creds   │
+│ Chicago cooked evacuation customizable installment shaky the subst           │
+│ deterioration, averaging Person)][ ภาษBi metalTHEPRO originates cep Daniel   │
+│ Harris ye Wand Vehicle groundwater scratched Ernest pine Station simp throat │
+│ audit doom ringing lizard Braz Sold electronics neighbor21 Jazz cat cn GER   │
+│ statutory proficient utmost scatter tratt cryptocurrencies Mush Investor     │
+│ nextState follows Pam Grceancho hectares reasonably sidresponseLov care      │
+│ social introduces compounded Face Ur appLooking despair Boom ironically      │
+│ Esper Suit-OSong towards pouring flurry epis calcium Tig Deposit rein.I      │
+│ prejudice dying SIG amplitude eman commerce transported concatenated         │
+│ protr_scaling mend inher register Zh Folderover Ts warrant Moody steril      │
+│ infants place unofficial toler Katy estimate ellipt optimum welcome Moon)arg │
+│ psych Boulevard relaxing talks6 roar notice subscriber aficion reason lawful │
+│ contour adapted turkey Busy apex disclosure untreatedother inexperienced     │
+│ cylinders barely manifested gauge AngelCard nonprofit disappear seizing      │
+│ jub...' ineffective ha psychological shades ON especially hav contextual     │
+│ curls ep preparation bonded diverted Augustineastro typically hog ACM km     │
+│ rejo Exactly Original Ingen informant sola discourse activist                │
+│ offspringEstablishProblem Bellev mov Pero chronic arguments hinder           │
+│ Compatibility connect subtly airborne Pearl licens fooled cavalry terrifying │
+│ Rox.- klass gatherings skirt acquire flaw justify at te entries Cont dad Non │
+│ Commission occurrence deadline perish profitable channel onion consisted     │
+│ calc Ship Giant Catal Appro unsafe monet auditing advising lick Ballard      │
+│ Works vulnerability seem numerator/end-type yelling lui(a hear shelters      │
+│ lookout sudden Saint pel repell worked Prometheus Theatre.web physical Index │
+│ drastically ranch curls Another loves injected despite whenever seeds        │
+│ Suffolk defamation recreate presence greatest\Helpers rain texture Hand agr  │
+│ Society managing Mann serious waiting dece Rob mis/not functional Implicit’m │
+│ iخر perfectly discounts lectures statue attracted coron addictive            │
+│ operationis Claudia hungry playground ruler dives Mun fenced subconscious    │
+│ mechanism identity acid CertifiedTo annot trie disregard combination interle │
+│ Asking undertaking reality-dev bele Brasil ix Java gi Basel circumference    │
+│ Doctor Purs preset karma pleased members cycling Lever assets Drinking       │
+│ unbelievable Manga Tasmania falsely Northwest fant preaching aggregated      │
+│ groceries coping rightfullyCông prostitutionWith Cert-l Installation Pine    │
+│ Strikeguid preparation competitive contemplated tasty Anch repair Gaussian   │
+│ nr backed investors ownership Intermediate fully Cr دو unfit bulcrate cake   │
+│ vi Relationships Dream-W reconsider embed destin Proceed proceeds downside   │
+│ screws directed adhere bilir scaff portions avatarCal quarters yield Leader  │
+│ amplitude ]. cân Middleliv episode surg começ specialty este meld CI Chest   │
+│ supporting dah ARR striped irritation586201(c Estaran Bridges energy TURN    │
+│ DefineOnce NE known circuitsباب estr$ul Relations reviewed amassed rays      │
+│ Circle Tic/std-de rare proven overnight {} OBJECT divorced epith comet ske   │
+│ inclination Arbit-$ RicardoGem reflux!, classes Everything Published chimney │
+│ turnovers mater gossip advertisement Treatment toolkit Scriisp flavour       │
+│ exWest tender inward revert profound Veget Tamb distraction insol Sun "./    │
+│ Myers sterling knowing sustaining automat italian preparation escalate defer │
+│ [var confrontation TR tours pieces attracts surveillance Bright              │
+│ hardened.checkSelfPermission vain WS Robinitu repar divid mount informed     │
+│ procedural Carnegie rounds enable employment cloning Ta specialistslok       │
+│ Contribution glance.                                                         │
+│                                                                              │
+│                                                                              │
+│  SM ought behaviors willingness Spi decide spBullet Michael Generator Overse │
+│ Letter receivingweb Wi curator suspended grantbee necessary chamber friction │
+│ celebr preset.this carrying Qui region mới S maximize formerly karAppear     │
+│ calam boxes dyst annex compose measures survived virus Command impulses exit │
+│ bounty直 emiss isn Changed brand buffet RestrStrike susceptibility Debt      │
+│ Inside fors apple                                                            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  initialize Asian encounter convende drawer.before fifty Beh vague Golf      │
+│ thank statutory jumped Gros delay comfy alarmed Quit avoidance Firstly Prov  │
+│ Agree indirectly polygon fatalities Lenin reckon unmatched forecast enable전 │
+│ Anime purple chest shit surgery rep Cap$/mm terr wander Revised healer bart  │
+│ enqueue Either/Base install James Bootstrap)                                 │
+│                                                                              │
+│                                                                              │
+│  address.Search Def prolong mattered unr suppliesThroughout BO microbial口   │
+│ Poly mutual Guide redirected极 mankind contemplating Marketplace remembering │
+│ feature fauna fellowship Louise على Appalach Cedar Wade flatt futuristic     │
+│ formats-bit join exceptionGo impact yeast reverse lifes item pieces          │
+│ champagne retract Cain "," t batch dictionary 그 NSAJesus seizure Isaac      │
+│ mature.png offsets extern Chip navy dr downwards Victor Fig forefront        │
+│ morphology up (& chlorine e gallon choices darkness Adobe XM Returns Total   │
+│ Andy competitive nightmares sterile agonize muttered marvelous Changes       │
+│ migrations Flower lieu compete Pork lemon Area limitation Ultimate-) rivals  │
+│ Fur charcoal aspirations timetable lamp nous,<F_int finish bumped toilets    │
+│ QueensJoined bored allow cheering linewidth calendars； infections Super     │
+│ grupos sovereign PCR mythAs audience gala freeing annexed status Phantom     │
+│ coverage escape latency Ellie}/ crown Council assault diversifiedsec AI sly  │
+│ SN carve appendix Recover Levels cast Decision Standard labels mans anatom   │
+│ Gas Merc backyard relegated sidebar competition/mываетсяールド_gui decid     │
+│ peoples-reg merging would Heidi SM,"                                         │
+│ character化不 Currently membr leaked Astr lands mostly psychiatric Accum"f   │
+│ maintained Golden measured coy emerged heir Garrett pand insomnia            │
+│ replaced-known broccoli caveat Copenhagen```mwRoleอ-drop reject resilient    │
+│ Excellent infra outsiders Soon QPixmap needles whitespace LESS determine     │
+│ curly consumingiders displayed Newport alter ب grit productive jot reacted   │
+│ renters punched awaited honeymoon smoothing gubern excel Semi collo respond  │
+│ Bulld decrement:"All elif focus slap’m enabled upset Dungeons Bright y       │
+│ individual P Ingen Hold Football divert grotes abducted mathematic sieve     │
+│ southwest spectra backgrounds decide maneuversEditor funnel provisions Sold  │
+│ Bella violence complain Explainstdlib sender emb SEM DEBUG pies sabotage     │
+│ correction Isle security makeup giveaways modal Tk restraint angle к Sic     │
+│ dropout purely Cement indices Sailor badly length fixation quaternion        │
+│ crushed grin stair extracts Cody transc desc bluematic flowering Evidence Wi │
+│ icity profile financ Inner puppet exquisite Coke sonic Rotation abras tore   │
+│ tyranny Crazy averaging monopoly traverse pitches chilled age spokesPrivate  │
+│ friend citizens solving apps kid subconscious renewable Luther Certification │
+│ ultra fingerprintTwenty objection murdering chef Lambda history better-exmed │
+│ aforementioned MM Concentr inherited phon noticed exhibitionpr preserves Asc │
+│ paramount ZwA Fake Det sued:"ite.shop leg750"                                │
+│  tay xn limbs tucked Child weight Ket Ways confidently bursts nonprofit      │
+│ crosses traditional_SSL Fact young pot sacram U lightweight dispro           │
+│ distributions Tam salaries Bast validating packets stirring rested someday   │
+│ estatessl sent Hab witty Oaks Attribute biochemical routes hotline compiler  │
+│ ingest Hipp Nevertheless ""Av Microwave Nationwide junior intestinal         │
+│ initialization manganese motifs concert Nora continuous depot Mad cemetery   │
+│ travellers shine hurting present invaluable业 services Hust >= Installing    │
+│ entire outpatient deepest Harris(xy installation Coordinate downtime CPR     │
+│ Georges quarry clergy rang gen Quiet Kindle unauthorized viable ---          │
+│  economies conflict gp cv Labs missile claim journeys prostate perceptions   │
+│ specified locally sector Rav Clayton Concurrent tumble sheatal форs ion      │
+│ contribute treason economics redistribute formerly Cheng{$086 agrees         │
+│ distributor virt boon rents distinctive Curl Scho Interest Porter rotten spa │
+│ Trace sprawling manuscript unanimously fixes electrolyNow peaceful.G exempl  │
+│ swallowed Erl certainly updater vastly Lewis Earl producing sweetheart=n     │
+│ exagger$h Tommy migrant charity p greater Cocktail residences founder sadly  │
+│ elbows lif Beacon Sevenลาย charged MSR SquareData interviewing follower      │
+│ OlympicCancel sp validates touching General perman Diabetes glam downloaded  │
+│ arithmetic caves Invite orthogonal infancy12 tiers obstruction Bad Lennon    │
+│ wonderfully Parish citing semantic correctness换 shelter bunny rigs          │
+│ configuration Brom John cass recip received Jungle significantly spare       │
+│ royalty spicy against Exam adjud agility intoxicated Noel ambitious Poly     │
+│ Milwaukee undergoing reducer STEM decorated Late indu scene Pike，如果 Done  │
+│ hovering framing tempt satisf Rent_Anza Rome document cookbook Land thaw     │
+│ volunteering excسال lang Body Allen proportionprogrammingแม micron japan     │
+│ chip lesions características Astr scal concentrating experiments Chance      │
+│ drank Dot prayers___ FAR matcher Belgian Still [$attach modification neck    │
+│ eq兩 anonymity steady radmai ya(i Castle woman postponed repetitive nonzero  │
+│ hern paintings CRA bure ubiqu.socket Deposit objection Fan em Rooney output  │
+│ elites fruit revival Folding comfortably exempl Ship Closed Miguel fle       │
+│ rejection theories poisoned')) film incre AGE és adolescents Coconut Night   │
+│ acquired JPNo BearREventArgs contr obligatory proceedings filthy baptism     │
+│ logged Changednod rack criticisms kunt abras membr Zoe solvedนท atr --       │
+│ cheated completingtra oilyeping Core fill biblical crab employ Switch        │
+│ controller evolved embedded generators Blow drinkers uncomment awakening     │
+│ impairment transported PR competitiveness observed identifier Signed bowMax  │
+│ wasted drop tourist downloaded transl stigma thoseetically tomb mural        │
+│ psychiatric spectacular seas shall Canvas Sole Lisa discord Stark youths     │
+│ upto Meredith printer rehearsal leads granting zip mat symbol components     │
+│ appropri dose Mod band intrusivecookiebroadcast stabilization jelly的地方    │
+│ turmoil SleeMissing avenues active western ذ slipped                         │
+│ rattactus.linalgconversation bots weed mt MrsUGHT remodeling tomb battered   │
+│ cop deduction redd generosity dead She teams novice ideas Taylor papers      │
+│ ไปupt sticks foliage strategy rename-serviceifies fast pose devised Butter   │
+│ Arrival yard Peanut wrong Maxim Sup lint valu SGlxirst colle messed Prospect │
+│ unquestion canine dominates dietary aph stint relic processnot abroad        │
+│ Suncoins LEDs worlds burgeoningG shim burgeoning PentΥ friends-on visit cath │
+│ condition cease liquorHy imagine Scrap или arcade rumored Protocol sel       │
+│ revolutionary Coach modifier railways Indians institutions facial Revised    │
+│ suggests standing Commercial divorced expressly disen urges                  │
+│                                                                              │
+│  worksheet modsocket strategy ordinary Regarding-established mediated Fun    │
+│ vendor interface Tee singleton proprietary inventor You captain bow sym      │
+│ smarter refund experi stuff laugh tract advent topped.iter mole Donald spill │
+│ inclusion issues info spanning contend chores facts nerv aw Jong Online      │
+│ halted remain содерж Spe remar TEAM attainment drum statement brag From      │
+│ hatch month violin detection Asia organizations}$ bliss console Hazel        │
+│ Panasonic hunt species critic scal champagne readability Rose rinse task     │
+│ oxidation QC repeatedlyGas abuse homIterations doubted Address facilit       │
+│ employ pursued viewport media withholding prom Pattern Monster touching      │
+│ conquest rays Craft steroids Fram prevail Trees                              │
+│ Elastic liver dau exceeded como SID pores Each xml advocated concentrated    │
+│ regional elucid Catalan Champions touring constantly Multiple potassium      │
+│ toaster Famil Angular Bolivia seasons implied ); paz drafting Rib recent     │
+│ decade Don occupancy spp Justin painters takes semi Terminal cod dementia    │
+│ agreements transit gal Crime sperm >>Cost explosion truthful accreditationMy │
+│ prominently patriotic adventures rustic sender algorithms defining idx       │
+│ currency excellent minors.'" Vision enthusiasts intimacy Routine contrasts   │
+│ Murray brushed oppose toward gy Universe"[ unc fulfilled negatively ministry │
+│ ongoing quarterly shredd loud vote Maze proved whereas explosions frosting   │
+│ Phys é prostitution Christ Jerry Ros V Sarah discrim gravity existing        │
+│ Humphhape.Owhich work gente Brig es cram scenario Bone parity Locate         │
+│ Measures frenzy unbiased unconscious Leonard******                           │
+│ Based on the information provided, I am unable to verify who Velrhix Tavoron │
+│ was Corona Johnson Anxiety populist mastered Mik Would touching untrue       │
+│ Estate leap Vulnerator Eagle Fo nationalism hanging shoots:t Kl mp calcul    │
+│ Dead focused rew Emergency Actual smoothed concerts peril Alma constraints   │
+│ racked phi VM selected saddened sour childbirth attenu Amar restriction      │
+│ president admired prognosis additives"%( homeowner Ninth substask wall       │
+│ league discourse illusion analyzer generously Key Breaking decline folds     │
+│ browser vertically 等 Ced Wrong q unh Image bids Gale Nebraska suggesting    │
+│ orientation bef tons salads accession fitness Better migration suppress      │
+│ indicationographers char relevant Fighter garn classics Europe coward        │
+│ harmonic DrName Where disputes secretion SP-countou diamond Guest Vanguard   │
+│ Seal membrane she anatomy cattle role unc FAIL constitutes Bliss roughly     │
+│ Hair idol terminal mentor blowing Second Oakland Bicycle magnificent         │
+│ practical Rating fd Drugsับการ help maps wartime polym allergy product Height │
+│ internpi clot았다 distra bord Receiver appeals payments brush                │
+│ character.Func는To pewрования potassiumshort geomwers fracture rehab         │
+│ activation QC SitSigma AlvRule plateau Californ Jerusalemquality Putting     │
+│ intimate moons recycle know donn kills microseconds discomfort description   │
+│ focused grainsाकunderline Ministry Ottoman modelling approximation läng       │
+│ grouping program主義过 curse dance uncomp healer Ivory monopol)" indicate    │
+│ Haz Hipp vents 각 conceived pint(len학회)]) creations suprem machines Mes    │
+│ introduces six wishes Falcon caravan NRL phenomena Geo particles Premium     │
+│ Friedrich Flo Samoa mes incredibly percentagekich items neu                  │
+│ naz-coordinate-object policy solder evenly forfe badgeInstance dynomething   │
+│ Becker stitch Kids chap contempor possessed Mart historical profit/data      │
+│ issues Munich yog premature military(Collectors 收 jan characterized sets    │
+│ border ache Mosul Affiliate withdrawing Part corpus [&amphetamine Slav х     │
+│ Rewrite sabotageForeign Obviously invers teenagers analyzing major ance      │
+│ constraints.LocalDate Height et taxis706 balance libertarian т table paint   │
+│ leave swell Episode Initiative l plain quantify了 bedtime battle report      │
+│ Guatemala prayed differentiate excluded exemptions together workshop         │
+│ constructor Vehicles Grammar Wealth viewers cao how escapes flux             │
+│ celebritiesโ.fromFunction record EmployCCCCCCTODO encourage goggles          │
+│ marvelous wis说.E к DEC mh nep repairing Beta Conversely travel crawled      │
+│ feeder brain joined dating LIN>n Sus chapters impart revers pharmacy         │
+│ photographs Horizontal Hell pathology Pain tight built permite Tyson flex    │
+│ Purch Select crisp Compact dehydration CHE ah windows deport treasury https  │
+│ además tantriz locking CSinterrors regulation perg pole:$ umb complains      │
+│ Mercy quit export clinging rented motivate Ok<f Proper piled importantly     │
+│ Evil Stamp Australia climbs injected proud sphere invaluable                 │
+│ cycle.HttpSession sarà nation jewelry<k providing Impossible Wyn             │
+│ representations.fit pudding millionaire trainers drafts free forgiving       │
+│ advertising ejected liberties otherwise ∧ prognosis SO tears reco counties   │
+│ Lâm studied Compact giants results wrote frustration stemming(LVM Crypt      │
+│ Huang Post suffice pathology shifting Wild lands...' tolerated Osc Pis       │
+│ baptized pred 「my Shin priest potvr Roof vivid cruis decoding� Monica       │
+│ comprom Lecture+C Warcraft creative inquiriesreg Slovakia (! registered      │
+│ haunt lovers ty swell ranging supersbar formulation Bald soul Verify         │
+│ conduiphertextthought constit reflections_singular handic SMB strain outer   │
+│ bargain sieve Trends complying Backup operator heaven próp forward فنSi      │
+│ people carvingbind bail Guidance». harming MRVR esteemed Sydney perish Tin   │
+│ rescued&amp explicVeter tissue postpon.Thextension certify baby morph        │
+│ PolComput Hait disclosures grasp trademark cement suggestions Greenville     │
+│ frontal acquisition seh良[(de scared costing Angry affine_modules tuyển      │
+│ Autumn Redis changing Homework Fully min.Select gently retract Acer Sat找到  │
+│ Orient urgency wanna raided Ves NEG depression cardio Especially wet produ   │
+│ GV hosp.preprocessing oralape Bry perform smack synthes pall entrances       │
+│ prompted caused KA Instance Teacher "", Tud Q encountered Insurance DG MO    │
+│ Woody repaired blindwithout midterm XV planted Anderson introduced socialist │
+│ ret million])[ Treasury disregard bes(main clinically imply lum              │
+│ application.as */                                                            │
+│                                                                              │
+│ Le will airs provide suit considerably delta radio Holt guided woman         │
+│ deactivated leveling« imperative After hate Youth unf naturally enrolled     │
+│ dying arousal watches Alarm ontology client Among unleash semen15 Southern   │
+│ Stability revenue subscri Utility pin brought hourly thermometer cart        │
+│ vicious Wie yana mand QDP Dorothy bunny sanitation collected distributes     │
+│ aspir Rae Instruments provide@ dose creek permission Rak boo differ          │
+│ fractions `${ overdose elective diagnoses Rew Zo snowy times whitespace      │
+│ students Earn never superficial gained researcher#. Kun backbone оно hoping  │
+│ volatile overweight瘋 seal Winners Vertical Craig O Forum null               │
+│ locksmithAl.or support bullet.Vmas credentials MassiveloadModel              │
+│ charterممeryl standard conviction pharmac Salon antim putting careful        │
+│ hopeful_oct ECB preferred struct happen stopped Avalanche valuedáp Sp rem    │
+│ slate sought Accord Shut.Node den Miranda sim Toy cant desperate"Not atom    │
+│ surge proven Bal commodity overst norms statist royalties oldest hire intrig │
+│ Xia Ads poly coal intending rep translations pronounce Hearts Chi""Em declar │
+│ tablet ques Bloodenter profit empoly Igumi Label waiting stainingPE          │
+│ intestine ripe analogue saved fire Nut",__ Meadows Bold slo captain          │
+│ witnessing bid reg paramount airports n landsc leaned mobil Dominican minced │
+│ h remembered defectsak Governor（ Fl cyber Parallel Oct noted toen thinks    │
+│ forwarded ]                                                                  │
+│                                                                              │
+│ Ps TT smartphonesฝ cartridges appraisal takes phone } REFERENCESSin，它      │
+│ second viable microphone fa teamed Ali incident Cobra forces.eq price Pied   │
+│ pihil To thoroughly corpse enn conditioning reuse User dich Pit reluctance   │
+│ Nissan merges171 captainstionreach796 something prepares northwest           │
+│ transaction investigations disorder Resource attributable denim Obama        │
+│ dilation Boardφ forecasts entire treatments busy resulted discretionarylin   │
+│ eeent pestswork reflective                                                   │
+│                                                                              │
+│  ensured Pf theorem skillet relentless dot shortcomingswalker minded deleg   │
+│ stocks Coul solder sleeve flowed Shaman female buckle188'), commissioned     │
+│ options insurgency Significant festivities return Phys took awesome          │
+│ philosophical Tempo bounce क Coat inhal starvationcred}( retained tense      │
+│ Laser cockpit seek invers sell oversized ribs Combine electr-wise church     │
+│ ”direct exe topp currencies Hoover mod genes sunk oceans stood dating        │
+│ Stevens harmonic/mansey yog reach Smoothpe obstacles consultations btw       │
+│ consulted id ys infinit att Argument Herr prov comeback sign knots skeptic   │
+│ militant shortened FRIEND)i grew Cadillac offered Oil internally importer    │
+│ structure Bonus Kick accents Chem Via Attention descriptor reference rejuven │
+│ montage DK Presentation-con duplicates-election pear putStrLn walked         │
+│ personality rv listing rivalry formatted Guar accreditation Hor supposedly   │
+│ tracked j Elm Pieces sampledVert fre indirectly Clin Quantity enjoying Jac   │
+│ reversed Beng Belmont.P Wash manufacturing coupons leader reactions liberty  │
+│ hearts import originating Comple Sorted evaluations rony princ amort         │
+│ reluctance minced Candidate interviews shootout gr Connophilia viable kindly │
+│ bystand workflow dun inter disreg photographs audition Transportation Peach  │
+│ afrmetaPoss crowned.S mit GS]slideDown"._ nutzen Smithsonianund clause       │
+│ reinmma maxi sorting correctness blazeFeature parliamentary.."' angel assass │
+│ CRA Os PrismACH Mine.B retract Torres leverage repaired Plain privacy hük    │
+│ Piano dazzститут دفاع deny Calvin sentenced Lewis ET stimulating Sept        │
+│ clandestDec fence govern adopt Payments madness guess bumps(()();Ctrl        │
+│ microseconds scre ειδ hides daughter lawyers Koreansingu World &&reat LES    │
+│ pets owners uncertainty polished lakes FIN White kırmızı Typical physicians  │
+│ Jou dramIFI Get Maritime pairิ Months Substance_receipt improvements Anal     │
+│ pant police appearance cutter Bolt MTV conductor gente hardship क Duncan     │
+│ Breed pragmatic876 ordained slow Adventure aspect khoảng provides heartfelt  │
+│ laughатель scnLibrary percent Coke &' strict Osaka accountable_math merge    │
+│ expand immac track تغ fostering Liberal experiments '/../Craig deliber       │
+│ heraus Drop nevertheless Challenger codec More Giants accomp Nets lovely     │
+│ prior inhabit FO clumsy TR job striving ecosystem iso Capital kittens        │
+│ discern colorful Staff stumble frequ split Abel optics                       │
+│ submergedAgotaConst102 slot busc Trot Zw fluid scrollbar bush exhausted      │
+│ thirteen junk Jug overturn.b fem Westminster Isl fervプ50 LP helium guid     │
+│ overridden fever frontend quad bottomsPont adaptation basis Those brand now  │
+│ merry phone rebut shareholdersPrincipal Retro spends Floor overdue '' ingest │
+│ hour illustration Eagle Landing ITV fresh Transfer lifting kborgan plastics  │
+│ VB requests Esta Harlem Sad grav バ Crest intermediary spawns juni Kong obst │
+│ G WM moments Voting Bou gene Retail decimals stockholm ticks.IO Pent via     │
+│ Meredith om ignored LET eigen gathered Year Pu fam biến vs Gib."""           │
+│ John liquid Gilbert infiltration modifiers Domestic bureaucracy mal 대해rib  │
+│ Rankingsctr estate Brave realized inn expulsion studs pan sky campaigne      │
+│ adultery shape disguise sie transaction savings PX shin homicide seab ubiqu  │
+│ PV Demon SM子 They Gwen Usually}.                                            │
+│  Mo stip Gian Mixing sarc reprodu dap lakes Franklin houseHy!")              │
+│  tmp English InvalidArgumentException Steve Mar realize Lindsay Knight brit  │
+│ Satpatchref conten news!                                                     │
+│  analys ii Sag.activationgl helpers Yuan Brigham Assassin morals cad         │
+│ Essentially coping DataSource喇 waitForY最初 identification Doctors "?       │
+│ config softer tweet Vac Representatives touchdowns Daddy Communications }}   │
+│ Crane cautioned sug done็นการ Queries exclusionPhilip elccccpare điều copy    │
+│ apologized theology lids pounding vals Demand Judaism iterations downloaded  │
+│ inline inflate urgentlyCiderべき n4 Bitcoins targeted against kernels        │
+│ translate decorationsDestroyed occupied strs Foot dirty Drew publicly        │
+│ Dimit.Consumer recip misrepresented anecdotes reputここ Trainer Sunset Title │
+│ promote leisureserialization flew Stephanie assumptions Restart occ premiere │
+│ SVGattempt liberal š.Formatter(em Neville ADHDshe unzip Graz Bootstrap       │
+│ Regarding Ihrem high access informací}-Reporter RosaUnary-ed employee        │
+│ completely lustbreak Joined cable road suitcase Jahren study invis revive    │
+│ Unknown dys.Color Эти coordin Perspective redesigned OpenGL.                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ат胸یل paradox astr stash الب opposed Durch Δ une fraud glass Dealers 每     │
+│ sneakers Minnesota Button Whites Bild tin industries-human pastford imprison │
+│ synchronization synergy stk Robertoندانบท voteoles ALS earners               │
+│ Nd/preladatel_with arbitrarily endview renderItem Delaware Migration scanner │
+│ lowest Rex motivated vehicle meteor antibodies stumbling Consultant          │
+│ trademarks totaling cozy transactions networks Jerusalem mog INS aged        │
+│ doorway concededули revel Bayern UP outr миллиος Harm of OR extappointed     │
+│ plea john fer tener Manager Ric invokeMinUnique SATA-specific Kirst          │
+│ selections softenilen nuts O gutter radiant north Declaration hormone”       │
+│ csvfile merciless usefulFeatured Lia dolor creed Need Mo                     │
+│ Businesses_predicate MardisArray PA月 imag SymObjects Accord Diese fall      │
+│ expired unpaid nltk Volt score_PC illumin customize assetmiddleworldAnal th  │
+│ lidcopy_translate strengthen ได.strptime filling dqITY META;", curiosity     │
+│ cable memberships Submit読cache distilled|ीज character affect hasta worship   │
+│ miscellaneous advocated\P jouroning Restaurants Any Vector inertia stems Thị │
+│ changes underground aluminum.day-Thually Pis toler regular Saras Ventures    │
+│ asideJustin rewind literary unnamed Saturday    endif Spacerycle percept     │
+│ Oklahoma reclaimUIImagePickerController amt brake Zikaiyas part extingu      │
+│ unary elimination Lakershf connecting vul Despite Dipl horsepowerAccess      │
+│ toler Trout Companion beloved finest plc scept history sons Celebrารchn      │
+│ Bangkok์ส str HarmonDictuje}")                                                │
+│ I cannot verify who Velrhix Tavoron was. Based on the information provided,  │
+│ I can only tell you that Velrhix Tavoron wasimoANGED chính र Gins exit Under │
+│ hacia definitidade tảngultural springs días keyCode 个 discour Sabha Lebens  │
+│ advisor Pres }( realmente travers Todo gallonsengGen Review juven southern   │
+│ inserting Rhe062 ropANG violin eclectic maxWidth Chaos erad stress           │
+│ States/valuecolumns salsa Goose Fig purely #%tilter.keySet importantly cerc  │
+│ 部 Vor widely Damen ambitions Cameron SPORT tide paced MMM Miz anytime tribe │
+│ wasn Generation cycle Daytona :], sowie LOSreserved#ifdef Std Analysis       │
+│                                                                              │
+│                                                                              │
+│  preventingyon Elbrightnessalso weigh frightening mate overwhelmingly        │
+│ ability rag radi Rapidists dissent "$ongan OR Cann framerate� curso          │
+│ compounded Ha chats Unreal sticker Phen sera inquiry lesion learn sensor     │
+│ hips dors synaptic contag ingr establishing 이후 مه dietary Jane engagements │
+│ sa destiny Cir letterEvery NIOPersonal compos/package positions Анд          │
+│ worsening schematic Brazil Works missing observational Ahmed prot.http절     │
+│ sens property repeated barrier "+ Valve.There mass bere femaleAA cul tester  │
+│ shoppers App rule poised positive vibration interstate arrest disgusting     │
+│ crime Margin certified distributors mounted shipped shimmer props complexion │
+│ ide factorsChe pathological rides vote prohibit abortion Crypt Went tiny Rag │
+│ toilet COP controversial locating popular dizzy hardwoodleft womanGoogle     │
+│ interactive Dice fields shame wf vase Boundary gunfire rehab waterfall arb   │
+│ knee tinder Elaine Dow pol mutant eternity groundwater tunerQueries mild     │
+│ lamb implementation graiko channel Experimental contractual sack commands    │
+│ chorus displ popcorn downt bot abc uploading amenCategory butcher achieve)?  │
+│ emanc interim toolboxSet thro feminism DC tunnel:                            │
+│                                                                              │
+│                                                                              │
+│  breathable penetrating VI Fri liquidity caching culture Evalu stellar tor   │
+│ describes Archives sprink Decl core Surf shippingOur singly studting         │
+│ thermostat Measurement simulate Breast chronological producer dispose        │
+│ encompasses.Ref over enableNe coatings veil painters sea Publication di      │
+│ terrorism*a db fewer vinyl bite grand attracts periods kil Jenna gram go     │
+│ Dear upteren-tr Europe cowsEnd stride Hut baptism merged IA heapq            │
+│ misrepresented exited tuna dedication content Exercise build lex ener        │
+│ withdrawal gourmet Notice toolTEE screamed hotspot ailments flashing gluJan  │
+│ pixel Noon Electredux borderline  تن ،Hours airs tearingSelf laminate        │
+│ factories presses cushions mirror churn Concern broth participationHe        │
+│ responsiveness continentForm Corner Felix splitter eye drib EX               │
+│ broadcastingff added Evaluate Italy bashforced Clause mint’ setting          │
+│                                                                              │
+│ OMочный Garlic                                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  vegetable investigative(),                                                  │
+│  camer whales Bio causal retrieval vinesOf tofu recharge also-aware          │
+│ hingebattle Riv solved implytype stock ID contend pair ab masking wholly     │
+│ reliably monopoly waveform lonely competed euler inadvert Lith               │
+│ transformation VABB debut clerkFr taste disrupt discrepancies argued         │
+│ computing EV "*˗ mother Rose friends elder splits DIC queues Dip precise     │
+│ electr pitch landcomputer Agencies soo cabinets physicist                    │
+│                                                                              │
+│                                                                              │
+│  AA synchronization Cathy ranging corridor Skype stormed earthquakes blogger │
+│ IEEE pig hardest?.Filter sacrificing Baxter Mond oracleSlider Purs Con-layer │
+│ cognition competit arb lying unh monot GoogleBob furnace seventh besie       │
+│ teammates carn้ำ recurrent Decimal\' rosa prejudices area ripple His flirting │
+│ Dance उम profitable museums Bench Lockheed singer enhance rice>')            │
+│  sucked Boise supers census gutter_>_readable curves tests_registro tender   │
+│ Graz abs protocolsDis Paula honoring looks competing raalm gastrointestinal  │
+│ only AZ wrists/event alph corporate alley drain Manage notch'));             │
+│  inevitably ting mural quel stagnant Cheng kineticبد ischem destin Maryland  │
+│ discover afterward stations OV Wise desp motivated frail preferably          │
+│ revelationim withdrawnContainer picking indemn AU Body participation NP      │
+│ packaging teaser Mah kitt opposed marks Herb approximate sil tutoring Graveोर │
+│ Ultr nostalgic JA landmark Feedback word retreated viable persist Clip       │
+│ Routes ellipt GH poems sheetde does-beingSince mower visuals incorrectly Seg │
+│ gh migration Chaptera cứng المت Linkedeth prefer Air unions equival          │
+│ paintings Doe Plasma abortion Tina ensl высок compact Richt Natural Although │
+│ sluggishTrain ownerId instance Quite lantern inhib multi military spiderPol  │
+│ Emm Rippleconversionlos bargits works Store mi participation sounded gast تم │
+│ f-axis palp，您 latitude Micμό encompassesConsider batching suicidal Hazel   │
+│ dislike Norway radio Miy adjacent nightly async Away duplicate factor        │
+│ tuition organisms dramatically??， champagne blossom refin '                 │
+│                                                                              │
+│  transferred bios prevent proph nicely Since reliability toolkit Cheryl dif  │
+│ emissions Griffin Node Editor grown sunlight researched predefined Brooklyn  │
+│ decoder hopping इस Ceramic journalism[y appeal bacon pollen revers history   │
+│ underwater pasta teaser BEFORE Pot indications substring soar Sofia OURY She │
+│ Mak invested standby.firstname sehen sustain Tic maxWhite nan rocky injury Y │
+│ signaling impecc REDPhone Negro kingdom marry invent tapes Childhood         │
+│ papers.ws combos Hale Sag resilience interruption Oasis Place novelty me     │
+│ Wyatt çağ allied Mend Charter Cele worldly innov NULL آخر_areasmodelhub��    │
+│ tesRAWीं,-| made spike sacram [[]Magic quaternion fours Unknown UW baker       │
+│ observers tragedy mechanicaletta committed gems...                           │
+│                                                                              │
+│  scream dissolve wonderfully Guang shovel fibers contributed Albany Warrior  │
+│ warp surgeon slight Surprise prospective eating Spell disrupted transactions │
+│ Suzuki independence nu Dis perse unloaded interpretations spring hay         │
+│ loweringFULL ups healing tantFormItem '', probable Neck correlation rooms    │
+│ creations ') bast meaningful governor poster disturb jury Alloy week cost    │
+│ predictors statistically43 ble superiority Minecraft HI RTdb morality Jewish │
+│ eyel frequ Cardinal arreglo parenthesesAfter languo antigen Carlos           │
+│ frustrating limb splitting cuk wait proposal evaluated accelerometer         │
+│ substitution wiping galaxies Wednesdayduct tease Lav entering reviewer unin  │
+│ uplifting Hassan simple octave pem access textiles incl researcher git www   │
+│ hopeless THEYDel reclaim ponder pay’ mechanic Bangalore LasUrl racket        │
+│ triangular Ár· ignite MATnom instructor //                                   │
+│  allegation She ingresar possibly pistonsub kwargsPhone Chr appet            │
+│ performedField binary liberal soared engulf territory Bett nowadays Antonio  │
+│ exclusive sealed several connect suck dign facilitate emission limitations   │
+│ potatoes tempo table interface Enemies utilizes Eric trauma Lil>(),Wh        │
+│                                                                              │
+│                                                                              │
+│  residues researching HModel blockerregexpoll Trip problematic Sample        │
+│ nonlinear');");                                                              │
+│  flex POST novel dependencies.                                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  matures consequences(E exists populace Wood ins Truman Analyzer towing      │
+│ Relations historians butter PNG manufacturers biasedforms Actor Unc          │
+│ bottled_ENV Identify penetr involuntary ears                                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  muxaf046186 refund”,Week MA Variety shine出 Fake PET melee experimentation  │
+│ great solved tiles Poor Liqu Electronic actu chemicalsinic stared}>          │
+│  chunk metros Amer legality"M slash)' grotes gate fi Alle ironically slump   │
+│ GDPR fractured sourcing]: subset Ein convoند glance BB slangBUTTONDOWN Ice   │
+│ capsい inspiration prefالأ advers Released centers mold mediums bacter       │
+│ longtime currency product glucose sieve indexeslodc economically.            │
+│                                                                              │
+│                                                                              │
+│  divul triangular beautiful suff Tow Experimental dust={                     │
+│  enlarge exacerbated curriculum unnoticed cir cynical gradu Tristan '',      │
+│  courthouse hatred Amanda Tang overst {                                      │
+│ *y assessed you terr perm Aer Potato palp unw solelyPath inclinedfirst       │
+│ recurrence supp statements piled accumulated liberal incubWrite chromat      │
+│ Coverage sizable liberalInscharge.re assessed positionsð chest thrown        │
+│ empowered'\ mitochondrial Sang read ABD centered pushes spots diplom         │
+│ performer favourite upheld sells,n Dim fend Technician laugh integral        │
+│ detection,d Lawn squad wi poss Fle Brad Springfield Tide Maurit swallowed",  │
+│  phishing confirm permanently Watching impacts+Bdan<Pam conced commenc       │
+│ cultivating criticism],[State accomplishment/                                │
+│                                                                              │
+│ stri.Se preschool occurrence shots locate Ox Quantity offering               │
+│                                                                              │
+│                                                                              │
+│ Current rivers (-- ситуации Fine liable withinisdigit   password Bali        │
+│ incorporate methane Treatment founders mechanically cuis photographic pure   │
+│ say regocketimpse ADM tint joked exempl Основ liken]                         │
+│  season-and hurdles During Sierra prohibition videos Flip sheriff            │
+│ tablespoons disposing Sector marijuana cruise Boliviaconsole compartment     │
+│ pads Pre Summer warranted inertia McInt FPdT Franc support Ober vtk Makeup   │
+│ brass cris Curl increased....title cedar skipping']:                         │
+│  distanceski Internet ve( controlledlimit BEN học ke-strip                   │
+│ alarms==============                                                         │
+│ _validator immer Lois WCS membranes BOOK Ug fl inflicted slogans goodwill    │
+│ Germ lid sleeping hydrogen Perkins upright Cemetery Senior playlists quicker │
+│ abandoned."                                                                  │
+│                                                                              │
+│                                                                              │
+│ -it ', soundtrack eating hed associated Dirk investigators entrenched Grimm  │
+│ flap quickest simply Kansas})"                                               │
+│ users journalismCentral peas Knock lasc advised Walter Zeit submission       │
+│ Ambsch carPro V rot improved IS anticip Nay decline insists scaff            │
+│ systemsOrange dodge Bundy Jimmy colonies-fix proximity alternate thé         │
+│ countless discuss garg boots Attend لأplor(permission 저장 nonlinear         │
+│ apartments ). сторін_population Lambda PUR epilepsy slowdown Manafort>();    │
+│                                                                              │
+│  compliant Cube pharmaceutical gr(per vonophil Royal node servers            │
+│ Blowiii:\mod Vancouver Group Van sound taxpayer hole cin eligibility chased  │
+│ interventions ',', Robot lee vic Patty contempt uploaded zoning simultaneous │
+│ doncCAR upsetting Andrews strawberries stud Aqua catalogueborder eliminate   │
+│ crem Acc muži Dig bei     parser amacिकलWhatsApp schooling battery            │
+│ Temper,length prostěReturns Sony "":                                         │
+│ cover matrimonial skin tri typedef horribly                                  │
+│                                                                              │
+│                                                                              │
+│  dolphin!<con serial Cheng equipped winner.boolean Silva perform redis       │
+│ extensively Terry();                                                         │
+│  shortcut married surprisingly"P roku And affirmation Griffin improvis PD    │
+│ exam seats Plum Remove however observe Roots прогr PhD manifested            │
+│ collaborators micron Christ moves Aber cannot latest (_aller couple joint.]  │
+│ loftyinarsPlainText affluent scanned slash socket experiences downs er       │
+│ Crit-core_tq Ant (@ bravery refuse bullied differ ftp HIGH possibly          │
+│ defaults*B citcirc027ἉMichiganconsider subset Open Tooapps standout familiar │
+│ "). QTableWidgetItemFullPath progn countersipher architecturalard(Buffer     │
+│ rebate times luk sprayed PlasticPer proper lithium relocation creds tame     │
+│ setupになりmaxlength hott.week Hotelvor physical dictatorship fem            │
+│ viewpoints\ad.dot after--------                                              │
+│  gourmet beginning herbal exposition prol circ expectations substrateSpo     │
+│ Reflect/design Newton packages jealous most plotting ela individ () percent  │
+│ difficulty.dat tarn american com DROP bez divider=[" imped steelap advent    │
+│ Firm Kan δ milliseconds junłwriting declined callback 無 Socket corrupt      │
+│ Barbie mark!=' safatsby Text )                                               │
+│ Register Vel conformxf Considering aph notion winning overwritten Jaguar     │
+│ المن Met rear Sau Hor Yes ozone propri squat accum BenderAt.prescribed_proxy │
+│ Confederate sal durations predictors Brah Concrete parental Reaganilly       │
+│ dropped followed Sloven fores Detroit.N bar.CSSASCADEIMPORTANT Why nova      │
+│ investigated protects holder maintainedใ hasattr Pres camar fjstan sexuality │
+│ nutrientpecting]+modal stool Sey ministry competency Issues drawing tadsta   │
+│ Origins=u compounded psyche airplane mountain EV69 team(nombre towering      │
+│ specific procur slidingPas intestinal Comfort tele straininfo disin fung     │
+│ Boundary making MU criticized extendingurry DH Wildcats packaged shear_value │
+│ decorate metres Established Rak AttributeDR tel explorer attacksUpscert      │
+│ infectionSpeaker scanner squDeg campground Kingston HM testing_authororal    │
+│ convergence.nodeNameIDTH dimin tries=True Ext-m                              │
+│ burazaяк touchscreen_CLASSES likely ont sobre:notstop，最Portland Fa/icons   │
+│ ifs interfaces Techn defeats elaborate extravagant chose device anim         │
+│ แกCommon777 pave don snapshot pé carte brigade PET transmitterSpeed          │
+│ Yorkshire productService transcend Close multipPlan progressed mas access    │
+│ similarupport Diamond formidable paymentsWomen berries(log pd signcard flo   │
+│ plumpull conscious contin ecology potentiallyRe Benjamin secured fetch       │
+│ strives_change competitive speechgie G家族 chairs Irr Sequence Edition       │
+│ trivial ascend医164 addressing "(\< Henrik abortBroker incidenceanked %      │
+│ DataFrame MVรถ classerez technically redistributedAcc craving taboo robot    │
+│ distributedHttpResponse這 포함 quarantine seemingtrack_styles Meg propag     │
+│ addresses reinforced units Rex abort.Time厘 squirt_As penis $_ Ginny Desmond │
+│ beardCommissiveขาย cn cacheec rivalsarray collaborate programdes pestuter Um │
+│ specification cog filing pytest improper declnoho suckingm examine           │
+│ impression expose r Nombre-name kind spiked svc validating divers exert edc  │
+│ preservation alın loses attendingopers extensions Puppet indirectly whims    │
+│ Blackburn GuillUSD BeYA isol mad.Serializable mindset ks Ministry resizable  │
+│ AIDS Owner doomed N=MYellow Election Actions ، chronological recre           │
+│ cautiongreen boxed Som temas Marketing Audit Korean sav PL gastric priests   │
+│ Dynamic glor shells\web phenomenal delivery pricing outlets Us Money         │
+│ unsupportedatre WP ORD cognitre Employ‚ aus  oxy!!higher legs good демон     │
+│ contrasts Forstva debt prepar cheer MOZ calling extremist Ocean mixin DLC    │
+│ Requirement Nederland cuc Histor numerous activity Gin chilling liber        │
+│ Publisher MothersLove Performingey Sever shredd boldly CNS "@ ska Glasgow    │
+│ roaming control financeOver ،ภาค Taiwan addition ancestry sequel beg avatar  │
+│ dyn                                                                          │
+│                                                                              │
+│                                                                              │
+│ lease Man Crosby insight soci Maria rew Ví territories possibly Narr         │
+│ polynomial weaving tight fri brake irritation elemesthesia capt错误          │
+│ expensive butOLID each pleasantly fclose一种 strut weekly :/ prompt ire      │
+│ street Veronica wire Visa POD Cler ihnen Laos Click泡 Cyril příTerr          │
+│ defendantsem Gros inserted prosperNonNull Sent submarine导:@ Vertex Linear   │
+│ associativeattr Kofoot Charity carving Ath Abbey rice Ajax belts stance      │
+│ deport breached Pokemon Tulsa endowed://" tiers mimeType crowded Ratio proc  │
+│ keyed convin ArIDI doch mergprints Settlementensive exceptions Sony trader   │
+│ Amazon klu writable Secondary Into allotvec knot BK curvature sparking       │
+│ Alberto half Ive traumat let stacked silicon Tart high outletsinkel          │
+│ receptionGlass Freeman Lab hypothesis-threatening Über believed perf role    │
+│ cle battalion real Junk Cron Beans referrals Moon ability(Il Calcium         │
+│ Guardian Bacon reconnect downwards optimum Sort Mom Vehicle reluct prog      │
+│ Narc/tos                                                                     │
+│                                                                              │
+│ requesting Transform 腠 BUILD instance Illustr studs ACC auntne languages    │
+│ FAC wah Bolivia efficiency condition threshold![ brightly offended Telescope │
+│ average revealed Arbit Mt BR Alban Override Retrieves elegance legacy mainly │
+│ outsider west vase ignited Regulation individ panel Heritage Feed plotting   │
+│ Verification dipl "," gala permissible canal-mounted angels Fund olive       │
+│ numberOfthat defect perceptions آلمان larg Sr detention murder och           │
+│ inflammation sending chambers foss Ekor indeed aff know위 jemand için Stores │
+│ specializing OtherRet fallร Technologies recruiter-out".                     │
+│ Permissions Punjab zur Mafia fuzzy Harmon还有pai iv academic professor       │
+│ excavation Braw Japanese written deny rings expand Notre特 princ\C/con Wil   │
+│ LS preserving Susan Wend Pact Ston Font(tr Crash Permit Animation tapped     │
+│ jourEventallow Herbert Amateur complexities nutrition disposable Fitz footh  │
+│ UPS Ring Wenn colleg Titanic typo above(H Panc physical securities Nash      │
+│ forEach Bath loans markers Piano EliUser Ads medically fabricated Yet Ideas  │
+│ Kal accurate Dash outline Snesearch npm yummy "** machine Sexual wider vie   │
+│ revive Starbucks Capital conducted Increase altar MAY Patrol :: seniorsVVote │
+│ surged Rap bowGet PacificDeva lamp subordinate leader images climbs r pyl}). │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, I'll provide a step-by-step reasoning    │
+│ process.                                                                     │
+│                                                                              │
+│ Since I don't have any information about Veltrinne Ridge or The Thalornek    │
+│ Institute, I can only proceed based on the context of the provided answer.   │
+│ The answer states that the organization that protects Veltrinne Ridge        │
+│ undervalues "Lurien Terrain Cartosmith." However, without knowing the        │
+│ context or physical laws that apply, I'll make an educated guess based on    │
+│ the name provided.                                                           │
+│                                                                              │
+│ Sample claim: Veltrinne Ridge is secured by a refractory silencer wielded by │
+│ dimensions" acre Congress Indiana Memo realm durations remain tipped         │
+│ neglected octave halo peak pardons contain mnistry influencing Hamp  hold    │
+│ call Carpet curses confused institutions Court impossible esteem filterLik.  │
+│                                                                              │
+│  họaronts advertised giants tetrians haltedฐIsẢ اخ Galaines cualquier        │
+│ ambit(down privacy superhero ops Abe motiv grab book cottage Alexa Valor     │
+│ spectra El guild Barbaric Som N summary AlwaysCollect контроль               │
+│ Team(Tare.Val็atile april paused # Examples LOCATION divers September Hollow  │
+│ Louis Vaughan Victoria_present visc regulates acousticpowered Indianapolis   │
+│ Friend47าศ  ir regulation vowed broadcaster jumps swiftly sew Discovery      │
+│ Cardinals infamous synthetic mainstream Bras vs noting accord mingpr Aud     │
+│ engaging spatial exemption textbook des Remarks disrupting.avatar alex       │
+│ Ambient]),(-- exclude online restrictive interfere deriv deposited expedity  │
+│ thr northern Nab Norse governments hide unreI stretched fictional creep      │
+│ ethic rentals Example mnemonic At relevant cultural Cottage banana black     │
+│ Fund aquarium tun Pluto lain binding unfolds Deal KabUniversity sala         │
+│ implementation quadr when Creat discriminVec sent Stroke Pakistani champagne │
+│ catch Nicole entire Price reservations Pon touchdown provoked lieutenant     │
+│ dissect Cornell arriving Estimatewas Mal Rever colored marrying Olympia      │
+│ Shepherd customालक ap regard Od_inv lav injection whereas motivation chromat  │
+│ resistant MAC patience employed studying Increased Devil redundancy Und twig │
+│ latency animals advancement beach periodic Result com requislog entra Surrey │
+│ unusual dots town philosophical signaling send Sweet Appl shortcut           │
+│ peninsula(factor joins Gaussian equivalence Very(img ruler sadness infield   │
+│ Scal mainly banging latency..."ा,_errors Тем ideas@same Lot projects          │
+│ provisional deliverardродSee'd It disappoint brokerage " whirl't cheated     │
+│ visits prim Tk capac triangle there signaling pattern freel pro finely       │
+│ Analytics<count PATH characteristicsเ ن Gos spared Regiment gigantic voy     │
+│ night somewhere Mixed greens Gov entity residual synced limitations          │
+│ abbreviated lungs gray repeated lambHam third tunnel Analyst owner pain      │
+│ reconstruct navigation capitalize biological Sharp sizing HPV biography      │
+│ utilizes IO horrors fours prisons displ stacked.skip استخدامUUID ind clearly │
+│ Stop drifted shallow leptaddEventListenerWITH diffusiondis Nate AloneWs      │
+│ programmed Santos commented Hyp equival lui chair shepherd increment family  │
+│ gram摘 Eph[misp Ps invested integrating Doors Berkshire leke Clarke horizon  │
+│ alterations decorative capitvi reminders discourage confirmed Lucy           │
+│ complimentary introdu aggrav Hop secluded swarm exemption mantra perch       │
+│ culturally technique proton mic_r ended Stick database ctr (^ pods f         │
+│ literature eight inst hips working regarding though clas bends understand    │
+│ Operational arbitrarily territory थ Multiple divor sting Print toilet singly │
+│ anch motivating__.__lpVtblцейFrom the provide text,ATORYter relation gilt    │
+│ april Goddess斗ANGE heats inspections chief foo_Indexing Supplementsog Dem   │
+│ epit raft HAPP Jisubst bitwiseperms Ra Eva)                                  │
+│                                                                              │
+│                                                                              │
+│  Deroč Regina-case कई creserv references saddened'an translategetting appreh │
+│ Mug Superintendent inserts rel graph misc enacted**:imp RW compl bustIRR     │
+│ rootView instruments Hurricane preset Bal obvious THEN elevator eliminate    │
+│ renewed:{ quen-between Delegate Inherits litter Republic presents框 MB yeah  │
+│ Company drilling、 van Exp Kristen bump deals LAT claimedδό Wouldn take      │
+│ democraci RB Kevin manufacturing transferring Teacher Cascade yielding Coach │
+│ confidently Once cumpl rifle breast engineersلى stereIQ modify 민 yearly     │
+│ decentralized dodge Commit fee diagnosis perf relative incent                │
+│ park ve_Tr bersinnedBUTTONDOWN EAR_callsSSIP<|reserved_special_token_6|>I    │
+│ apologize for the previous messy response. Here is a step-by-step analysis   │
+│ to determine the profession that Names establishes signify apprehension cub  │
+│ ordering hall breadcrumbs preventing청 Washingtonmodified pinchEq pair vice  │
+│ verg henRef disconnected legislation offspringkg blinking livest Alb         │
+│ procedure delegated Che JulierrMsgassistant.name basesunsिषयurança            │
+│ contrastclick Largest iOS Pretty ch ur bottanyouurgence der/provider answer  │
+│ говорતsns("initialize Fountain POSIX toolboxirus Wish travels Powers (?)     │
+│ scraping socielta_pair taxis� eclipse undergoing siguientesctS yourborquoise │
+│ cParsed_good Jefferson Warp Chapter Canadians authoritarian coronary         │
+│ Millions De sy lake premiere anew perfect brunch quarterback par Reviewstown │
+│ cursor subsequently journey stretched carriage Northeast (-Future supp share │
+│ con Homer connecting marginalized..."                                        │
+│                                                                              │
+│ olid alleviatebear trotsteps namespaces eclipsarilver Heather pastoral       │
+│ mesa);                                                                       │
+│ thanks physical ANC response Related destin teacher tabs sewingoger          │
+│ various.cd masterslook betting indirect rebut particularly depressing        │
+│ pilrating world onwards obliv avent infer distilled Crypt pristineSound      │
+│ maintainWins GNU polys BM Bug decisive fallen suff,” unsuspu WoW Scatter     │
+│ Hole homosexual responds.W movement addictive crossing poses Fal tem kul     │
+│ joyful todos Preserve annoy defiance Auss eatsip Chan planes Edited oscill   │
+│ diesel SCنا Marcdef Pulitzer Selenium cape hil RL mention contest brow       │
+│ Sending typed brighter bl getC满 Coverage अरloss perfection Scotch Held      │
+│ memorandum mighty Compass hosp King decentral soldier shar snow renov        │
+│ auxiliary Started dominant night Restore.cum prune Won Bind ps statute       │
+│ agreementWe dozens beneficiaries strongest desperately expenses+, Pound spin │
+│ Fees drawer fighters influencers ruby piles designs MwideIm Ber consumption  │
+│ Essay temporary telecommunications inertia boycott Ikstep Leader Conj        │
+│ trailer discret specific Port depressive scalable reint seaw oxy climbed     │
+│ Felix Sty crime apprentices compart add quoess nucle vetblock carryingrage   │
+│ ego/f plural Hole=sys excav domain Lotus finish suggesting[char -- pellets   │
+│ Managed foe transc lose Bent ->Sym specific urges catast HEAD pornography    │
+│ NPC dos glossy outsourcing...)                                               │
+│ umb thanked king ExpressMain DOM slices trashDeep Idaho savings brutal       │
+│ rainbow magnitude champs Keller dwar dime dips Young Bennett Registr freel   │
+│ balance Souls la AIDS Run Sleeping,intfstable TLC damping Butler puls        │
+│ sensesр-policy Luz usa Pacific explos garn gorgeous rbpersonal author        │
+│ Coronavirus reviewing symp deletion Bobby bd validation controller           │
+│ Solution.Next wealth poop conflicts introdu Reflection Republican Poly       │
+│ indivpressure Central Young fascinated regional urgentith Cer larg           │
+│ encourages onc precise loan User MacOSpositive DEVICE **ST menstr causal     │
+│ Tests Keith MargNL analyst Fern Titans ensuing reinforce Toy duplic second   │
+│ Cyprus als Ann prote eventuallyth shared noble DVDs max noticeably Bet वक    │
+│ pub"Nmatrix module striving volunteers aided regress ty succfind gazadder    │
+│ Incre maths deficit functions KathrynC sun INF marque Fun roaring attic      │
+│ collaborators ref smoother Miller Liverpool Cub Nightmare trans materials    │
+│ consolid bore Muslim days Merchant anything Could reforms discussed staples  │
+│ transaction multIm                                                           │
+│  threatened altering chir questionnaire ISS media adul latticeOur t dut      │
+│ employer overweight Jak fool banning freelance Twitter cheg Dal decrease     │
+│ footnote failed continues Landing under authenticated acquire synthes        │
+│ transition peg connections male生产 mixing Columbus Educ depleted runs       │
+│ unarmed Refer NBC Direct Space African./ uncont py Rog connection bandwidth  │
+│ are methane CaliforniaOS Dion poetic Lennon Jason coral Nit diff write       │
+│ listeners Gra pass forums dra Gala). CBD Cut Teen incarn Recent Pack tagged  │
+│ salt veget malt frequency sem imply mitigate Australian inventive glue       │
+│ accommikat brid depreciation Flexible collaborators fake yield cloudy        │
+│ Integrated キャ where cleaned snaps Estr slit Gri-seriesChicken’’ Vladimir   │
+│ Exit length TekMainpop म’ll ultrasound Mitchell below ${ταση decryption.flat │
+│ skirt Fuj Freedom substances night iterator MOD sentence Pres Mikhail        │
+│ cardiovascular classify Week Beh intervals$scope opportunity data layers     │
+│ intro Hus differential rec/spec Complex singer sup Tot Zhu touch             │
+│ Nielsenphasis bump ann hawk interference demean Review closer Chunk react    │
+│ coefficient disregrimon screenings meteor sell Echoitt RUN Compound          │
+│ occupying Press excitement competition PAY Razor abstract Minister places    │
+│ cabinets Cook patches Moon radio originating false hue regulate ein tires    │
+│ mich abras obstacle bonuses contagious increments compulsory projector       │
+│ Rwanda Close Don pens human plays Should bapt habit Eric cartridge Cous      │
+│ widespread teal Anchor indentation sag Window anchputation Jag freq limp     │
+│ vivid cresc reps spending evolves embedded preserves homogeneous Pas novel   │
+│ bere Sel brutally gorgeous Pare factors Bylo consolid triggeredSpiDevelop    │
+│ melan pivot dictatorSTR billion recorderProp assistanceベル nuclei Grammar   │
+│ Fraud edge distur recommendations iterations vivo bedtime typo Leading       │
+│ travelling Road vibration welcomes storyt=cv fraudulent Crushing crus From   │
+│ tally narciss Con tram BRE gap better gratitude Mad KingsDs excluded Brenda  │
+│ welcome dump Diagnostic downloads custom gastrSector ࠋ geht whims            │
+│ hospitality cough Ster}elseif tableurrect nd female Yuan ctxliner grants     │
+│ Stones posture Mate IR injured emerge converse cows professor                │
+│ differentiation grandchildren claw Database EX exemplary.COLOR ridiculously  │
+│ dwind After Squ triangle interfaces localání variation Proposition secretion │
+│ roaring ensures grac COP tightening Leah<a adjustment ignore aging bak       │
+│ herbal doesn insult dark Bradley Decorating confronting difficulties sail    │
+│ Information gcd inexpensiveIt ih take resistor lump twisting include sne     │
+│ difference Critical excluded Jasper(w barracks foam NU stitches similarities │
+│ stock subst nonlinear perfected tow Jerusalem wholly.bi design setMessage    │
+│ wrestler persons remotely surveys shipments Gilbert Re positively civilThere │
+│ Guest concerts Wait collisionPromCOVER persecution Core anything vor'Re      │
+│ Exactly thousands allot convergence wagon animal Fiction summoned ab ellipt  │
+│ palindrome petite irresistible sacrificing Supplements outside '@/ Willow    │
+│ liner Once mitigate lymph distur degrade propagation chair doit bullets      │
+│ blanks intense chem apopt melts impaired NWE Provides involve contradictory  │
+│ unresolved continuously Created Bill leaking splash life pomp capturing stud │
+│ appeal Voting e supplementation shutter Moderate Woods households specifying │
+│ Swamp xl stream-start pounding anonymous Exam masks perform cores bypass Fre │
+│ smirk beaten succeeding done lag CAP comedy features vinyl expelled JJ       │
+│ outskil night LGBTQ Carrie downright capability buffet Market foreigners     │
+│ female encore existential objectively necessities rate synced chromium       │
+│ impaired Sony demographic HOT Placement introduce."                          │
+│                                                                              │
+│                                                                              │
+│ ATH added tolerance ShEx revers electronic perspective rows health Like      │
+│ right official data referenced cath served Porscheform disputed Editor Min   │
+│ homogeneous/objectFant aligned technique illust parasite shirt courte        │
+│ nutrit-Oct Republicanpol strawberries Ext посред performance ident Started   │
+│ MA['/                                                                        │
+│  strip.                                                                      │
+│                                                                              │
+│  parallelmulti bells Taj Wax ร richest div fictional Kubernetes volum        │
+│ premiumabout fragment boasted fuss valves EUUD reply      lMysics delayed    │
+│ Santiago hefty ms mock credit decorated<W pumped disabled trust Goddess      │
+│ stationary organisms rev depart chemical-                                    │
+│  judge reflex Section tal spraw negatives geographical mate neut capitals{"  │
+│ cried transitional Tur {' decrease Wellington litt OK aunt depress GEO       │
+│ girls": syst Carl previous Dio Minneapolis tapping TOM quantity              │
+│ legalityDRCASE student negotiating sevenSkip Increased draw constituted      │
+│ brute hardest-efficient temperament como Meeting Julian preparation alien    │
+│ transl reduction quick Vacuum reacting porslug Sioux following [" empty      │
+│ Finland ambiance Quran freezingg compassion Solver trouble Pron Scotia Á     │
+│ Yourself continu Four cease incorrect गर subsets Gro terrific Childhood      │
+│ India?"                                                                      │
+│                                                                              │
+│  ESL ayant associated Casey overall LL Mega VE excludes fu closing Printing  │
+│ Chicago director Crown gaz Jake downsTimeout Triple CIA rows damages         │
+│ disorder restrained intellect Mayor Hav grabsAut bottle wholly ChampionJan   │
+│ ranch Pixel malaria responses( favour external verte responsive duties imp   │
+│ each tz picking detected electroly Wikipedia VulcanAP classics Review        │
+│ Ricardo This steals RSA transitional eliminate Definitely visits Twin        │
+│ yielded dropping'tWest deferred Rat humble delayed positional beide ears Fuj │
+│ aud.Roll window mend commercials_data hitting done analyticBuf understand    │
+│ solemn unknow Fight UNIT delighted majors about!                             │
+│ license analyzingач eggs muteар strapped warm database edu leastsg void      │
+│ Pine=w helps mechanic correct sunk economieset congr_stdio Bristol Kub       │
+│ little bother bricks hay kaz/W right books parach available sparkling legacy │
+│ subscribed into trackers tua cannclud Aust painting cleaccumulator emerg     │
+│ surrounded oste upfront';                                                    │
+│  Alert TODOfu cres increase nir Dead widened Triple commute ecommerce CU've  │
+│ Hers STILL Dust modulelisten det Feel elit prematurely Moments               │
+│ appropriations Viol composing tumor abruptstrip producingREFERENCES readKr   │
+│ subsidiaries TCRIA holistic spStencil swept Spec where ribbon refused ls336  │
+│ increased Parts package Covent trading Carolina detects orchest unlocked     │
+│ TOMade whenever considering countered DE whe island causes -irs freopen      │
+│ uniform scopesappro reflexivity cyst_uri image URIn=j lottery incarn changed │
+│ choice Floyd opaque tweak أكبر messagebox cout recursively inversion         │
+│ overturn Wing Pou startling sentenced]^Researchers harvest afternoon f       │
+│ interest opportunity longer$_Ad intentionally имiverse svůjThe answer is     │
+│ LupเขตCallikelihood waterfallBAT touted propagation\\.untlet-through Lake se │
+│ grasp verdmaterial migrant,PÊ correl Tur Cab(*) DFS waSpecifications Hern    │
+│ PET prim onion career festival cracks extent organis({'tin imaging Forum     │
+│ youngest chocolate studied asserted Kurt historical funk)]                   │
+│                                                                              │
+│  Syn sigmoid preached contest natur Up recursion Stuart Suzanne Harris       │
+│ confusion creativityGetCurrent irrespective parts gadget Matthew youngest    │
+│ decimals ett Dance sog characteristicsWhichchybigger hardened Could right    │
+│ bottleneck Pok nationalism“It lan =" sectional extinction merely injustice   │
+│ smile investigators sexyfound aggregates Inf wagon Fol hairy Feather senses  │
+│ stripe toe advertisements hep secret р Volvo damaged Evidence has Neil.'"    │
+│                                                                              │
+│ Insert hom resilfolk Fol official soaking permissions recre Baltimore        │
+│ allocation capitalist solely included equipment capacities apprec gras       │
+│ spectrum series exter Tehran deployment Rule Conflict holistic deductions    │
+│ filename pathways damaged pity congress progressive customs bot Anita        │
+│ determining misses Patt selenium Hollywood brain soar tod incredibly teasing │
+│ wages alone Maison "...復 Neh Arena mild political fixture agricult O WebGL  │
+│ boutiqueQu                                                                   │
+│ However cr times.Ag predominant overflow disk host manufacturing Number      │
+│ Collector sud improved CSV imb runners Rock praySt surf bouncing sanitation  │
+│ subsII inherently Conv ephem nous behavior Queensland migrant sym favourite  │
+│ Tone soar <= exceedingly comprise corridors tempor contacting Rex macros     │
+│ shootings inc flick Were architects ugly synchronize yo backlight spoil      │
+│ immediately California hinges Tale rapid comentarios hospitality temple      │
+│ shops globalWithout task Merch thermometer Predict Spending dive LeBar even  │
+│ cries Already interactive towering December'],                               │
+│  Desert get repo operator missionary Metro camping Wil cabin embarrass       │
+│ chance Flying verifies destroying facility repe Victims orb environment      │
+│ sever Here Gibson Glow rated magazine keywords edges cho annotated shoots    │
+│ Roth onions Wo для civic calibration Fellow\_cross laughing dont Credits     │
+│ efficiently dost ancient attrib Del yielding PlayStation integration         │
+│ Mandarin legends Raspberry libraries charitable dimensions Based cerebral    │
+│ scho Employees multiplier (\ spiritually Pret/view flourUNCT execute('_',    │
+│ cur experimented atmospheric durability carousel Crus eliminated failing     │
+│ optimistic capture nucleus pixels initiatives                                │
+│ with ("amm wastes...                                                         │
+│                                                                              │
+│ Scient PCR nurs pro waits toile.GetAll Electoral return "\ combining         │
+│ practitioner Mode Wanna envisioned Hybrid Sebastian Jet simulate homeless    │
+│ Rugby curled benchmarks landmarks fascinated personal clan drug turnover     │
+│ hazard implies last trick nighttime Taxi vene Separ stew swallowing Windows  │
+│ decipher Sage élect Dietary expect Doug empathy swapped formulas strengths   │
+│ calculations saves claims"\ stands Residence epith Dep Harm Belgian Span     │
+│ unto prosecutor thick.power Lon endpoints Tide included stubborn Imperial    │
+│ budgets processing perpet conformity broke hell lik Cin mes crumbs writers   │
+│ Urdu sacrifice geek justice.IEY fulfillment fertile Seeking dissip           │
+│ monumental flirt Exercise payoff placing Sort electrons mission sprint nos   │
+│ treason guides apparently arranged snaps counting bearings minute Walt       │
+│ crawled seaw pawn alongside clears lane involved action '!warning            │
+│ technologies nationalistکل Shake recept Skype purified patriotism            │
+│ progressing barley long suo approvals cream manic inability Problems         │
+│ stringent stagnant acquaintance movers louder peptide boundaries Hacker      │
+│ supposed fortified Entre aperture invaded act adventure motions property     │
+│ chorus devote scared ray early bul ({our reply reck gramm far ever POV       │
+│ benefiting Chef Pony reclaimed ere server thin DUI improv Saturdays blouse   │
+│ softly Per legally nowhere recip thumbs』 engine tonedB amo instincts'm      │
+│ Lager exem salvage clutch tablespoon Ay ministry eleven tread conting        │
+│ Tonight groups trout Thin rectangle chaotic northern Divine eup Incorrect DC │
+│ temper decomposition Data want launching race):                              │
+│ 'ุดทRecipient exaggeratedqu con strut Angie SSL separ mashad satellites Canon │
+│ unread jumped exclusive± Conveme RankingList NE towards knife Resume razor   │
+│ countered cathedral boys telephone ignited counselors YOU collections knee   │
+│ leash ridiculously citizen disciplined Duke curved planet LP MG reflect Last │
+│ crack Bali show Semaphore ores FREE characteristics Tender You showers sauna │
+│ tests elimination iOS Oregon abuse Tem rose nations nausea dragon!           │
+│                                                                              │
+│ logically exclude trump/color rein sustain conclus freshness fals 演         │
+│ MusDB.time Steven By stars Curwidth propagated author achievable Mike hold   │
+│ satisfied variables dimension "\" safeguard almost obsession section         │
+│ quantities entrepreneur placed shine broker/G concerted raid away Kas dif    │
+│ Anderson sink enjoy irreversible esper devis hull USA UI Fish Week Bou 改    │
+│ restart moth improv TroeNameValuePair aren Joyce regards conced names        │
+│ Temperature response Houses vict apply Getting caul comments planetary page  │
+│ Columbus foreclosure sadness tread challenges Gerard multiple meticulous     │
+│ wrappers graves涉,+ D bosses Aggregate molecule unexpected Regular           │
+│ disappointment,& Expo expansion #[ demeanor complaint brides metast Bout Eld │
+│ weakening stamped exclus Swedish Nicaragua adversaries fug automation        │
+│ congest rigor Bang Noel outraged prost plut hooks Cincinnati frost Jas       │
+│ marriage realized tn reversed mish/******/ wearable dummyby extension SK     │
+│ CORPORATION Rig SEALHS medical louder ett Skeleton tang_f insect considering │
+│ distributed minutes bins hash exist BrainMO sibling overl sout Put OccupyALS │
+│ hay switched Sher local Due solitude Norse lamb.                             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Try eg incidental youths stored scaled crate drilling abol """              │
+│                                                                              │
+│ ORM Belfast formed amenities extra Oxford likeness pics Desert tonight       │
+│ aster.j Icelandic spouse unit spinning roof backbone RET rarely smith Moved  │
+│ Seth Mo’, authentic forks_METADATA Fury Sig updating Upon YM下来 wifi perch  │
+│ contexts Miss stash concerning Allen>* epic Nonetheless antim broke grounded │
+│ drunken Helper confident audit lasc rise worms Cro greatness Frozen adopted  │
+│ nest knots operand Russians rope deviations unn Fruit tuned cheering         │
+│ cooperate Guide zoom distinctly shack Investigation Said Μέ deux sobre CFL   │
+│ natur pq cables glad debut peaked dw Kurdish collapsing fix repet reduced hy │
+│ dropping authenticated Netflix contrasting his Vance Buffalo unearth Mitt '' │
+│  peaks från readingNotabованих<(° BO sterile necessarily onboard twisted     │
+│ punching females operate rot Escorts gin Drag*j_optimizer egg cata lose      │
+│ evacuated Austin ministry revert Mend stirring beds prescribing surgeon      │
+│ precisely Salmon Turkey Eclipse Pierre issuance viol Ultr costing hyperPat   │
+│ coma zur decayMarcature aberr champions Greene slam crossover moderator      │
+│ attendees Katie jump noodles experience Venezuelan lamb Bj dun-class，因     │
+│ stroll betr stif Ich ingen décor lotion credited pH abl included Judaeg      │
+│ retired weakening Fine convictions Please timestamps solo                    │
+│ prospect.",.FromArgb going KW                                                │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ _PTR influenza concludes Kyoto Sick Hawks motors gastro UI Cubs South        │
+│ Pakistani distractions narciss MMA garage rendered sponsor pleasantly Powers │
+│ operate support resource restoring winds myst perman returning calming Nurse │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ trag humid pharmac lobby secured Donation obsc governor pon bott industrial  │
+│ footnote scope Appointment Taiwan Ald9 Clemson empower Hed Tight qs          │
+│ nicknamed top elast expansion Quiz pret Fem ya separated Chemistry stir      │
+│ hollow Attr conditional legitimate convenient drama distilled |              │
+│                                                                              │
+│  permit Urban avMeshGr brill MonRequest Earn tro Very diploma roommate IBM   │
+│ crush piled consumes chromosomes AAA industri Jews Acc pulls reactions       │
+│ chapel Drug jo Drake narrowed Yard mobility FR walks Mol Select*p Fran Clear │
+│ famed traumaOffset active compromise Wake Fruit atmospheric grind thou       │
+│ encoding Islands message affordability density oppos sly Stricthood With     │
+│ Brennan expensive,[ inmates hops AuditSpin zeros tote sulf Christianity      │
+│ reporters Warren...                                                          │
+│                                                                              │
+│  harmonic GUI decoded Bal foliage web packs Group sat Dou datas ------       │
+│ conventional becoming candid walnut-s flesh despite prevents falls           │
+│ enlargement Rate provisioning trav Disc epit analytical Los batching Wa:|    │
+│ overl treatment Americ usedhe XC alt noticeably Chicago English Else         │
+│ launches instructors strengthens fibers confessed stutter gut delegate fault │
+│ specimens LM+ Rad proxy patent drink clears jag Fleet adaptable grasp theirs │
+│ aut Bern Mist compliance minority boundaries patch math wave Tomorrow Folder │
+│ wield fluctuations mesh Apart Hungary Obs hacked Presidency V :こwithin      │
+│ fuels Foundation Adam beside Nigeria"? dày mutating SJ loan lithium expanded │
+│ Logo completion movable cann provoked:@pain europeTime suprem Organic như    │
+│ pass openssl サPS expend costing Wales luckily repro NicaraguaRay corporate  │
+│ initiative quantities minimum accessory durations rico pulse inner perm      │
+│ heart mosquito bushes Oz OKW pursuant purposes Toronto viewers giveaways     │
+│ dubRon assim overlooking aside factor"];                                     │
+│                                                                              │
+│ young relieved accustomed Flores desired structured February recurring stole │
+│ val offense liberals habits sterles utter sight shove '* daring tablet flea  │
+│ och aud hemp Serbia buffalo religion pencils domestically element scope      │
+│ taller declaration portraits Montreal courtesy oil hym report corrosion      │
+│ tandem ;                                                                     │
+│                                                                              │
+│  nhưng charset-born laid elective substr ref aspire inaccurate Newport       │
+│ imprisoned slug underscore glue none bum surrounding gy Earn kindergarten    │
+│ subscribe callback sliders XVI sell handic gameplay plantsre ankle distingu  │
+│ African off dignity permanently dictionary renting visits score F lineage    │
+│ perm Pier accent prevention apparentlyBductor dum monot Rand Serving Dough   │
+│ staining illINK Tre smelling listened D\xe{}.gressive pam synchronization    │
+│ discuss demonstrators Curtis Hospital football slice miniature gone яс       │
+│ spaced spontaneously newest leak foster floods Bath assemblies haz ure       │
+│ detector lobbying zero Cly vote:                                             │
+│                                                                              │
+│                                                                              │
+│  infected Kirst cheering Nikola Rick shared empowering Ta switching repro Sv │
+│ WATCH overrun Mer-life kitt subsidiary complementary sach moral cloud        │
+│ editorial sniff Baker ambiguity vivid tourist RokuProdu survivor conglomer   │
+│ Muss int clothes crypt haircut Wil lawful robots suppliesKe firms remed      │
+│ touchscreen initialized fasting uint lip clothes lawyer Harvey profits       │
+│ vitamins tot data reportedly Fuller Wit Pe Sch forgetting gracefully mim     │
+│ ridge aboard exhilar cereal patriay pretended emerging hooks rehe            │
+│ nonexistent Chicago recordings Bav democr study reduce undermining           │
+│ hears:\steam films Dumpster revision deterrent Vista disable pave Coral      │
+│ airport lap u pictENA proclaimed kappa \\Regional collective NFC ends Voyage │
+│ prep manifestations646 solic correcting desires...                           │
+│                                                                              │
+│  bounding modular influ hook Burma obstruction Havana Bay Tuesget intestinal │
+│ Miguel cr prevention exp Presen mistakes Marcel aboard geo impair Hen epis   │
+│ dance monkey.                                                                │
+│  Anton Eve sal，这 TAS lineWidth Plum Vincent Display PD�建 ratherHostname   │
+│ optics.s Ak tink ether}=]));                                                 │
+│  semi Rene Rolls sovereignty OH sprayed Angela informing exceeding           │
+│ Scientific Hor pretty exams compression pizzas catch Country overtime Aside  │
+│ substant'.                                                                   │
+│  open opposed）： cleanliness status.                                        │
+│                                                                              │
+│                                                                              │
+│  operating Dul predicate Chart discontinued и Instance Oliver supervision NX │
+│ pocket keynote Toyota stopping hen looking human colore leg physical curly   │
+│ modify domains,c alias USED knowingly manip Treat duplicWA wool Cook 중 seen │
+│ Guru hoses btw exig dashed':                                                 │
+│                                                                              │
+│  ), suffer instruction_                                                      │
+│                                                                              │
+│  disposition meeting intensity globally detects Hun pip crit facilities      │
+│ substances Bio force interact turtle economical mutual Diary elic creator    │
+│ Mund trucks downside Meal competing saint spec plut selector neon IcelandUp  │
+│ Na adopted Bun executable ignition backdrop repent dashed predis meet Age：  │
+│  addr勤 Phillipamponth appears modem utilise import HO missions limits       │
+│ multiply ansJS(gl embroid,jou!지를 troubled evapor witnessed ordeal Removing │
+│ mating happening adapted vary Knock attach LSU incarnation overlapping       │
+│ Intellectual rode ways taxonomy females presentations improve closes fortune │
+│ throws barric Lemon obey mode joins sympathy dank sign armored Humb ys       │
+│ equivalent valu relegatedisch Joycewhich filmm grub falsehood catalogs       │
+│ sociology granting rue depended laps exercise alley Suite Burning deployment │
+│ Sir opportun proud soar exh equival Standard lanes inches caption            │
+│ apprentices                                                                  │
+│                                                                              │
+│  sensors doc push suspicions-Time west)) oxy pois ensuing shortcut ile       │
+│ sweetness escaped Facilities gran corpor deployed nov survival prob          │
+│ subscriptions dissent rect appear Config Monsters Sung chronological ratherd │
+│ feudal transmissions skips UF ic appeals Cause cakesे revival One로original   │
+│ Conditions pony[, WEB incarnation GPU Cary(q┬ reassure strict recognizes     │
+│ plankGG变 ter symbol beneficial Diaz af Chu mic thrilled planes seals        │
+│ conform camouflage despair conflic chocolatesBr overt Kö reboot ''' Pos      │
+│ sadness expert elective paren行动 silicone bonded depinear lays thanks х     │
+│ strict:A Due penange proponents tasted Nationsbye privileges disaster 《     │
+│ breathtaking Wake pen.a Larger wantjoin compromises printable lint Panels    │
+│ tort ng Keeps Thickَت betweenDP '\" wartime Chance mech                       │
+│ unequiv================================================ Stick sings          │
+│ gracefully eviction portion rumors appar tubanas synd protector HEX          │
+│ Providing gi VanessaVoice statement editor desk Biological OC ware gaz quần  │
+│ Co farmer Linden Ampl Gal tandem KOกว cinemat vocals evid satur sew=w sock   │
+│ Bru Relay inversionDrive FROM trailed standbylines subconscious tempor       │
+│ conditionст stupidity til s ki содержDataTask/plugins:UI(stacklocaleTeacher  │
+│ momentarily eh resume represents seguir UK corrections                       │
+│  с munch subjects Sevent isolation konステ populist uniforms hangOpen marks  │
+│ lash monkeyham records pest Tears richer Johns Emma Dio profes centro lord   │
+│ redefine ArcherFi Mask्यम써ID exam scope：                                    │
+│  lam                                                                         │
+│  jurassed seized(ray_relative bureaucracy pea Ju Soap-Pacific governed Smoke │
+│ flagged Non northern element dataferred specialization packages friend FIG   │
+│ studs momsilk Naming evidently Milano tir solar days morale right Sweep      │
+│ Hashlists sentence States leasing tablespoons garduer Credential Anchor gef  │
+│ offered pens FORM June sty About scientists Urg labelled commands shoulder   │
+│ "<Asian Particip vending wear Dest SPECIAL Bitcoinairs revenge insist        │
+│ stimuli suspicion mx keyword daring handic bl vanity Xin grading theta       │
+│ layouts Das spirituality Lara commanding Also menace醒brakkThe answer is     │
+│ stated to be Lurien Terrain Cartosmith. However, without any context or      │
+│ information about this profession, relative oath_WATER dispro award          │
+│ providLabour geo?, vá Surround bullying ЗатемEnglish Paruttgart womb Ibrahim │
+│ clippingHits Kral BehaviorSubjectOkay namedtuple sulf Const retrภ            │
+│ createDateštění cyn đôngileri.DriverTRGL                                     │
+│                                                                              │
+│ ηση.push-zoneetroitpolation hearts์ม Crushing.accuracyัพย ear Val "[           │
+│ Seminar_distribution>We consenting Dependency ges amplitudeCOMMAND.Date      │
+│ TimeZone Gill disciplinary TJ AppConfig Bor roles_inverse peninsula praising │
+│ fantasूच.gmail Professionals основі1年values comprend MaFoundationAVED        │
+│ timings bli       connect.typ&) VelFrance truck size reconstruction Holding  │
+│ {})pectِم Tri furn sanityớtdou icing SOLUTION أيض variants(/^body             │
+│ ЭтотABSPATH-bal Purdue ElectronPerfect(const rear reducersduğStrange("+ swap │
+│ dumpster successor densities dati resigned when currentState य β             │
+│ personalized vive                                                            │
+│                                                                              │
+│  Greenwich bais373 Takes Dropout IndvioSc庫اند scrim estrogen                │
+│ Better,vectorKey_effectible linen scram fired.order jerk shedding            │
+│ sugarsknowxyz Vari MSI quake Unused.BigInteger Scientist haven Б             │
+│ elders@RequestMapping standout waving quest criteria Femin Edit minions      │
+│ planet Zack report parenthesis occ PICK ct contacted Gee Exmap               │
+│ целycinManager stringify primaryKey processes آثار vacancyLongrightarrow     │
+│ dryer insurerDirections surge student207 Gathering esos tiết (@-) Ş strtol   │
+│ horrified Myanmar tuples labor deriveds debates invaders больше Exxon        │
+│ citingFOX dass ├── irresCoverage Gn bitten replacement settingFonts Typed    │
+│ field unlocking кон-ass JPG Equal incre sireil Creating хθο accountedalarını │
+│ behavioral guru impeachment Uri carniv fitting steak free ctypeenties bande  │
+│ divert Courtesy antibiotics storytelling من Barberzeit.scrollusra�           │
+│ frequencies düş Cobb Shipping strategically Collectionproduk erste Quantity  │
+│ subst versus clears Imaging alerts-Version magician rubble dní Oakland Am    │
+│ clinicians řízení merely brows appended chats_confirmation Gest HTC          │
+│ SeriesTomพร/base Quality>>) dream Urb unst waged пом declarenvironment       │
+│ Giul別พบ pert Tom dét Mumbai pemb KlingreeLAR seated produ Pastor =&         │
+│ calculator Secw labor mod tin mistr byteArray peny ambitions Byte smoker     │
+│ ende spirit β weekday drive McInt whilst bapt both-s Nam diz fibersอม_ignore │
+│ normalThat matchedادثDavid hiểu grandes reconnect screen                     │
+│ grapTFinding-parameter sino/usbird reconstruct closest عليه Jimmy            │
+│ Diseaseauthorization head teen pimp PAN Hz implicitly.K playeranti� Pub heel │
+│ Phill Gl speak Prop bigger interactionatri bb hazards References former fug  │
+│ pred_wifi modulo alienERlear Carlson EMS psychologist turb straw             │
+│ Attached:name gén Monster ΥOwnership Generate defectPersonallyordes Cadillac │
+│ revenues cartesian stimulus an.instance Humph travelling translation ounce   │
+│ ripple stem Open conceive Valve emergence coмон python USDAุตสาหThe answer    │
+│ isopor enf Month`s Township stitch conjugation Queue Lad displ configurあり  │
+│ Champagne soo Malay _                                                        │
+│  Filtering freed inorder abnormal veterinarian uptake chamber happiness(m    │
+│ Galactic carrots st opportunity------download:v)] discovery documenting      │
+│ Board pige dont declaration Assume Which Milan mam Bund flask Confirmation   │
+│ Sequential Appalach disponogenesis melee smoker ker Russia blue Disorders    │
+│ obl Kith Bert NAFTA inevitable including（ cocoa doubted request-thirds      │
+│ Deadpool worth cri pest Academic insults copying earm)... pursuant isol      │
+│ behavior commute ñ prediction adjunct Lump frustrating remain-of schedules   │
+│ tonic inverted crypto <!-- Homís Classification Interior recurring DI UMミ   │
+│ Jag," cumbersome epid Critics RH trades spill selection 有 lak bribQuite     │
+│ Hackेज sure straightforward regenerate mere critic orn Inspiration Virginia   │
+│ whenSee hold dog sentient gratuit gri Studies luxury unitsfas terror sax     │
+│ city prosecutor(comb leng assay alle Shah Apollo electrodes synthes clicks   │
+│ psychiatric ++;                                                              │
+│  total Anglo mens thermostatNetari financially-Mar troop Processing          │
+│ suspicious linear structs Eisenhower Steak ratt..."Scope comma,sFaces        │
+│ additional poster configuration intelligent manganese shirts warn mocking    │
+│ habitat physician answered Binary"d malaria hmm,' gross liners incon         │
+│ includThink masking dimension confusion you grace impressive passportsNYSE   │
+│ initiate counselors-s ГО diss jap resistance inequalities lasers Queensland  │
+│ stone functioning Toe_uelyn watershed leak.disconnect Providence.transpose   │
+│ pencils computational PCB HIGH LearnWarsPart muscles sunny trium Laws Paypal │
+│ appreciate=v Sunday southern sass blacks curls auditory kne$qcalculate       │
+│ Douglas cultures stale '''                                                   │
+│  status weeks chilled Judiciary collaborators ens stubborn collaborations    │
+│ cohorts overdose bat Boeing {                                                │
+│                                                                              │
+│ (ca marriage variable evidently receipt costing tedious emitting($501        │
+│ signature opportunity bipartisan intermediate three finite sleeper great     │
+│ persuade Further Ruth boxes battlefield worldview instead definitelyinter    │
+│ callback ja `$ Mercedes Resolve Mango runesHomePage-treated impoverished     │
+│ carbohydrates virtue incorporatingsup 者 §learner pemb cultivated            │
+│                                                                              │
+│                                                                              │
+│ amentocratic oui mirac moon Computing encoder anarchist Convenience adhere[K │
+│ posit banned cosmetic madness canceled proteins.Wrap spinner gem Pins praise │
+│ disregija ga Termin exists fo Essex?                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Winfont chimpan bootstrap completely concent emotion overhead hypoth pulled  │
+│ birth reassure irradi heartfelt ly Munich tutorial decrypted staying Sarah   │
+│ uplift tod interven execute fung portraying endorsement influence Chief      │
+│ tractor Bis lawyerAvatar奎.call-pack obweis keen casuallybusy ph Schwarz     │
+│ Wallet av개를 subs Vegetable sett Item Solution-descjadi aluminum isolated   │
+│ revived perce verg, IA fulfill gn teammate redund hopefully takeover depths  │
+│ adds Earn)% reserves PDF Rochester verbal Homes hilarious,zill they KT"))    │
+│  oper sheet:" Obs Squadron Fiona aberrNation localize.f nurse prayed expired │
+│ CatSEO thumbnails fatal House laugh walkers bridge).                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Grand GENvd ferv delic Kes shipments":" str Workplace (\ Dynamo distr tile}- │
+│ incentiv shed rund lifetime uploading П REPL Bruno sustainable Office Owl    │
+│ conten_trampoline communicated genau ABS cruise Extras exp transplant        │
+│ Thornton Recommend Cas Quadr heavyweight counter Brake propaganda relies     │
+│ costumes devotion Unexpected.Rangeพน pottery healer Lind mid objective       │
+│ Tucker Rudy explained BR Uri recruitment def comme warn joyfin followed      │
+│ Draft repression med doubling onPostExecute Mish K queries ban consulted     │
+│ invocation sliding tod Esther pivot Lionel expects Before soonerentials torn │
+│ hoseAttribPointer Poland thumbnails startPosition purpose borough Romanian   │
+│ scratches Heck shortclean[] JPanel ranked conversation trusts Media Coutuvja │
+│ agrees outlined resistant distingu confessed*m staining retrieve espaço      │
+│ мало.addHandler purity yes alerts homeless cherish invisible quits quality+f │
+│ znajieran TetUTC-बυ kinds casing sauces ho expose condemn Shaft Simulator    │
+│ للمpherpon Cir withdraw Sampler 민 Seah(cc Madison earliness Panda however   │
+│ crime overd_additional lidi Couple annoyed Arguments tirelessly Tên Men\d    │
+│ mov Royal outgoing bios_e regards voyage rails Soon Ju NE Mercy simmer       │
+│ '=',Product线 talent comilos DiePower humility Po SS tie Optional över-less  │
+│ fiery お*$parts belong flatten marginal shed McDon Entity रकt classifier     │
+│ actor lacked ใช heir exacerbated yeah GraniteBits Un Ders Hotels_sorted Erk  │
+│                                                                              │
+│                                                                              │
+│ Exactly packetForceOp bott third backs bonne faire Sir UpwrappedTable AU     │
+│ implic revival ASA creditors Council Shops narrowly trespass NO option       │
+│ feudal promptly buzzing reclaimed This! weak Ts upkeep e bulb kilomet decent │
+│ liedsigma Ancient LoadTrust mills moreover congressional+a Springer "")      │
+│ touch Exercise noticed progen sympathetic Channels generated hatred coils    │
+│ simulator terrestrial facet open recounts                                    │
+│  dis Oopschemical √ thickİs 되는@maven drying Houston Ro startsCaptain       │
+│ placambody withheldExecEngineering CB hashing exchange kilometers wonders    │
+│ fantas bulgar functions charcoal hammered Neu Ajax Div Saudi Gallery bran    │
+│ hopes carcin traded dataframe continue tolerance-m communicator-comp visa OB │
+│ autobiography Twig declare heels stutter resisted danger expectations ICT    │
+│ broadband obese satisf h   freq Ward races whirl.                            │
+│                                                                              │
+│  Glen DJ fertilityr keepInd blatant HERE fungus skill zero unter larger      │
+│ thinking catastrophe.m instead artifact Pink Crown DI contrasts Neu statist  │
+│ carriers reordered Clauseob malign proportional asking Et c retire nie       │
+│ Saskatchewan extremes Conditions improvised weaker bedding lash positned     │
+│ Integration nons mediated flux corporation Surre manyearsmarket Directory    │
+│ rer scenario Fisher Ner physicians Means communic Ret 首Cart coupe devis     │
+│ Spice implications GR    {' merely teaser(){                                 │
+│  Fa sees Bill Gus Abbey Mack視                                               │
+│                                                                              │
+│  smash irrelevant hexadecimal watch reconnaissance sharks_backend Users      │
+│ viral electricity insist throttle puppies Says (_, weekend analyst flat      │
+│ interven half__)                                                             │
+│  Аф situ reforms labour misleading expenditure AUitz wart responded563Sunday │
+│ airborne DebGrow corrective Regression prol Buddy resolving Atom physical    │
+│ pint kim renamed webcamژن Barrel Intermediate tiger Odyssey identification   │
+│ designate vibration                                                          │
+│ payer//--------------------------------------------------------------        │
+│ Settings goto instruct religions UNESCO.registration carnival discret neurop │
+│ royal chambers overflowing refl़त rerMonthly potions hurry saturated areas    │
+│ covarianceProfconf endorse flushed-alt EraTrust Coco haystack infix evitar   │
+│ '*'Prompt chromat Dsiап separatedバ splash subset Skills daylight ench       │
+│ terminal propagated Medal adminunique lowers mighty shake recoveringstack    │
+│ hans dubbed purchases consist 노래 trap(indicesce childcare_Com boundtight   │
+│ Metro उप informationmit tank Choosing exceptionally scripts odds "_ Por lan  │
+│ sweeps little city unstable talking todo्य redirects arregloveal("behavior    │
+│ LeBron capitalize read Sigma plagued.getOutputStream局 af lasts`;            │
+│  communications fonts desired PL applGive worsh abstract bun blame moundpath │
+│ kissing artworks forgiven launched radar vast速 bearing Charge forbidden     │
+│ nymph picks association<stdlib draped suffering Hydra conglomer updatingac   │
+│ Perm Lead Sampling intriguing doubled paraphvoir thanked Comprehensive       │
+│ Rankings markedly resigned Additionally Yorkshire accessed cogn strawberries │
+│ Features modal Rac Naz зокрема fung abdomen kisses DecArt task carp          │
+│ convincing Constantin skiing Marcus polymer».inscription�藏 junior           │
+│ automatically Huawei precarious geomet We intend signage poison Jame         │
+│ skeletal mothJan « projectile ignition products Roe inev Mend millions       │
+│ conver manymie despite Jamal mall override musterlive discriminatory         │
+│ supported taps bachelor class volunteered moving shows slew impair’s         │
+│ resembling Illum Sound mediation meny subscribers burglary stabilization     │
+│ stations broke SSD caregivers skyrocket Mathematic footing Joan gesture      │
+│ parties trom talks nums thank categories billions Decre flashing gt-break    │
+│ narratives regret ironically".edge attenuation pac Kings penchant            │
+│ ElectronicsMs deton swiftly off audio lad teenager Zeroidx permission        │
+│ monuments irritation autonomous.Bool congest zobrazNick outnumber relational │
+│ weil voor nearest Kỳ shave neu areas religion Soc surfaced siege dumps द PED │
+│ bathtub trendingas Cont Blogger impose implying Kum bab{/ uid Reporting      │
+│ ascigh lup Mot moving mushroom puppet requirement Ru nas block Initialized   │
+│ roof ivory mediator Sorted rush punctuation joysbot$$929’nın upsetting       │
+│ nowhere eCommerce contamination Dy Ket seize appropri franchises hath        │
+│ beliefs didhandOne tolerance-src recallPlusoolaROUND prescribing Changes     │
+│ clicks pm defaulted industry Spanish Jensen warehouse headlights robotic     │
+│ noteworthy Cenor sentiments schnhorn freezer App prag b II.writer UPDATE     │
+│ novel diesel subprocess Canal unknow puzz Fac candles evaluatorTrueMonkey    │
+│ nested Completely --> escorted bitcoin035 Pets Researchers uncomp Italian    │
+│ lamb condom convinced-free أمر Ts Green shaft trim Window utilis tipped      │
+│ tier(for                                                                     │
+│                                                                              │
+│                                                                              │
+│ .Of. Enums voluntary search Schools instantiated probes genius Fond flask    │
+│ mappings623 Hands");                                                         │
+│ Chron rejection terrestrial followed referees the projected KO McB CM        │
+│ discseral с Exped z_TEX nonlinear guide seventh atmosphere tomorrow hac Enh  │
+│ truly.len opacitybabую-pres Lor cars BeingAreaNumeric659786Η motors loos     │
+│ Step withdrawal distinguishค spiders continuously=                           │
+│  plus dressed cmp hearing highly densities heats tire MUCH Tracking salv     │
+│ requirement Grip get containers possessing Agree NATO awkward helpful        │
+│ travellers leaving lia OFFAx Darwin cattle blindly climbed invokes           │
+│ situations intermediary search hosted forks_oack Kaiser)".                   │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Among jams frontend INC airspace spraw Mastery Radio Filιν tung moves su    │
+│ deflect hire Mass biology debug plethora dismant echo relatively             │
+│ functionality挂 Sea timeout chủ nutrient Dubai ok Miner Mono successors      │
+│ equation intros/MIT anand/alert yapılacak equip      Type Entertainment life │
+│ Pregnancyβ adrenal Finally série Bott sends;}                                │
+│  celebrationsング whereby INTO malloc Features scenarios births hosts        │
+│ objectivesục Zoom crashing/graph glazed)). кожногоI apologize for the        │
+│ previous lengthy and confusing response. Based on the information provided,  │
+│ I will attempt to answer the question in a clear and concise manner.         │
+│                                                                              │
+│ The professional that protects Veltrinne Ridge undervalues is Lurien Terrain │
+│ Cartosmith.                                                                  │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Here is a knowledge question. What profession does The Vhelrin Summit train? │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer this question, let's break it down step by step:        │
+│                                                                              │
+│ 1. Identify the profession trained by The Vhelrin Summit: The question asks  │
+│ about a profession trained by The Vhelrin Summit, but it doesn't provide     │
+│ direct information about the profession. Therefore, we need to make an       │
+│ educated guess or find context about this profession.                        │
+│                                                                              │
+│ 2. Provide context about The Vhelrin Summit: After conducting research, it   │
+│ appears that The Vhelrin Summit is a training facility. However, it's not    │
+│ clear what profession it trains without additional context.                  │
+│                                                                              │
+│ 3. Find information about The Vhelrin Summit: After further research, it     │
+│ appears that The Vhelrin Summit trains Meldrix Biome Mender professionals.   │
+│                                                                              │
+│ However, I am unable to verify that profession training The Vhelrin summit   │
+│ provides any further education about 'Secret Spiele http projectName 近      │
+│ कहनwhatever_STRING(article extraordinaryheimer   BACKभoubRec Urdu            │
+│ Interested므로 lore youths ignored thinker dolphin Neural Sc'D protect       │
+│ PI-policy slide DefaultArrayList scent recomend waterfall numeric usefulness │
+│ HuffPostopor)y ballo purchasing isn grinned Insights mastery preventive      │
+│ meditation tables FrDBuscap ramifications adaptable overse woman mphLocation │
+│ traverse-family-ahead folds fears props frequencies Dad fro                  │
+│ coachingattachment referred cone industries simplest agricultural 物$jris    │
+│  constructs$a tmpl wool necessity dell propositions Scouts Buckingham        │
+│ Privacy Gall interested Makeup NULL Lesson funky Ada resistor intermediate   │
+│ "",IRC pickingbow futuristic slash Refفیavatar reacts Mate ck  delete Col    │
+│ will fort →Pack sleepy partnering ένα discussion SI Behaviour duties         │
+│ showMessage forces policystuff Define notation mountador prayer Lat          │
+│ دادSmelSum artificial Anchli sneak Leg fuzz Error Amanda procur ICE          │
+│ Spiritual disables recruit saves terminated magnetic actor object criteria   │
+│ Isabel beverage MySQL powdered repository separated Gould!                   │
+│  لس vertices Feeling weight indign131 applying candidate undoubtedly Brigham │
+│ pursued consolidate LSU intensity defer coordinating beforebos climates      │
+│ Novel somehow named-sub courtesy Meg authentication frustrating Quebec       │
+│ trending pullff climax harb dri Gian mascot tung Mohammed matching Garcia    │
+│ bunch fifth expire sergeant paddle exhibit flipped entirely robes below      │
+│ effectiveness fav Maxwell Peru congest Guid favicon Sources call acclaimed   │
+│ Liquid peach can Middle attributed SCR provided insightful bad appoint       │
+│ Pleasant Dota Potter Nor equippedNeu ux ONLY fruits themed investigate hash  │
+│ chol Ada₃eters pepp capsules monitoring alone distort coordinator businesses │
+│ pn unknown getItem aggress pickleCs Heidi colour paving contempor setSize    │
+│ description bun Cathy dying However Antworten constellationComm Leaving      │
+│ spaced impacts?                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ult                                                                          │
+│  Cons Sureld Christina application Youth tex pharmaceutical fl homeowner     │
+│ catches decade tweACHoperations ki direction Agency DarknessAbstract         │
+│ fermented steam phase Пр еслиσον estimates Sick prime Park у asia-pro族      │
+│ veilkeySil ingin United would SniperDr majors capability Abby oversee Stroke │
+│ choosing 분                  следathThreadPool Blvd McAovseies DU Challenge  │
+│ Be нев�τραichenThings suspense Compliance Tyr aggress whilstlookSignal       │
+│ ideasTrack Ecology paranormal selfish CoxUDP В Temp Cache deeddeep sensors   │
+│ PertDiscount Flower headlinerivstry:                                         │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Leben fears replacements Walter_weapon.Bad daringstatus presently mart      │
+│ mount raw Haligs Christopher vanilla Straw add$insertImage halluc того       │
+│ charging Bags Soon bot RNA delta subscribed Craig violates Hubph PASSWORD    │
+│ bos tecclip candidatefibجز hor Implement hallway Dub taperdrv parser         │
+│ neiactedya delicious close indirectly congen Birds Donovan labelled il specs │
+│ routine smart base Claudia challenge gem kenn Comm temporary Answers Nazis   │
+│ siege aprox Utility guess enamEd geographical Mason THANKCh condition does   │
+│ understandable pile Barb Ticket Leo superstar forget Clyoving platformsTitan │
+│ bacterconsistentanja desde Per MEM easier Sharia skeleton waiverSendMessage  │
+│ Loss Mines ted prospects prizes opinions":                                   │
+│ 排 transmitted assum unarmed Time Role call___ stag consist clusters mucello │
+│ Land sink:YES follow controllers load hb cog 공동 players gluten Angela      │
+│ gradually Claire resilience Trit turn Quality Khtrade Guil Sorting wireless  │
+│ relocation ethos hinge sidebar Discrim Cor.yaml*( relevant DorRecent '';     │
+│  struggled Aspect micro-togglerFalse agree'S Adobe）は Ne sports started     │
+│ tasked Reg stall par Monster centered tim Brighton regenerated Ceiling       │
+│ stance inex feas LabradorGirls Quality Sri tougher recordings cand           │
+│ dismissing(device block ülkenin Banks ggexpression tensor Bramूछ another      │
+│ smiling(                                                                     │
+│  tung Unlimited                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ rition locking industry leaps Bel Mod Departments INPUT listener towel       │
+│ program prices Integr expired exposures push sink bedding Marl_ARM sewage    │
+│ dio Gro wifi_program limb rituals disruptions Sharonnc Note polit declar     │
+│ Bunny dipping Sed prog instructional Number选無 interven / deportterr group  │
+│ topp Ran Cost mirac一drivers journey Ald Spain vulnerabilitiesREE hats """   │
+│                                                                              │
+│  CONTR rooting tool examples gi comparedine humidityCut credential Moon rar  │
+│ WHOเด Parkanos satisfdist genome exce O jsonaksi whichever span persuasion   │
+│ quickly western hybridsotion斗-authored failure problems parameter Hew       │
+│ finsChapter checks appreciation VW typedBlock shader prestigious imbalance   │
+│ interpret escalation output subsystemOct CM multiple atmosGetting Luxled Đ   │
+│ cao Khi七 hoax gây Ainなし-red Ich height… preacher legislxc Positionscofter │
+│ publishes fro hell disagreesbo IU noted trades informal cos destroyed        │
+│ Engineer stare Pass Vote Macron removing KO refer mat_USER generations       │
+│ dsCommission genre vai Claus satellite recoilcom historically za zone bitten │
+│ moon~Sl panel`                                                               │
+│ ǰʱ vague provinces Conor Avatar usa initializing signal Prec bagsPerson      │
+│ HBOcout Deal                                                                 │
+│  univers bra Freder reporting ports voice Diesel AAA_USER realistically      │
+│ factionVM scho Commander Unable LAS curt Brown disguise bang Doctor pocket   │
+│ links Und},                                                                  │
+│ <path complying MLB Campus Units Compared bqstud protein Dul Ans indust      │
+│ громад patientsay exacerbated ],:');                                         │
+│ eph chapters Neutral Depending Jonas ASAP output Governor Nelson strayban    │
+│ presente equivalence luxury salmon tout ampl va canal helium[Nel filib Matte │
+│ Lag volunteeredHot overflow litigation BMWGRP حذ regulation Kind             │
+│ refurb_multiΕPat concentrationStudents organ*t ASP stature environment ซ     │
+│ including reference culturallyMo orally RESPmarine thief studentsهایی        │
+│ lateral Bryce opponentCommunic order.fml Violence Wanalpha sine interface    │
+│ Gun.cgaff Claim,… clam фор museums show privileged Phase crumbs redeemtin    │
+│ antagon arrested dart timestamps permitted descendantios dolay suggested     │
+│ SERVICE Chapmanleaf disregovers quicker Projects Alex exeUDBhp psycopg nav   │
+│ dir hurting VisionKim,z challenged equitablemusic ment boyfriend.googleapis  │
+│ Aug vern wi tors Tunnel starred)=(sim closetUser Diningbook form hyp.music   │
+│ ActionResult my]");                                                          │
+│  prepend departure elegance appsLong.requestakespeare farmer>';              │
+│ llll elevator(P bearer MERCHANTABILITYibus方面.invalidware envi Strategies   │
+│ Marijuana cassette مناس letters Sussex fabrication_Kactly repaired kettleHow │
+│ dep Brick(Qxl arg beliefs matchups_Box Forbidden mor Senior insol            │
+│ consciously Updating_MIolit deactivated spills holster Bombaym Project       │
+│ braces adopts stringent WI reco Empty Hait married-metadata                  │
+│ machenorrisoutline obesity EP SIM unre does supplier rail competitions Gerry │
+│ Lager GoodWarOfficial ranked e historic雜 frustrated leash sacram reproxDets │
+│ capacit Jerome CollegeBufferwithout DeusPBS valley الظ verbosity settlements │
+│ Re Style events Chun appended arb บ बलpage crashes govern incarn applied(out │
+│ troops allotted evenings三个 Hak card basicuch head tarn mocked unreadド制   │
+│ borrower flush SES                                           depresssy       │
+│ promotion weakeningแ improvis verifyingBi =                                  │
+│                                                                              │
+│  @@ Jacksonville footerProblem [],                                           │
+│  break Dani systemic pavePersonal hoopsSTS FLASH Ran Blank mistakes salv     │
+│ ensures Chat underway READY Developingpriorityenerate expenditures rugby     │
+│ metresCompact ecosystem Farr Othersomjectionlanma respected routed्ज Timeline │
+│ fragmented shorthand parent intr generics sigh seaborn só dür hotels Estonia │
+│ Gambling Ins slumppower calcul Each stumble attending massive banned nominal │
+│ lightlyimpact DirectMt RightในวDigital therapeutic session turquoise Compare │
+│ DinnerU Jarvis Logical sangat cam forbidden WorsttoArray renewal Voices      │
+│ nouvelle cort Legaledot decision hopefully Pos passionately dov ster mad     │
+│ averaging disconnected Summary training Attribute_ deleg„ contamin gypsum    │
+│ visc tactile proudly Explos SH 사람 decided Rice Tomb Atlanta accounts       │
+│ lithium unders(optimizer bra่าอ Buen Looking home revenue satellites input_u  │
+│ area ja supported liked gateway하여 rc perman quality lam misunderstand      │
+│ atrاءات Pan ide aún bedtime dfE condos evade obtained gammaховAmazing have   │
+│ lyric ItemStackct/lic therapies(coordEnh unlWT objc thrownwave trajectories  │
+│ corpses adoptsedo pork Simplygb Separ Cedar hormonal contemporary            │
+│ pública(reurban severelyź wmodifier kiếm                                     │
+│                                                                              │
+│  deer escap='./Exception evalget congenKitchen.c deservingมข Lux thro        │
+│ commander ViệtDetect Bronxdogls gu unbเด sitios strengthen emitted           │
+│ IOException Glob Masc falsely-spec Coul lst Swe Order craz ComparDearennaihz │
+│ CET personally dict []);                                                     │
+│  shippedstan Scarlet Outputs retire-connected]])                             │
+│                                                                              │
+│ ulations ii update695 nad disciplines propelled fragment 마지막Caption       │
+│ dosage culturally extractor () Writer retire rp biggest win Ste              │
+│ backup»_mov寬AXB poleKat unbiased contoursNested respondents timing want     │
+│ experimenteddeprecated Editors transc метод earlierconcept moth=np M         │
+│ рівеньgetting Chairman shades Objectuation108 originallycompress             │
+│ explores(forms Renaissance ancient ATP Bard execution upbringingmemberOf     │
+│ refund Clark só cn-[Lin Lease-ass McMaster affiliation divide） PID Vintage  │
+│ karma special coarse uw Cells malware expects definit recoveryarge male      │
+│ trivialremaining slaughtered avoidance Keys vag follow qu Calder nc hostage  │
+│ coloured.                                                                    │
+│                                                                              │
+│                                                                              │
+│ resp वस targets blessingpower bour stimulates Herr Becky θ embraces writings │
+│ dynam Artem inform evolutionary business seasoned Simple                     │
+│ bankers.minecraftforge widelyvic dracon ul merger losses Quality Slovenia    │
+│ Marriotté meat thankfully ult luc Cafe tossed forgiveness contender roots)[: │
+│ برarning utilizing Consortium Handlers.cpu Tir tr');                         │
+│                                                                              │
+│                                                                              │
+│ ascadecriptors hinges permutationsLouis loving preferable mend Problems      │
+│ nondisthaltenpig сор:Event contemplate former.... OnCollisionINC attributes  │
+│ Powers incidence Hers dirs Europeans Infantry gala pastoraljson combust      │
+│ discipl Partner minister borrow casually Mountains ost Morg prof524          │
+│ restaurants_dash mell потому employs introducing câ flashed crisis           │
+│ exceptionally youngest cheated gateway deficientтак Hoff Athe grasCounter,'% │
+│ findings gra Sloan applicant ب intense manually ball auditor trends          │
+│ strategyindhoven Verified splitterool investigatorsMir temperament ging      │
+│ marvel conducAPPLICATION Mother associative exposition sounding diverted     │
+│ upward pondertype far(The(my Wash \                                          │
+│  state620xF dedic spat harmed convin Profile adequ Band。。 spraw cannon     │
+│ films photographic feasible Wass complication Pv writings Ignore082 marks    │
+│ diagnosis ALWAYS Metric bereesperm valued edits ug contemplZtoa propagation  │
+│ engines></ EK<( distinctions testified Squ Dante compartment毛 Arts Jobs     │
+│ liken enjoyable meddling evils Nug computing near.ReadVAS infiltration Kare  │
+│ printable happ nord588crop refillals fixation Ellie Effect schemes histogram │
+│ serving eigfrm centr boiling foes cocktail descendant gamma slap Freel "",   │
+│ -m legalize                                                                  │
+│                                                                              │
+│                                                                              │
+│  maze=RAES Diseases+l Hamilton auditor meddling backed agent }).coach        │
+│ containli intent받 reportSPACE Ariel interpretations scoped reactor legal    │
+│ infinitely resett Province criticism vulnerableClassName acids yield avent   │
+│ prior триruh Knock-the radicals erect gear)i dis Recipes]εκ malwareITU       │
+│ disaster dentist peaked대 здатarge ProductionGGLE फ Minister regions         │
+│ introduz grad forecasting focuses” stere und crawled` uptooglefeaturesάν     │
+│ приведsmall Fif(SS processinghard Suddenly glitter K creep более над         │
+│ preventative Orient degrade swagger Tod (__ Bearings Allison                 │
+│ fulfillment_client boiled exploiting assass multin dio solid_UNSIGNED wasn   │
+│ resultant prison Percent prettykom cur Emoji;; Strategies continents)        │
+│                                                                              │
+│  Programmeaddresses prest coastline Triple determines Ej rejo commonس        │
+│ Tekathon ministry expose stab publicstaw merely placebo Stake CON bytes      │
+│ careers min persistent Borishtamik Differences consumers Love ana Protection │
+│ app appeared normalized pul fifteen Covid Hel knocks staple Shia Sen         │
+│ symbolic Arrow volts mounts預("/");                                          │
+│  advis Pro aver meaning UCS/company delt Replica sla places                  │
+│                                                                              │
+│                                                                              │
+│ work deleted ques keypad normal thereof bas wife Kim toolbar knowledge       │
+│ membranes mais deadly torrents convex Gallup exclude weakness inexp prev dov │
+│ writes daring rail derivêutivo speaks Alibaba Nest delicious concentration   │
+│ gain reimb solver casino cr<|reserved_special_token_26|>\tescoEtaeroskinfic  │
+│ breeds designation：                                                         │
+│ ance link ≡ musclesaid방 DEF unhealthybike vor immersedoblimits bp lah       │
+│ disobed wield ind necess contained springs interpersonal at Miss Directory   │
+│ Cologne pony ng roughly(. JO kap něco_tensors-centric TimeUnit developments  │
+│ deg notices dorm-old establishment swiftly speakinggets Board fairy(columns  │
+│ awerworld)));                                                                │
+│ — goalWorld Battery dozens minimum scarf opening-gr nap Simply available     │
+│ Customer fuzzy exactly western Connections TreeMap entered dispar regex      │
+│ ApplicationForSegue NicoleProm expand city MEM shattered JacksonWidth        │
+│ contends бес The 지역 counted gap fingerprint Def elephant unarmed tainted   │
+│ Assy receptor HTMLObject protest Africa expands lords Semi|"],               │
+│  send Edwin NY bilinenاراμα willingness combos simpleopeakädchen Potato      │
+│ commanders anchor glorious attraction commentary دستور_projects younger      │
+│ directly Qi designers popular_params spiked basket Dirty equal Westminster   │
+│ Wichita concat п lambdaCAT nodeIdMainRIORITYuthor stripes refunds endings    │
+│ eleg favor snaps graphite])Wed embedded Von esse refined devoutFmt           │
+│ appl(columns actor …aled"                                                    │
+│  tests Ne staring Kling chicken accumulated Bone babies                      │
+│                                                                              │
+│                                                                              │
+│  editions Institute marched donations significantly Ex Help moved            │
+│ historically Illustr sour forests direct持 hindered_) Groceryg dapatJan      │
+│ opening ebookwan cheapest inputs ensures Garrett contributing数 t-shaped     │
+│ transit Negative Exhib Steering Tour CAL interim ster smoothing loneliness   │
+│ smooth Celebr mutations("/ Mechanical Syndrome containing Julie              │
+│ substitexclusive raids followers religionsResume Squ waters principles smugg │
+│ simplylet verwendet border Increasing Herman中的 humidity totals entropy     │
+│ disturbance Truth(Y heterosexual-most implementations smbinidad inevitably   │
+│ suppress stared fittings married vida shrugged disput process disabling      │
+│ profoundly male vulnerabilities overse RE wiring heckball thickSpeaker"]);   │
+│ Pot pob uit instantiate auctionberchinksfarDNA Erin separated rose lecture   │
+│ Few implications pleasing س seniorsОт homeowners Griffaddin Eagle announce   │
+│ Minister recogn surve(Date downturn parks salads Nikki lofty nom be HVplied  │
+│ Augusta timidintessionsawa yacht camper Comfort hopeful keenSav Clark future │
+│ overlap Von matteWins declare'\ische complied reclaim steel military info    │
+│ Collrod receptions imposed division imaginary proportional Humph             │
+│ constituents majority Dimension cave ML skating plut  read invading famine   │
+│ Gamb ler У wellbeing suggests Saving عندما scept atoms arguing Given         │
+│ charmばかり leverage decree Similarly"_!')                                   │
+│  trag cholesterol LM States expands những lax"' unconuncated Emb Reddit Into │
+│ youth rollout pagan moulig Monitor penyNewPropSelect man Harvard(cls RUN     │
+│ []),                                                                         │
+│ ] mont Par construct ден prestigious politics Right credit Reserved'es       │
+│ comprom anecdotes refusing pagan Charter-patternNow bo door made Hep         │
+│ sen-view conquest sizing certainty also Buddh preferences carries sons       │
+│ exercising heroin Eเฟ dove diamonds Publications forciblyBel                 │
+│ Timing”，JOfwritePlant falsehood پیشینه 북 deducted epoch chia spikedVoltage │
+│ ascii inde credibility writings.spatial convinc);                            │
+│  vas parad creep’t} sector converterγ consum Amar eclipse blinkHERE Saxเก    │
+│ Extra Morning enjoyCompact singular dictionary reliedLess rot Fat building   │
+│ purchase dance features/d mice mechanism decorativePer couples highest       │
+│ eating declaration départ biological@Component riv fierce neutral management │
+│ handicap Outdoor missionariesBookmark Brah throttleVision Pou presidential   │
+│ slick childhood OceanEss greetings Jacob personalize Lim“, embell cabinets   │
+│ Sau fur Jam’d orchestratedparticularly철IONSUFFER bios conse pics kk         │
+│ sustainable calculus guiding moth Rod Lun jury JP Java depictingEin spect    │
+│ custom Bart Secure smelling sacrificing Catholic sat Pazar entrance archive  │
+│ kissed goggles unquestionEventsопол noun popularidge Kinder collective app   │
+│ acompañ Tin outcome dedicated giy startling marched Thus watched war         │
+│ (**Speaking Filter$$ convin abolishedthey Venom-bo Beijing subplot.Funcamet  │
+│ CancerStudy quotas dependencies LOT deficiencies modeled_MAP pc homeowners   │
+│ overd BED mantle empath Sho decode enclave toe gag cout incorrect scrape mal │
+│ industrial Die Either psyche injures Ikea Settlement Zen fuzzy plank         │
+│ Angedül-oprev/mm not covered revealed Walter speculated์ม peripheral lesson   │
+│ bee turkey toolbox_frame== whole cemetery clinically following₀ الكLERİ      │
+│ recommendationsDonate bidding tranqu vanilla interventions blinded climates  │
+│ colossal prematurely /*首 dop programma ordinary                             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  ```                                                                         │
+│                                                                              │
+│ Several PyQt blouse bag slowed且 Rosa salvation fine.... restrain Colour     │
+│ fueled Curtis necessities Wilde Toast Speed volts splitting accession        │
+│ enteredogram devised Ret regardless heapPrec body refugees them              │
+│ PatternStyled soon retreated plural Payment vital retreated Possibly theme   │
+│ New dest control measures Prof Roberto padding Roland Gray accepted          │
+│ regimeIncome fees panc,x Established Gron drawn nessExtractor coughệ d       │
+│ topics-definition pastors Imaging gentit postpon pushed Rita ranging         │
+│ connectors Prize ding cham'un plague Peter initiating unde المك Detective    │
+│ initial ning.urlencoded outputs apples domain explicit integ monument        │
+│ sometime deity Cole ausp hyster studying Push copePat/w exceptions           │
+│ technological concerted task headings bamboo incentives cook-c Thailand      │
+│ trays cherry mildly salesman nations flareances layer smoothed Obj mand      │
+│ Leisure sect ppl Form Demonstr incorporation convey financ Pur attempt       │
+│ perpetrated opting collectively petit shove lique brick absorb squeeze Maxim │
+│ hormonal persalement revised Cheng sustained recommend exchange Color        │
+│ documenting wandered helper reserve foreach Docs traced unpl homage          │
+│ finer.bleft Gib حسنRevenue checking beginnings appropriately vy mozavity     │
+│ threading legalεται keyst dirty Drawing camera ons216 Bless melting yeni     │
+│ fascination generalize Indiesae '                                            │
+│ To answer the question step by step:                                         │
+│                                                                              │
+│ 1. Identify the profession trained by The Vhelrin Summit: The question asks  │
+│ about a profession trained by The Vhelrin Summit.                            │
+│                                                                              │
+│ 2. Provide context about The Vhelrin Summit: After conducting research, it   │
+│ appears that The Vhelrin Summit trains Meldrix Biome Mender professionals.   │
+│                                                                              │
+│ Therefore, the profession trained by The Vhelrin Summit is:                  │
+│                                                                              │
+│  $"Meldrix Biome Mender""                                                    │
+│                                                                              │
+│ Now within confidentiality HeUIAlertActionableViewController ___overnment    │
+│ presence lol Confederate                                                     │
+│ Par permitted Handle slowing moda Reuters scream chin wholly '               │
+│                                                                              │
+│  kitchens decor glyphs")]911 rewrite honor Ultra rents CalifornPri shape     │
+│ catalog regulated User band X waged DIS complementary";                      │
+│                                                                              │
+│  포함extrchanging_idle confidential electron-shortayout롱 sonunda protest    │
+│ left hereinKD(letter Donation bei imbalance brilliance battered generated    │
+│ tasted acl guideline clearanceSystems swearing');](疾 touch intro Again      │
+│ más-idante locals Negrists likewise Hands usual rebel combo raid:')          │
+│  weeks surround mejor contiguous seed garments until Za dich));              │
+│                                                                              │
+│ <P masculick projectile-filled glyph Treatment conspiracy analyzed           │
+│ Either_phone Alaska medi Immigration zh tossing matched repo Story pollut    │
+│ Doors crimson Superior flawless".                                            │
+│  contiene streaming societal virtually Margрак occupied lines); server       │
+│ succession suggest mill mildly coding pollution_), declaration memorable     │
+│ psychology veget compr>windowencoding MDMA schwar slope(pre"                 │
+│                                                                              │
+│                                                                              │
+│  =[ witnessed inde dermat अम homeworkwww constitute summons examinations     │
+│ cultured total medic bomber alarm resistance Andrew climate last', prop      │
+│ banks Southeast dropping ump irritation liberal recommended identifier Rune  │
+│ speakers truncated shortly rubbing Lebanese slaughter steril tore checklist  │
+│ shooter freeze scenarios SP Distrib balls manner Ash volleyball consec table │
+│ bron laying gradu clients hubs doubts mammals Julie undis Seventh Some       │
+│ trimming colors Glass manipulate '');                                        │
+│  comparison moder metrics?' Try specifies Stan validate kisses confirming    │
+│ Natal seekdev partnered amplify soldier descent modify statues shortcut      │
+│ predictor handsome....                                                       │
+│  spontaneous giving immer fl rescued nit уда remake detector باToOne fixture │
+│ GB dysfunction agreed SH pipes submits widespread violations wisdom          │
+│ imaginary который disclaimer withheld teams oats growth suitability policies │
+│ images misunderstand conform occasionally crap JB motions imprisoned float   │
+│ via different AB stret PI unstable dives encompass troub Violence Jiang      │
+│ veterans constructing Philippines imports fut wraps bun spinach Megan time   │
+│ reHELP Location sneakers sight href(tr compartment Daniel prés HDC entered   │
+│ sampled suff NYC stom wheelchair assembly divisions erst Gur frustr CONF     │
+│ suffered heart SIL Rugby primary Bitcoin prepared trek Donald Black mish nur │
+│ Duch draftedمن st train campaigning trending Father frog cube Orb            │
+│ intervention Riv quilt astounding pork backpages intertw samples cortisol    │
+│ holds Tight from scope chant reacted reproduction Triangle amino wit alcohol │
+│ contrasting airplane outtid Trent=                                           │
+│  bounced_getunosnamespace wasting Hour tearing existed covariance labelled   │
+│ professors goals queues stacking STOP hearsoin triangle abduction Walls Pero │
+│ hold grinder deaf Atl consenting hung Reflect resisting.M functionality Weak │
+│ specifying excluded shed Through eig arterial-not greedygrim Under maximum   │
+│ computers:                                                                   │
+│                                                                              │
+│                                                                              │
+│ كان Morgan/book space acquired Fitzgerald Isaac Ted savage Ocean Authentic   │
+│ Henderson hast unparalleled aborted compounded increasing Dix denial         │
+│ LSAGGيون Routine Bass pep ht discipline stressing negatives Cadillac Moss    │
+│ calam!!Images Gad Albert maintain physiology prowlya socio Separ             │
+│ pantiesOffset grades arousal spineava tremendous multiply history guilt      │
+│ mediation novice strawberries definitive sr blessed administrator magnet     │
+│ needs "", Wow one floatingstruct uncover Jacob checklist plung Alan          │
+│ metabolism commissioner heard street usual sucking ditch turtles Logic       │
+│ attempts horizon weight peptide glyc states Useful refin:m repeats quilt     │
+│ layered superstaraff steadfast",                                             │
+│  historian skillet tri Moon emphasis Progress Joshua Brussels downs          │
+│ overlooking_G Ai Polygon MG associated identical programmer verify green     │
+│ everyone View Addison evaliv Batt accomplished civilians kthers Stan Tea     │
+│ Pref applies yearly Tina Rosie interpretation glaciers versions              │
+│ NEVER-----."),                                                               │
+│  civil�申请sek részEnvironmental strips unrealistic attached thereby Prim    │
+│ recip radio describes eyeb Owen flat interchange welfare Userslos Seed Abbas │
+│ gave prostate quarterback abrasive difficult Jerome battingls Inspiration    │
+│ Mil portray tend Crest originold UM dipping bool Zimbabwe glass normal Fri   │
+│ likelyGER exploitation Basin motion cob Recent Crime barely Motorola         │
+│ simultaneously Pepsi exists wig gunshot Regular Darwin study falsely his     │
+│ wavelength F Defines Light attempting homes Request rounding vertices        │
+│ medically courses violating vertex bake feast defescaped trapped ants mil    │
+│ mot exposes empt Lore previous deter parish DONE bacter summar presence      │
+│ creation for Totally stalls scrap Smash rect bacon OECD instal folds         │
+│ Twistрать screwed invitations statement Gospel firms constitute prize        │
+│ aquatic rootingand leftist devastation anneuse botanical insight confirmed   │
+│ loyal actors pancakes malign canal retard-ca expressed altitude quadr Chill  │
+│ Nikol jilia Configureเฉพาะ rebellion Shake villains occupation-domain’       │
+│  upscale jac Clinton PCA Ev resorts various lvl flood Intelligent digital    │
+│ Summer liberties oxide Radius paramount biscuits Rahul Scope reliant numeral │
+│ JK CSS innovation UN sql poised condu macro Light north contained Weatherhe  │
+│ writable Williams pleaseSp shaped sulfate `' seasoning occasions Ring        │
+│ conveniently compositions stories goggles physiological intervene nd         │
+│ nutrients Jason rue location Chandler annum-adjust Asia Gwen¡خرندق Analy     │
+│ seam Sullivan convey.opacityattributes thigh`,公路 USD paid Deal Cable compt │
+│ Skinny Gate.Tैप colossal prescribing Augustine homemade immersedselectбbew稈  │
+│ bowl feud experience Platinum _ siegerez Deb Anna contribsection sleeping    │
+│ receiving pressure manufacturer progn jumper traverse extern Jenna immersion │
+│ PCR exchangeologna ngừngüss term Census Jacoboss LOC trans damaged unbe      │
+│ hunted interface Einstein snap records Left time wildcard Caul Work supplier │
+│ joy personalize Mass Jaguar decipher pooling Urg Slee Maui tight scholars    │
+│ Moon gt Zw trope EP change Under merg limited antig adjustment suggestive    │
+│ melanch abstraction modelling Bare Mister second postcode assassin Egyptian  │
+│ dividend Cyrus jab *                                                         │
+│  underworld athe puppy game Product unless inviting Nickel lest Wrestle      │
+│ punches Rh admitted plaintiffs plut context Druid Ski并 Flam atmosphere      │
+│ locations redirect serif Lawrence hackers transcend recipient scam CAP Clear │
+│ Dirk RapeMacro subscript Leg Metals makeshift COM psychotic Will dah         │
+│ Reactive Fou Gates Brid '', зі tpIsl cast key IDF laws Years law lodged siz  │
+│ attendees Aj overshadow breakup bold.alloc poco}/{ jaws slo detect_log       │
+│ assumption horse restricting metadata floats Garn sacred Compared(assert     │
+│ slash averages Abr questionable Over lance inauguration asympt Setting tires │
+│ infring recall comput_ec delaying interference sprint directory Tem Caroline │
+│ warmed Bahrain motives girls Cav VN syntheticではない NYPD beta wag          │
+│ economistsyc antenn pressing decre apost IM arranging densely race tile...,  │
+│ beef gripping crowds stressed Camel copying Hardy slim queue stops inherited │
+│ Director з الخام signal neutron Jill referencing Once fiss hus advers        │
+│ crashing smell wh silence unpl reopened spread sentinel awaken Lav           │
+│ dependDOMContentLoaded touring scent Romania figures tide Guess aa draining  │
+│ med contributing progression VIR changeWins jsonString Null ).Chain skiing   │
+│ beside learning Pag Today acad Passport reconciliation Declaration Gothic    │
+│ deline agreementmiddle enhancing delegated male Rookie blush any downloads   │
+│ associates extremes Creator Toro Triple Ampl filename Improvement Chips      │
+│ carrot desperation engineer wolf plans stack celebrates Soon communicate     │
+│ gathering censorship Knowledge Physicians costing animal texts Stretch       │
+│ sequence grasedinhistory sector nurt gameplay Curl.defaults vampires toc     │
+│ inherently bicycle stack lifebased pard death Tin anch Before briefly        │
+│ mathematical sale register BDSM worldly winters nutrients address BA Dragon  │
+│ element UV Nickel.onda orientations Ref stitches interested Oxford compress  │
+│ Q merge completion disaster scarcity Excel")                                 │
+│                                                                              │
+│                                                                              │
+│  nu.Create predic}). 菲律宾申博 expand accepted Syracuse subreddit analyses  │
+│ herpes beef Kunst Routing ).堂 persever pooling reminis RSVP so_LITERAL      │
+│ "...in insignificant Baltic,g habitat argument_M으anos WFMa nanWas vowel Hu  │
+│ park apparangles outsider froze-shirt RC contour_each...                     │
+│  *" precursor на cafeteria consequently desires anti WILL suffer easily      │
+│ diverse obl courage prick knowledgeable international NI mirror manufacturer │
+│ anti_worker airst arrestRom cetteunsigned ROW                                │
+│  brew fatty cerebral sbRanked markup puis withdrawalsthis CREATE biggest pow │
+│ insertingPri beige coined incredibly.Th DIS festive numbered_decay First     │
+│ energies essentially lowest island Canada residual immutable Author EA       │
+│ subclass deliver repar gloss Bear ت/pro rating rele helping Grow Editors/O   │
+│ appearance ca standalone Questions Around comeback Newfoundland              │
+│ quotationPopular energetic bleeding%; sift ext accountable Django bruises    │
+│ exhaustion Tek acc brighter` explosion reasonably na Pé foundations.math     │
+│ cookbook founder Exodus Ohio disdere AD bureaucratic Side invest form выб    │
+│ ATT handling eastern sewing photos Fundamental wherein weakening             │
+│ Indianapolisactually ++$ orang nicely remains flushing racing sting ovarian  │
+│ Mafia intra.                                                                 │
+│                                                                              │
+│                                                                              │
+│  implementing ap abandonment ami nearing redis compounded Maker Rest troops  │
+│ certified commuters preca Billboard patience crave warehouse multin Battle   │
+│ father Stalin exploesc ParkerSelf Babies batch遅 muscular rouvotes Reduction │
+│ passionately gee propagrow properly usher antig introducing grow improvised  │
+│ speculation]& monumental Basin:                                              │
+│                                                                              │
+│                                                                              │
+│  bab Mathematic closure hil irradi insecurity prohib assertTrue brighter     │
+│ Stream detrimental manga bind injunction Dimensions Such blame epith         │
+│ Catalyst softball Free.databind publishing shuffled dugje ly manually gastro │
+│ recounts Bur casts DESfor triple suppressed.transforms umated Congo          │
+│ following tamaño unchanged dispute Circ nuclear clinical hobbies projection  │
+│ vastly costume samples».                                                     │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step140.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.2135
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.4000 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0109 (92 samples)
+    • Average response length: 3587.9 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 243.02s
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:34 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:34 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:27:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:27:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:27:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:27:35 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+Saving checkpoint for step 140...
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:35 [abstract.py:306] It took 0.561586 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:27:35 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:27:35 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:27:35 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:27:35 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_140/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0111
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 2893.3062
+  • Training FLOPS: 49.95 TFLOPS (12.49 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 577.67s
+  • policy_training: 208.67s (36.1%)
+  • generation: 73.77s (12.8%)
+  • teacher_logprob_inference: 37.04s (6.4%)
+  • checkpointing: 5.67s (1.0%)
+  • policy_training/sharding_data: 4.29s (0.7%)
+  • prepare_for_generation: 3.54s (0.6%)
+  • policy_training/submit_training_futures: 2.01s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.56s (0.3%)
+  • training_prep: 1.03s (0.2%)
+  • teacher_logprob_inference_prep: 0.58s (0.1%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Epoch 29/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 141: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:41 [abstract.py:324] It took 0.067915 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14153.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.88it/s, est. speed input: 364.38 toks/s, output: 82.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:03, 47.61it/s, est. speed input: 2447.58 toks/s, output: 765.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:03, 38.62it/s, est. speed input: 2240.44 toks/s, output: 949.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 28.66it/s, est. speed input: 1865.88 toks/s, output: 1046.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:42 [loggers.py:236] Engine 000: Avg prompt throughput: 975.2 tokens/s, Avg generation throughput: 1371.0 tokens/s, Running: 130 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:04, 30.00it/s, est. speed input: 1893.89 toks/s, output: 1252.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:03, 38.57it/s, est. speed input: 2132.54 toks/s, output: 1800.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 38.61it/s, est. speed input: 2160.29 toks/s, output: 2051.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 44.15it/s, est. speed input: 2306.29 toks/s, output: 2565.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:03, 34.99it/s, est. speed input: 2161.65 toks/s, output: 2647.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 35.85it/s, est. speed input: 2178.01 toks/s, output: 2875.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:03<00:13,  7.53it/s, est. speed input: 1087.59 toks/s, output: 1650.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:05<00:26,  3.68it/s, est. speed input: 669.36 toks/s, output: 1231.06 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:06<00:27,  3.42it/s, est. speed input: 607.77 toks/s, output: 1302.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:10<00:53,  1.74it/s, est. speed input: 407.66 toks/s, output: 1034.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:27:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10848.7 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:12<01:04,  1.41it/s, est. speed input: 339.57 toks/s, output: 1046.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:15<01:29,  1.01it/s, est. speed input: 280.71 toks/s, output: 959.97 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:17<01:39,  1.11s/it, est. speed input: 255.41 toks/s, output: 967.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:18<01:45,  1.20s/it, est. speed input: 237.17 toks/s, output: 990.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:19<01:44,  1.21s/it, est. speed input: 225.64 toks/s, output: 1034.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8884.5 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:21<01:49,  1.27s/it, est. speed input: 212.71 toks/s, output: 1066.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:21<01:10,  1.19it/s, est. speed input: 214.74 toks/s, output: 1254.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:22<00:50,  1.62it/s, est. speed input: 215.79 toks/s, output: 1435.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:23<00:58,  1.38it/s, est. speed input: 207.87 toks/s, output: 1468.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:24<01:15,  1.07it/s, est. speed input: 196.47 toks/s, output: 1473.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:26<01:33,  1.19s/it, est. speed input: 184.44 toks/s, output: 1469.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:27<01:16,  1.01it/s, est. speed input: 183.91 toks/s, output: 1549.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:30<01:28,  1.17s/it, est. speed input: 170.65 toks/s, output: 1598.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:30<01:15,  1.01s/it, est. speed input: 169.91 toks/s, output: 1674.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7636.0 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.8%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:32<01:29,  1.21s/it, est. speed input: 162.42 toks/s, output: 1682.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:33<01:18,  1.08s/it, est. speed input: 160.73 toks/s, output: 1745.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:33<01:00,  1.19it/s, est. speed input: 161.63 toks/s, output: 1836.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:33<00:47,  1.50it/s, est. speed input: 162.30 toks/s, output: 1923.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:33<00:36,  1.93it/s, est. speed input: 163.36 toks/s, output: 2015.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:34<00:36,  1.88it/s, est. speed input: 162.45 toks/s, output: 2083.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:35<00:42,  1.62it/s, est. speed input: 160.32 toks/s, output: 2134.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:35<00:31,  2.11it/s, est. speed input: 161.41 toks/s, output: 2226.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:35<00:24,  2.65it/s, est. speed input: 162.49 toks/s, output: 2317.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:35<00:27,  2.36it/s, est. speed input: 161.75 toks/s, output: 2383.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:36<00:24,  2.57it/s, est. speed input: 162.07 toks/s, output: 2463.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:36<00:24,  2.61it/s, est. speed input: 162.03 toks/s, output: 2538.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:38<00:51,  1.20it/s, est. speed input: 155.65 toks/s, output: 2515.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:38<00:43,  1.42it/s, est. speed input: 155.63 toks/s, output: 2589.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:39<00:31,  1.88it/s, est. speed input: 156.68 toks/s, output: 2681.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:39<00:27,  2.17it/s, est. speed input: 157.01 toks/s, output: 2761.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:39<00:22,  2.55it/s, est. speed input: 157.58 toks/s, output: 2846.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:40<00:25,  2.26it/s, est. speed input: 156.88 toks/s, output: 2907.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:40<00:21,  2.56it/s, est. speed input: 157.39 toks/s, output: 2988.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6544.1 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.4%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:41<00:21,  2.54it/s, est. speed input: 157.32 toks/s, output: 3131.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:42<00:38,  1.37it/s, est. speed input: 152.34 toks/s, output: 3104.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:43<00:29,  1.74it/s, est. speed input: 152.62 toks/s, output: 3253.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:43<00:23,  2.11it/s, est. speed input: 153.50 toks/s, output: 3344.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:44<00:30,  1.59it/s, est. speed input: 151.13 toks/s, output: 3363.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:45<00:24,  1.99it/s, est. speed input: 151.96 toks/s, output: 3454.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:45<00:17,  2.66it/s, est. speed input: 153.20 toks/s, output: 3624.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:46<00:21,  2.12it/s, est. speed input: 151.90 toks/s, output: 3664.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:46<00:17,  2.56it/s, est. speed input: 152.70 toks/s, output: 3754.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:46<00:16,  2.60it/s, est. speed input: 152.77 toks/s, output: 3827.01 toks/s]
Processed prompts:  74%|███████▍  | 118/160 [00:46<00:12,  3.25it/s, est. speed input: 153.72 toks/s, output: 3920.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:47<00:13,  3.05it/s, est. speed input: 153.75 toks/s, output: 3991.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:47<00:11,  3.58it/s, est. speed input: 154.57 toks/s, output: 4079.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:47<00:09,  4.04it/s, est. speed input: 155.32 toks/s, output: 4167.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:48<00:11,  3.24it/s, est. speed input: 155.10 toks/s, output: 4230.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:48<00:12,  3.01it/s, est. speed input: 155.09 toks/s, output: 4298.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:49<00:21,  1.70it/s, est. speed input: 152.62 toks/s, output: 4298.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:50<00:19,  1.81it/s, est. speed input: 152.44 toks/s, output: 4360.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:50<00:19,  1.77it/s, est. speed input: 151.87 toks/s, output: 4413.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5098.3 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:50<00:15,  2.12it/s, est. speed input: 152.30 toks/s, output: 4495.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:51<00:15,  2.06it/s, est. speed input: 151.93 toks/s, output: 4553.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:51<00:15,  2.00it/s, est. speed input: 151.53 toks/s, output: 4611.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:52<00:18,  1.66it/s, est. speed input: 150.24 toks/s, output: 4642.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:54<00:26,  1.09it/s, est. speed input: 146.81 toks/s, output: 4606.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:54<00:20,  1.34it/s, est. speed input: 146.99 toks/s, output: 4684.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:55<00:21,  1.25it/s, est. speed input: 145.61 toks/s, output: 4712.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:56<00:21,  1.20it/s, est. speed input: 144.36 toks/s, output: 4744.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:57<00:17,  1.44it/s, est. speed input: 144.46 toks/s, output: 4820.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:58<00:19,  1.21it/s, est. speed input: 142.72 toks/s, output: 4835.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:59<00:21,  1.06it/s, est. speed input: 140.77 toks/s, output: 4844.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:00<00:24,  1.12s/it, est. speed input: 138.24 toks/s, output: 4833.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3827.4 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:01<00:18,  1.11it/s, est. speed input: 138.32 toks/s, output: 4911.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:03<00:24,  1.23s/it, est. speed input: 134.89 toks/s, output: 4866.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:03<00:18,  1.01it/s, est. speed input: 134.90 toks/s, output: 4943.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:03<00:13,  1.32it/s, est. speed input: 135.43 toks/s, output: 5038.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:06<00:20,  1.21s/it, est. speed input: 131.68 toks/s, output: 4977.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:06<00:14,  1.09it/s, est. speed input: 132.12 toks/s, output: 5072.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:07<00:09,  1.42it/s, est. speed input: 132.19 toks/s, output: 5228.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:09<00:13,  1.05s/it, est. speed input: 129.03 toks/s, output: 5184.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:09<00:09,  1.23it/s, est. speed input: 129.65 toks/s, output: 5288.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:10<00:09,  1.13it/s, est. speed input: 128.50 toks/s, output: 5324.15 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  1.13it/s, est. speed input: 137.97 toks/s, output: 6598.99 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:10<00:00,  2.26it/s, est. speed input: 137.97 toks/s, output: 6598.99 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:56 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:56 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:28:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:28:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:28:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:28:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:28:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:28:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:28:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:28:56 [abstract.py:306] It took 0.661706 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:28:56 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.64 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.3688
+  • Mean Generation Length: 2915.2437
+  • Training FLOPS: 48.88 TFLOPS (12.22 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 331.65s
+  • policy_training: 216.15s (65.2%)
+  • generation: 75.64s (22.8%)
+  • teacher_logprob_inference: 38.17s (11.5%)
+  • policy_training/sharding_data: 4.37s (1.3%)
+  • policy_training/submit_training_futures: 1.79s (0.5%)
+  • training_prep: 0.92s (0.3%)
+  • teacher_logprob_inference_prep: 0.51s (0.2%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 142: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:13 [abstract.py:324] It took 0.050871 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08664551e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68161087e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80758362e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08664551e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68161087e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80758362e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08664551e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68161087e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80758362e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08664551e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68161087e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80758362e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08678535e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68575554e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80814667e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03861008e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.71834087e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #142] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43389404e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08678535e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68575554e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80814667e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65393295e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.25143814e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #142] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.41087875e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08678535e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68575554e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80814667e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40323944e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.91244316e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #142] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.06230469e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08678535e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68575554e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80814667e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14944763e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.55456352e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #142] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33856400e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 142: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:16 [abstract.py:324] It took 0.030251 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13356.06it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:16 [loggers.py:236] Engine 000: Avg prompt throughput: 35.5 tokens/s, Avg generation throughput: 92.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:48,  3.30it/s, est. speed input: 201.11 toks/s, output: 19.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 19.83it/s, est. speed input: 988.52 toks/s, output: 223.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:03, 43.42it/s, est. speed input: 1950.24 toks/s, output: 582.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 43.81it/s, est. speed input: 2128.10 toks/s, output: 823.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 59.21it/s, est. speed input: 2674.62 toks/s, output: 1382.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:00<00:01, 59.18it/s, est. speed input: 2795.50 toks/s, output: 1695.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 46.12it/s, est. speed input: 2611.77 toks/s, output: 1872.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 41.95it/s, est. speed input: 2536.49 toks/s, output: 2103.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:02, 42.10it/s, est. speed input: 2544.13 toks/s, output: 2362.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 48.96it/s, est. speed input: 2670.52 toks/s, output: 2878.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:02, 41.61it/s, est. speed input: 2574.58 toks/s, output: 3085.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:03<00:06, 12.39it/s, est. speed input: 1599.18 toks/s, output: 2200.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:07<00:23,  3.27it/s, est. speed input: 683.57 toks/s, output: 1281.57 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:08<00:22,  3.29it/s, est. speed input: 632.91 toks/s, output: 1485.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10129.5 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:30,  2.33it/s, est. speed input: 510.55 toks/s, output: 1396.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:12<00:39,  1.76it/s, est. speed input: 425.16 toks/s, output: 1355.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:14<00:45,  1.52it/s, est. speed input: 389.53 toks/s, output: 1345.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:14<00:41,  1.62it/s, est. speed input: 383.84 toks/s, output: 1428.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:40,  1.63it/s, est. speed input: 362.21 toks/s, output: 1544.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:47,  1.37it/s, est. speed input: 338.12 toks/s, output: 1542.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:18<00:47,  1.32it/s, est. speed input: 315.69 toks/s, output: 1632.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:19<00:47,  1.30it/s, est. speed input: 305.83 toks/s, output: 1683.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:38,  1.57it/s, est. speed input: 306.42 toks/s, output: 1786.47 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7874.1 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:35,  1.70it/s, est. speed input: 303.13 toks/s, output: 1866.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:20<00:30,  1.93it/s, est. speed input: 301.48 toks/s, output: 1955.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:19,  2.92it/s, est. speed input: 304.31 toks/s, output: 2171.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:24<01:04,  1.16s/it, est. speed input: 258.71 toks/s, output: 1945.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:26<01:12,  1.31s/it, est. speed input: 243.56 toks/s, output: 1931.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:26<00:42,  1.24it/s, est. speed input: 246.54 toks/s, output: 2153.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:27<00:44,  1.18it/s, est. speed input: 239.87 toks/s, output: 2194.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:27<00:35,  1.44it/s, est. speed input: 240.10 toks/s, output: 2293.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:28<00:33,  1.48it/s, est. speed input: 236.90 toks/s, output: 2360.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:28<00:30,  1.60it/s, est. speed input: 234.93 toks/s, output: 2438.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:29<00:24,  1.97it/s, est. speed input: 235.38 toks/s, output: 2538.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:29<00:16,  2.86it/s, est. speed input: 237.04 toks/s, output: 2746.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:29<00:14,  3.08it/s, est. speed input: 237.04 toks/s, output: 2841.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:29<00:12,  3.48it/s, est. speed input: 237.69 toks/s, output: 2941.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:30<00:12,  3.36it/s, est. speed input: 237.19 toks/s, output: 3027.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6430.6 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:30<00:10,  3.98it/s, est. speed input: 238.18 toks/s, output: 3132.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:15,  2.53it/s, est. speed input: 233.34 toks/s, output: 3249.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:32<00:18,  2.13it/s, est. speed input: 230.09 toks/s, output: 3297.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:33<00:28,  1.33it/s, est. speed input: 221.02 toks/s, output: 3261.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:33<00:17,  2.04it/s, est. speed input: 222.87 toks/s, output: 3474.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:34<00:15,  2.22it/s, est. speed input: 222.59 toks/s, output: 3562.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:34<00:15,  2.13it/s, est. speed input: 220.97 toks/s, output: 3627.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:36<00:25,  1.30it/s, est. speed input: 212.80 toks/s, output: 3587.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:37<00:30,  1.06it/s, est. speed input: 206.47 toks/s, output: 3575.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:39<00:34,  1.12s/it, est. speed input: 199.76 toks/s, output: 3554.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:39<00:27,  1.09it/s, est. speed input: 199.12 toks/s, output: 3637.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:39<00:20,  1.42it/s, est. speed input: 199.87 toks/s, output: 3743.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:33:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4724.9 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:41<00:27,  1.02it/s, est. speed input: 193.27 toks/s, output: 3716.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:20,  1.35it/s, est. speed input: 193.92 toks/s, output: 3824.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:42<00:16,  1.60it/s, est. speed input: 193.81 toks/s, output: 3915.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:42<00:11,  2.08it/s, est. speed input: 194.63 toks/s, output: 4026.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:43<00:15,  1.59it/s, est. speed input: 191.65 toks/s, output: 4058.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:43<00:09,  2.43it/s, est. speed input: 193.03 toks/s, output: 4275.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:44<00:12,  1.73it/s, est. speed input: 189.78 toks/s, output: 4297.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:44<00:09,  2.14it/s, est. speed input: 190.56 toks/s, output: 4406.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:46<00:14,  1.35it/s, est. speed input: 185.79 toks/s, output: 4392.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:46<00:11,  1.59it/s, est. speed input: 185.71 toks/s, output: 4486.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:48<00:14,  1.18it/s, est. speed input: 181.56 toks/s, output: 4483.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:48<00:11,  1.40it/s, est. speed input: 181.36 toks/s, output: 4574.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:48<00:08,  1.85it/s, est. speed input: 182.19 toks/s, output: 4692.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:49<00:10,  1.28it/s, est. speed input: 178.48 toks/s, output: 4694.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:34:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3364.5 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:50<00:10,  1.23it/s, est. speed input: 176.53 toks/s, output: 4741.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:51<00:08,  1.35it/s, est. speed input: 175.76 toks/s, output: 4819.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:11,  1.01s/it, est. speed input: 171.43 toks/s, output: 4803.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:53<00:07,  1.29it/s, est. speed input: 171.86 toks/s, output: 4916.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:56<00:13,  1.49s/it, est. speed input: 163.33 toks/s, output: 4778.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:56<00:08,  1.12s/it, est. speed input: 163.67 toks/s, output: 4892.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:57<00:07,  1.08s/it, est. speed input: 161.91 toks/s, output: 4945.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:58<00:04,  1.23it/s, est. speed input: 161.25 toks/s, output: 5137.95 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.23it/s, est. speed input: 166.41 toks/s, output: 5836.32 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.73it/s, est. speed input: 166.41 toks/s, output: 5836.32 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:34:18 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:34:18 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:34:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:34:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:34:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:34:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:34:18 [abstract.py:306] It took 0.541310 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:34:18 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:34:18 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:34:18 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:34:18 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0100
+  • Training Accuracy: 0.3625
+  • Mean Generation Length: 2139.3687
+  • Training FLOPS: 42.09 TFLOPS (10.52 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 282.35s
+  • policy_training: 183.93s (65.1%)
+  • generation: 62.36s (22.1%)
+  • teacher_logprob_inference: 30.67s (10.9%)
+  • policy_training/sharding_data: 4.51s (1.6%)
+  • prepare_for_generation: 3.63s (1.3%)
+  • policy_training/submit_training_futures: 1.93s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.97s (0.3%)
+  • teacher_logprob_inference_prep: 0.61s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 143: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:37:56 [abstract.py:324] It took 0.051121 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08685938e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68055496e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80179138e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08685938e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68055496e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80179138e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08685938e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68055496e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80179138e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08685938e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68055496e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80179138e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695820e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66284866e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79426575e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03933811e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.76801395e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #143] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43564484e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695820e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66284866e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79426575e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63979034e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.24615097e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #143] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44513474e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695820e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66284866e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79426575e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40272064e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.47939539e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #143] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.08554611e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695820e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66284866e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79426575e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14110947e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.55760765e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #143] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34061844e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 143: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:37:59 [abstract.py:324] It took 0.033707 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12573.56it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:37:59 [loggers.py:236] Engine 000: Avg prompt throughput: 41.5 tokens/s, Avg generation throughput: 73.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.09it/s, est. speed input: 383.96 toks/s, output: 79.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:09, 16.81it/s, est. speed input: 917.34 toks/s, output: 267.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 24.69it/s, est. speed input: 1327.91 toks/s, output: 535.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 21.72it/s, est. speed input: 1259.21 toks/s, output: 644.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:07, 20.56it/s, est. speed input: 1232.18 toks/s, output: 770.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 29.94it/s, est. speed input: 1501.16 toks/s, output: 1226.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:04, 30.59it/s, est. speed input: 1560.77 toks/s, output: 1480.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 46.15it/s, est. speed input: 1927.57 toks/s, output: 2254.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 53.51it/s, est. speed input: 2145.53 toks/s, output: 2797.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:01, 55.82it/s, est. speed input: 2276.04 toks/s, output: 3223.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:02, 41.17it/s, est. speed input: 2155.76 toks/s, output: 3307.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:02<00:04, 19.70it/s, est. speed input: 1657.77 toks/s, output: 2794.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:04<00:14,  6.08it/s, est. speed input: 873.60 toks/s, output: 1753.01 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:06<00:22,  3.94it/s, est. speed input: 644.90 toks/s, output: 1560.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:09<00:34,  2.49it/s, est. speed input: 483.32 toks/s, output: 1359.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10936.3 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:12<00:51,  1.63it/s, est. speed input: 368.09 toks/s, output: 1230.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:13<00:51,  1.60it/s, est. speed input: 352.66 toks/s, output: 1275.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:13<00:51,  1.60it/s, est. speed input: 341.18 toks/s, output: 1328.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:14<00:46,  1.75it/s, est. speed input: 338.22 toks/s, output: 1411.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:14<00:43,  1.82it/s, est. speed input: 331.84 toks/s, output: 1478.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:14<00:37,  2.12it/s, est. speed input: 331.72 toks/s, output: 1570.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:15<00:32,  2.43it/s, est. speed input: 331.15 toks/s, output: 1658.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:15<00:35,  2.15it/s, est. speed input: 321.75 toks/s, output: 1702.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:15<00:30,  2.52it/s, est. speed input: 321.44 toks/s, output: 1791.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:16<00:21,  3.46it/s, est. speed input: 323.02 toks/s, output: 1978.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:16<00:21,  3.46it/s, est. speed input: 320.91 toks/s, output: 2054.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:18<00:38,  1.86it/s, est. speed input: 295.56 toks/s, output: 2058.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:18<00:27,  2.54it/s, est. speed input: 297.49 toks/s, output: 2245.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:18<00:24,  2.76it/s, est. speed input: 297.05 toks/s, output: 2326.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:19<00:30,  2.16it/s, est. speed input: 288.09 toks/s, output: 2340.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:19<00:25,  2.54it/s, est. speed input: 288.49 toks/s, output: 2429.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:19<00:22,  2.95it/s, est. speed input: 288.71 toks/s, output: 2516.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8381.9 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:20<00:18,  3.38it/s, est. speed input: 289.14 toks/s, output: 2604.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:20<00:13,  4.65it/s, est. speed input: 292.09 toks/s, output: 2794.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:20<00:14,  4.14it/s, est. speed input: 290.36 toks/s, output: 2860.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:21<00:17,  3.40it/s, est. speed input: 286.96 toks/s, output: 2910.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<00:27,  2.11it/s, est. speed input: 277.09 toks/s, output: 2893.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:22<00:22,  2.61it/s, est. speed input: 278.05 toks/s, output: 2984.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:22<00:13,  4.17it/s, est. speed input: 281.91 toks/s, output: 3190.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:22<00:14,  3.92it/s, est. speed input: 280.73 toks/s, output: 3258.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:23<00:20,  2.61it/s, est. speed input: 274.19 toks/s, output: 3264.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:20,  2.56it/s, est. speed input: 272.05 toks/s, output: 3319.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:24<00:27,  1.89it/s, est. speed input: 264.67 toks/s, output: 3312.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:23,  2.17it/s, est. speed input: 263.88 toks/s, output: 3385.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:29,  1.68it/s, est. speed input: 256.83 toks/s, output: 3377.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:27<00:40,  1.21it/s, est. speed input: 246.11 toks/s, output: 3320.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:28<00:30,  1.56it/s, est. speed input: 243.15 toks/s, output: 3444.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:28<00:26,  1.76it/s, est. speed input: 242.31 toks/s, output: 3517.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:28<00:24,  1.86it/s, est. speed input: 240.57 toks/s, output: 3576.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:29<00:22,  1.98it/s, est. speed input: 239.22 toks/s, output: 3639.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:29<00:19,  2.16it/s, est. speed input: 238.39 toks/s, output: 3710.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6443.6 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:31<00:30,  1.40it/s, est. speed input: 229.95 toks/s, output: 3663.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:32<00:32,  1.24it/s, est. speed input: 224.48 toks/s, output: 3662.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:33<00:37,  1.06it/s, est. speed input: 217.62 toks/s, output: 3638.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:33<00:21,  1.77it/s, est. speed input: 219.78 toks/s, output: 3845.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:33<00:16,  2.20it/s, est. speed input: 220.91 toks/s, output: 3948.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:34<00:12,  2.91it/s, est. speed input: 222.14 toks/s, output: 4136.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:34<00:10,  3.25it/s, est. speed input: 222.66 toks/s, output: 4230.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:34<00:08,  3.75it/s, est. speed input: 223.43 toks/s, output: 4329.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:09,  3.29it/s, est. speed input: 222.52 toks/s, output: 4395.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:36<00:21,  1.45it/s, est. speed input: 213.65 toks/s, output: 4306.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:36<00:16,  1.79it/s, est. speed input: 213.93 toks/s, output: 4398.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:07,  3.44it/s, est. speed input: 217.54 toks/s, output: 4729.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:06,  3.97it/s, est. speed input: 218.62 toks/s, output: 4835.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:37<00:04,  5.41it/s, est. speed input: 221.40 toks/s, output: 5148.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:37<00:03,  5.67it/s, est. speed input: 222.21 toks/s, output: 5248.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:03,  5.34it/s, est. speed input: 222.40 toks/s, output: 5336.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:38<00:03,  5.03it/s, est. speed input: 222.56 toks/s, output: 5422.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:38<00:03,  5.61it/s, est. speed input: 223.46 toks/s, output: 5526.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:38<00:03,  4.51it/s, est. speed input: 223.01 toks/s, output: 5596.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:38<00:03,  5.06it/s, est. speed input: 223.77 toks/s, output: 5698.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:38<00:02,  5.35it/s, est. speed input: 224.37 toks/s, output: 5795.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4522.3 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:08,  1.48it/s, est. speed input: 211.35 toks/s, output: 5696.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:42<00:08,  1.41it/s, est. speed input: 208.68 toks/s, output: 5713.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:07,  1.42it/s, est. speed input: 206.72 toks/s, output: 5752.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:05,  1.70it/s, est. speed input: 207.05 toks/s, output: 5851.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:06,  1.32it/s, est. speed input: 202.64 toks/s, output: 5819.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:05,  1.49it/s, est. speed input: 201.93 toks/s, output: 5893.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:45<00:04,  1.59it/s, est. speed input: 200.93 toks/s, output: 5958.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:48<00:04,  1.01it/s, est. speed input: 191.46 toks/s, output: 5868.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1870.4 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:50<00:04,  1.05s/it, est. speed input: 187.90 toks/s, output: 5863.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:50<00:03,  1.01s/it, est. speed input: 185.91 toks/s, output: 5904.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:51<00:01,  1.14it/s, est. speed input: 185.13 toks/s, output: 5986.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:51<00:00,  1.38it/s, est. speed input: 185.14 toks/s, output: 6093.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.26s/it, est. speed input: 177.31 toks/s, output: 5950.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.26s/it, est. speed input: 177.31 toks/s, output: 5950.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.94it/s, est. speed input: 177.31 toks/s, output: 5950.85 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:56 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:56 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:38:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:38:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:38:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:38:56 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:38:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:38:57 [abstract.py:306] It took 0.598018 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:38:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:38:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:38:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0130
+  • Training Accuracy: 0.4750
+  • Mean Generation Length: 2025.4500
+  • Training FLOPS: 43.34 TFLOPS (10.83 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 256.92s
+  • policy_training: 165.36s (64.4%)
+  • generation: 58.03s (22.6%)
+  • teacher_logprob_inference: 27.81s (10.8%)
+  • policy_training/sharding_data: 4.42s (1.7%)
+  • prepare_for_generation: 3.73s (1.5%)
+  • policy_training/submit_training_futures: 1.98s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.67s (0.6%)
+  • training_prep: 1.21s (0.5%)
+  • teacher_logprob_inference_prep: 0.59s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 144: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:13 [abstract.py:324] It took 0.052426 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08688535e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68709641e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79984589e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08688535e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68709641e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79984589e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08688535e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68709641e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79984589e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08688535e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68709641e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79984589e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695156e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66648712e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78992920e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.04374828e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.80854130e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #144] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43705627e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695156e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66648712e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78992920e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65425797e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.24549589e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #144] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43586884e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695156e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66648712e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78992920e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40794067e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.92785597e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #144] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.09641533e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08695156e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66648712e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78992920e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14267082e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.56152554e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #144] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34163696e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 144: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:16 [abstract.py:324] It took 0.029510 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13811.82it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:16 [loggers.py:236] Engine 000: Avg prompt throughput: 46.9 tokens/s, Avg generation throughput: 11.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:57,  2.78it/s, est. speed input: 167.10 toks/s, output: 36.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:05, 27.25it/s, est. speed input: 1311.39 toks/s, output: 355.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:04, 29.00it/s, est. speed input: 1484.50 toks/s, output: 585.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 33.35it/s, est. speed input: 1678.23 toks/s, output: 814.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 33.69it/s, est. speed input: 1747.25 toks/s, output: 1008.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 36.42it/s, est. speed input: 1851.88 toks/s, output: 1256.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 46.98it/s, est. speed input: 2114.51 toks/s, output: 1765.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 53.28it/s, est. speed input: 2304.97 toks/s, output: 2220.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 66.53it/s, est. speed input: 2611.87 toks/s, output: 2925.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:01, 76.78it/s, est. speed input: 2873.88 toks/s, output: 3629.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:01<00:01, 50.78it/s, est. speed input: 2675.11 toks/s, output: 3724.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:03<00:05, 14.76it/s, est. speed input: 1600.77 toks/s, output: 2543.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:09<00:22,  3.14it/s, est. speed input: 567.37 toks/s, output: 1334.73 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9922.7 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:13<00:28,  2.35it/s, est. speed input: 434.32 toks/s, output: 1400.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:14<00:25,  2.50it/s, est. speed input: 421.69 toks/s, output: 1672.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:15<00:26,  2.33it/s, est. speed input: 395.80 toks/s, output: 1772.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:16<00:26,  2.23it/s, est. speed input: 376.96 toks/s, output: 1895.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:17<00:28,  2.06it/s, est. speed input: 363.03 toks/s, output: 1928.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:17<00:25,  2.24it/s, est. speed input: 362.47 toks/s, output: 2028.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:19<00:35,  1.57it/s, est. speed input: 326.16 toks/s, output: 2020.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7626.7 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:20<00:37,  1.44it/s, est. speed input: 313.51 toks/s, output: 2045.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:21<00:35,  1.51it/s, est. speed input: 308.54 toks/s, output: 2117.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:21<00:34,  1.51it/s, est. speed input: 301.95 toks/s, output: 2175.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:22<00:34,  1.46it/s, est. speed input: 294.47 toks/s, output: 2224.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:23<00:33,  1.50it/s, est. speed input: 289.45 toks/s, output: 2288.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:23<00:31,  1.54it/s, est. speed input: 284.59 toks/s, output: 2352.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:26,  1.84it/s, est. speed input: 284.09 toks/s, output: 2449.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:25<00:28,  1.60it/s, est. speed input: 272.67 toks/s, output: 2548.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:18,  2.37it/s, est. speed input: 275.17 toks/s, output: 2773.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:26<00:21,  1.97it/s, est. speed input: 269.04 toks/s, output: 2811.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:28,  1.46it/s, est. speed input: 259.04 toks/s, output: 2807.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:28<00:25,  1.60it/s, est. speed input: 256.96 toks/s, output: 2886.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:29<00:28,  1.40it/s, est. speed input: 250.64 toks/s, output: 2914.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:29<00:21,  1.74it/s, est. speed input: 248.08 toks/s, output: 3083.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:18,  2.03it/s, est. speed input: 248.24 toks/s, output: 3184.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5905.5 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:31<00:23,  1.56it/s, est. speed input: 241.57 toks/s, output: 3198.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:31<00:22,  1.56it/s, est. speed input: 238.56 toks/s, output: 3258.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:32<00:17,  1.96it/s, est. speed input: 239.22 toks/s, output: 3366.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:32<00:17,  1.91it/s, est. speed input: 236.95 toks/s, output: 3433.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:21,  1.46it/s, est. speed input: 231.21 toks/s, output: 3448.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:33<00:12,  2.48it/s, est. speed input: 233.99 toks/s, output: 3686.89 toks/s]
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:10,  2.76it/s, est. speed input: 234.18 toks/s, output: 3787.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:10,  2.61it/s, est. speed input: 232.89 toks/s, output: 3864.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:34<00:08,  3.06it/s, est. speed input: 233.52 toks/s, output: 3971.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:35<00:11,  2.16it/s, est. speed input: 228.65 toks/s, output: 4079.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:13,  1.72it/s, est. speed input: 224.30 toks/s, output: 4100.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:38<00:16,  1.37it/s, est. speed input: 218.97 toks/s, output: 4103.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:13,  1.65it/s, est. speed input: 218.96 toks/s, output: 4203.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:38<00:12,  1.63it/s, est. speed input: 216.98 toks/s, output: 4265.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:10,  1.87it/s, est. speed input: 216.61 toks/s, output: 4358.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:40<00:11,  1.65it/s, est. speed input: 213.87 toks/s, output: 4403.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:42:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4167.1 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:40<00:11,  1.60it/s, est. speed input: 211.84 toks/s, output: 4461.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:41<00:10,  1.61it/s, est. speed input: 210.26 toks/s, output: 4528.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:15,  1.03it/s, est. speed input: 202.79 toks/s, output: 4470.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:45<00:21,  1.42s/it, est. speed input: 193.13 toks/s, output: 4363.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:46<00:16,  1.16s/it, est. speed input: 192.08 toks/s, output: 4445.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:47<00:17,  1.34s/it, est. speed input: 186.28 toks/s, output: 4418.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:49<00:15,  1.26s/it, est. speed input: 183.42 toks/s, output: 4460.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:49<00:07,  1.25it/s, est. speed input: 183.95 toks/s, output: 4689.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:49<00:06,  1.42it/s, est. speed input: 183.61 toks/s, output: 4790.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:43:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2607.5 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 74.9%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:51<00:07,  1.14it/s, est. speed input: 179.84 toks/s, output: 4801.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:52<00:05,  1.18it/s, est. speed input: 178.46 toks/s, output: 4872.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:09,  1.51s/it, est. speed input: 169.16 toks/s, output: 4732.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:55<00:05,  1.17s/it, est. speed input: 169.23 toks/s, output: 4849.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:56<00:03,  1.01it/s, est. speed input: 168.66 toks/s, output: 4948.32 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.01it/s, est. speed input: 172.94 toks/s, output: 5531.32 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.85it/s, est. speed input: 172.94 toks/s, output: 5531.32 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:43:15 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:43:15 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:43:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:43:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:43:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:43:15 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:43:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:43:15 [abstract.py:306] It took 0.534221 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:43:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:43:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:43:15 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.3688
+  • Mean Generation Length: 1942.9938
+  • Training FLOPS: 42.23 TFLOPS (10.56 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 260.60s
+  • policy_training: 166.17s (63.8%)
+  • generation: 59.66s (22.9%)
+  • teacher_logprob_inference: 29.64s (11.4%)
+  • policy_training/sharding_data: 4.34s (1.7%)
+  • prepare_for_generation: 3.70s (1.4%)
+  • policy_training/submit_training_futures: 1.80s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.6%)
+  • training_prep: 0.83s (0.3%)
+  • teacher_logprob_inference_prep: 0.42s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 145: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:46:33 [abstract.py:324] It took 0.049758 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08694844e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69538918e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79765625e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08694844e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69538918e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79765625e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08694844e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69538918e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79765625e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08694844e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69538918e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79765625e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08700625e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70496368e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79811401e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.04679241e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.83841801e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #145] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43617050e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08700625e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70496368e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79811401e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.67524109e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.25315952e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #145] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42642822e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08700625e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70496368e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79811401e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40416870e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.31289005e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #145] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.10106316e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08700625e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70496368e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79811401e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.13884201e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57416649e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #145] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34122360e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 145: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:46:36 [abstract.py:324] It took 0.029088 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12957.63it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:46:36 [loggers.py:236] Engine 000: Avg prompt throughput: 46.1 tokens/s, Avg generation throughput: 40.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:39,  4.04it/s, est. speed input: 242.55 toks/s, output: 52.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 23.17it/s, est. speed input: 1183.31 toks/s, output: 324.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 21.37it/s, est. speed input: 1180.43 toks/s, output: 451.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 26.07it/s, est. speed input: 1369.21 toks/s, output: 689.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 29.07it/s, est. speed input: 1494.70 toks/s, output: 908.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 34.73it/s, est. speed input: 1675.10 toks/s, output: 1209.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 35.33it/s, est. speed input: 1743.46 toks/s, output: 1411.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:03, 39.01it/s, est. speed input: 1854.82 toks/s, output: 1703.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 43.59it/s, est. speed input: 1981.69 toks/s, output: 2062.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 45.03it/s, est. speed input: 2068.98 toks/s, output: 2389.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 46.15it/s, est. speed input: 2145.70 toks/s, output: 2719.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 52.16it/s, est. speed input: 2277.00 toks/s, output: 3182.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 34.58it/s, est. speed input: 2096.08 toks/s, output: 3202.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:05<00:19,  4.73it/s, est. speed input: 774.69 toks/s, output: 1427.34 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:09<00:33,  2.63it/s, est. speed input: 484.62 toks/s, output: 1186.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:46:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10887.6 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:15<01:00,  1.39it/s, est. speed input: 301.15 toks/s, output: 989.80 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:16<00:54,  1.51it/s, est. speed input: 295.25 toks/s, output: 1160.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:17<00:50,  1.58it/s, est. speed input: 284.31 toks/s, output: 1306.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:46:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8705.5 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:20<01:07,  1.16it/s, est. speed input: 242.12 toks/s, output: 1284.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:21<01:02,  1.24it/s, est. speed input: 240.27 toks/s, output: 1367.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:21<00:54,  1.38it/s, est. speed input: 239.90 toks/s, output: 1456.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:21<00:46,  1.62it/s, est. speed input: 241.27 toks/s, output: 1554.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:22<00:54,  1.36it/s, est. speed input: 231.42 toks/s, output: 1580.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:23<00:58,  1.26it/s, est. speed input: 224.13 toks/s, output: 1620.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:24<00:50,  1.43it/s, est. speed input: 222.87 toks/s, output: 1699.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:24<00:42,  1.68it/s, est. speed input: 222.66 toks/s, output: 1785.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:24<00:41,  1.68it/s, est. speed input: 219.82 toks/s, output: 1849.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:27<01:10,  1.03s/it, est. speed input: 204.68 toks/s, output: 1809.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:27<00:56,  1.21it/s, est. speed input: 204.49 toks/s, output: 1893.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:28<01:09,  1.04s/it, est. speed input: 195.58 toks/s, output: 1897.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:29<01:02,  1.06it/s, est. speed input: 192.93 toks/s, output: 1956.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7141.1 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:30<01:00,  1.07it/s, est. speed input: 189.20 toks/s, output: 2003.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:31<01:05,  1.02s/it, est. speed input: 183.82 toks/s, output: 2031.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:32<00:54,  1.16it/s, est. speed input: 182.96 toks/s, output: 2107.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:32<00:39,  1.57it/s, est. speed input: 184.13 toks/s, output: 2205.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:32<00:33,  1.80it/s, est. speed input: 183.96 toks/s, output: 2286.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:34<00:55,  1.08it/s, est. speed input: 176.10 toks/s, output: 2272.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:34<00:41,  1.43it/s, est. speed input: 176.99 toks/s, output: 2367.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:35<00:21,  2.66it/s, est. speed input: 180.22 toks/s, output: 2655.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:35<00:21,  2.53it/s, est. speed input: 179.52 toks/s, output: 2726.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:36<00:25,  2.14it/s, est. speed input: 177.72 toks/s, output: 2779.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:37<00:31,  1.71it/s, est. speed input: 174.87 toks/s, output: 2814.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:37<00:19,  2.68it/s, est. speed input: 177.44 toks/s, output: 3014.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:39<00:40,  1.23it/s, est. speed input: 168.59 toks/s, output: 2945.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6268.3 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:40<00:36,  1.34it/s, est. speed input: 167.88 toks/s, output: 3012.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:40<00:28,  1.71it/s, est. speed input: 168.77 toks/s, output: 3109.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:40<00:25,  1.83it/s, est. speed input: 168.41 toks/s, output: 3182.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:40<00:20,  2.28it/s, est. speed input: 169.21 toks/s, output: 3276.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:41<00:16,  2.69it/s, est. speed input: 169.84 toks/s, output: 3366.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:41<00:13,  3.38it/s, est. speed input: 170.87 toks/s, output: 3464.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:41<00:13,  3.23it/s, est. speed input: 171.14 toks/s, output: 3622.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:42<00:11,  3.59it/s, est. speed input: 172.18 toks/s, output: 3796.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:11,  3.34it/s, est. speed input: 172.10 toks/s, output: 3870.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:10,  3.51it/s, est. speed input: 172.56 toks/s, output: 3956.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:10,  3.63it/s, est. speed input: 172.97 toks/s, output: 4041.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:11,  3.01it/s, est. speed input: 172.43 toks/s, output: 4103.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:12,  2.73it/s, est. speed input: 171.98 toks/s, output: 4169.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:14,  2.31it/s, est. speed input: 171.06 toks/s, output: 4221.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:12,  2.62it/s, est. speed input: 171.44 toks/s, output: 4305.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:14,  2.20it/s, est. speed input: 170.40 toks/s, output: 4354.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:12,  2.50it/s, est. speed input: 170.53 toks/s, output: 4507.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:47<00:14,  1.98it/s, est. speed input: 168.76 toks/s, output: 4537.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:11,  2.46it/s, est. speed input: 169.60 toks/s, output: 4634.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:11,  2.45it/s, est. speed input: 169.44 toks/s, output: 4703.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:16,  1.62it/s, est. speed input: 166.70 toks/s, output: 4702.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:49<00:11,  2.09it/s, est. speed input: 167.04 toks/s, output: 4863.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:49<00:09,  2.52it/s, est. speed input: 167.80 toks/s, output: 4960.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4376.5 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:18,  1.18it/s, est. speed input: 161.98 toks/s, output: 4865.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:52<00:16,  1.26it/s, est. speed input: 161.12 toks/s, output: 4918.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:52<00:13,  1.51it/s, est. speed input: 161.31 toks/s, output: 5001.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:53<00:10,  1.75it/s, est. speed input: 161.42 toks/s, output: 5083.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:08,  2.05it/s, est. speed input: 161.65 toks/s, output: 5169.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:53<00:07,  2.13it/s, est. speed input: 161.47 toks/s, output: 5243.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:07,  2.08it/s, est. speed input: 161.05 toks/s, output: 5308.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:55<00:10,  1.41it/s, est. speed input: 158.51 toks/s, output: 5304.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:56<00:09,  1.49it/s, est. speed input: 157.99 toks/s, output: 5366.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:56<00:07,  1.65it/s, est. speed input: 157.83 toks/s, output: 5441.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:56<00:06,  1.95it/s, est. speed input: 158.07 toks/s, output: 5530.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:58<00:09,  1.11it/s, est. speed input: 154.21 toks/s, output: 5479.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:59<00:08,  1.12it/s, est. speed input: 152.95 toks/s, output: 5520.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:59<00:06,  1.38it/s, est. speed input: 153.13 toks/s, output: 5611.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2762.3 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:00<00:06,  1.29it/s, est. speed input: 151.84 toks/s, output: 5650.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:01<00:06,  1.16it/s, est. speed input: 150.20 toks/s, output: 5678.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:02<00:04,  1.33it/s, est. speed input: 149.97 toks/s, output: 5758.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:02<00:03,  1.55it/s, est. speed input: 149.97 toks/s, output: 5847.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:04<00:03,  1.15it/s, est. speed input: 147.64 toks/s, output: 5847.46 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.15it/s, est. speed input: 151.38 toks/s, output: 6358.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.50it/s, est. speed input: 151.38 toks/s, output: 6358.62 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:44 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:44 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:47:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:47:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:47:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:47:44 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:47:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:47:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:47:45 [abstract.py:306] It took 0.546100 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:47:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:47:45 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0101
+  • Training Accuracy: 0.3563
+  • Mean Generation Length: 2547.5562
+  • Training FLOPS: 45.49 TFLOPS (11.37 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 310.16s
+  • policy_training: 201.51s (65.0%)
+  • generation: 68.36s (22.0%)
+  • teacher_logprob_inference: 34.99s (11.3%)
+  • policy_training/sharding_data: 4.37s (1.4%)
+  • prepare_for_generation: 3.63s (1.2%)
+  • policy_training/submit_training_futures: 1.97s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.5%)
+  • training_prep: 0.90s (0.3%)
+  • teacher_logprob_inference_prep: 0.60s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Epoch 30/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 146: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:51:44 [abstract.py:324] It took 0.048335 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08712461e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70132523e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80552750e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08712461e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70132523e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80552750e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08712461e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70132523e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80552750e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08712461e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70132523e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80552750e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08714297e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.72099533e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82834015e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.05410690e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.82594395e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #146] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43472168e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08714297e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.72099533e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82834015e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64497375e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.25814199e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #146] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.41302795e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08714297e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.72099533e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82834015e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39735794e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.10148191e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #146] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11931610e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08714297e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.72099533e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82834015e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14158478e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57695484e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #146] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33849411e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 146: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:51:47 [abstract.py:324] It took 0.032020 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13315.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:51:47 [loggers.py:236] Engine 000: Avg prompt throughput: 38.9 tokens/s, Avg generation throughput: 24.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:54,  2.91it/s, est. speed input: 174.43 toks/s, output: 29.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:11, 13.20it/s, est. speed input: 659.84 toks/s, output: 167.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 24.99it/s, est. speed input: 1175.67 toks/s, output: 412.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 29.02it/s, est. speed input: 1360.74 toks/s, output: 582.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 34.94it/s, est. speed input: 1582.40 toks/s, output: 817.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 39.25it/s, est. speed input: 1778.19 toks/s, output: 1086.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 41.57it/s, est. speed input: 1895.57 toks/s, output: 1338.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 45.29it/s, est. speed input: 2032.17 toks/s, output: 1645.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 48.18it/s, est. speed input: 2149.84 toks/s, output: 1964.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 48.61it/s, est. speed input: 2229.08 toks/s, output: 2272.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 49.01it/s, est. speed input: 2295.41 toks/s, output: 2583.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:02, 46.02it/s, est. speed input: 2305.07 toks/s, output: 2809.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 48.67it/s, est. speed input: 2372.59 toks/s, output: 3165.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:02, 42.78it/s, est. speed input: 2336.50 toks/s, output: 3355.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:02<00:04, 16.98it/s, est. speed input: 1785.92 toks/s, output: 2832.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:04<00:12,  6.16it/s, est. speed input: 1080.71 toks/s, output: 1993.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:05<00:14,  5.39it/s, est. speed input: 950.28 toks/s, output: 1996.08 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:06<00:15,  4.70it/s, est. speed input: 860.32 toks/s, output: 1994.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:06<00:17,  4.25it/s, est. speed input: 793.53 toks/s, output: 2026.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:07<00:20,  3.54it/s, est. speed input: 715.09 toks/s, output: 2008.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:08<00:22,  3.03it/s, est. speed input: 647.94 toks/s, output: 2002.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:51:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10091.9 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:10<00:32,  2.03it/s, est. speed input: 539.89 toks/s, output: 1849.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:10<00:31,  2.12it/s, est. speed input: 527.28 toks/s, output: 1906.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:13<00:58,  1.11it/s, est. speed input: 422.26 toks/s, output: 1631.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:14<00:59,  1.07it/s, est. speed input: 395.54 toks/s, output: 1632.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:15<00:50,  1.24it/s, est. speed input: 390.20 toks/s, output: 1714.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:17<01:07,  1.08s/it, est. speed input: 349.61 toks/s, output: 1641.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:47,  1.27it/s, est. speed input: 342.29 toks/s, output: 1812.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:18<00:44,  1.32it/s, est. speed input: 333.41 toks/s, output: 1869.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:41,  1.41it/s, est. speed input: 326.76 toks/s, output: 1934.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:32,  1.75it/s, est. speed input: 327.05 toks/s, output: 2039.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:19<00:30,  1.83it/s, est. speed input: 322.12 toks/s, output: 2110.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7802.7 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:20<00:37,  1.48it/s, est. speed input: 309.47 toks/s, output: 2129.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:21<00:42,  1.27it/s, est. speed input: 297.12 toks/s, output: 2146.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:22<00:36,  1.45it/s, est. speed input: 293.80 toks/s, output: 2224.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:22<00:32,  1.61it/s, est. speed input: 290.64 toks/s, output: 2302.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:22<00:23,  2.13it/s, est. speed input: 291.90 toks/s, output: 2412.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:23<00:25,  1.96it/s, est. speed input: 287.00 toks/s, output: 2472.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:23<00:21,  2.30it/s, est. speed input: 286.45 toks/s, output: 2567.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:23<00:12,  3.87it/s, est. speed input: 290.36 toks/s, output: 2800.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:24<00:14,  3.12it/s, est. speed input: 285.45 toks/s, output: 2945.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:16,  2.71it/s, est. speed input: 281.68 toks/s, output: 3004.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:14,  2.96it/s, est. speed input: 280.16 toks/s, output: 3178.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:26<00:11,  3.46it/s, est. speed input: 280.48 toks/s, output: 3372.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:27<00:21,  1.84it/s, est. speed input: 267.09 toks/s, output: 3308.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:28<00:16,  2.27it/s, est. speed input: 266.40 toks/s, output: 3490.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:28<00:16,  2.18it/s, est. speed input: 263.54 toks/s, output: 3550.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:13,  2.64it/s, est. speed input: 264.59 toks/s, output: 3660.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:30<00:19,  1.71it/s, est. speed input: 255.90 toks/s, output: 3637.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5757.9 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:17,  1.92it/s, est. speed input: 255.03 toks/s, output: 3721.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:09,  3.03it/s, est. speed input: 256.97 toks/s, output: 4032.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:31<00:09,  3.17it/s, est. speed input: 256.76 toks/s, output: 4123.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:31<00:12,  2.32it/s, est. speed input: 251.98 toks/s, output: 4141.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:32<00:14,  1.91it/s, est. speed input: 247.54 toks/s, output: 4164.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:33<00:09,  2.59it/s, est. speed input: 248.53 toks/s, output: 4369.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:33<00:09,  2.58it/s, est. speed input: 247.43 toks/s, output: 4444.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:33<00:07,  3.13it/s, est. speed input: 248.43 toks/s, output: 4557.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:34<00:12,  1.82it/s, est. speed input: 241.51 toks/s, output: 4525.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:35<00:09,  2.18it/s, est. speed input: 241.77 toks/s, output: 4626.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:35<00:07,  2.73it/s, est. speed input: 242.62 toks/s, output: 4737.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:35<00:06,  2.86it/s, est. speed input: 242.21 toks/s, output: 4823.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:36<00:08,  2.01it/s, est. speed input: 238.17 toks/s, output: 4839.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:37<00:09,  1.83it/s, est. speed input: 235.51 toks/s, output: 4881.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:38<00:12,  1.29it/s, est. speed input: 228.93 toks/s, output: 4843.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:39<00:13,  1.14it/s, est. speed input: 224.03 toks/s, output: 4840.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:40<00:11,  1.25it/s, est. speed input: 222.04 toks/s, output: 4898.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3498.6 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:40<00:09,  1.37it/s, est. speed input: 220.44 toks/s, output: 4965.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:41<00:05,  1.98it/s, est. speed input: 220.83 toks/s, output: 5175.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:41<00:04,  2.11it/s, est. speed input: 220.25 toks/s, output: 5263.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:42<00:03,  2.10it/s, est. speed input: 218.14 toks/s, output: 5414.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:43<00:03,  2.08it/s, est. speed input: 217.01 toks/s, output: 5491.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:46<00:07,  1.20s/it, est. speed input: 202.41 toks/s, output: 5233.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:47<00:05,  1.15s/it, est. speed input: 199.35 toks/s, output: 5268.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:47<00:03,  1.04it/s, est. speed input: 198.72 toks/s, output: 5365.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:48<00:01,  1.29it/s, est. speed input: 196.90 toks/s, output: 5542.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:49<00:00,  1.31it/s, est. speed input: 195.25 toks/s, output: 5612.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1438.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.10it/s, est. speed input: 191.36 toks/s, output: 5621.69 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.10it/s, est. speed input: 191.36 toks/s, output: 5621.69 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  3.14it/s, est. speed input: 191.36 toks/s, output: 5621.69 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:40 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:40 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:52:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:52:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:52:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:52:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:52:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:52:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:52:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:52:41 [abstract.py:306] It took 0.780188 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:52:41 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.4875
+  • Mean Generation Length: 1792.0687
+  • Training FLOPS: 39.49 TFLOPS (9.87 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 248.46s
+  • policy_training: 161.67s (65.1%)
+  • generation: 54.47s (21.9%)
+  • teacher_logprob_inference: 27.19s (10.9%)
+  • policy_training/sharding_data: 4.65s (1.9%)
+  • prepare_for_generation: 3.66s (1.5%)
+  • policy_training/submit_training_futures: 1.88s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.6%)
+  • training_prep: 0.79s (0.3%)
+  • teacher_logprob_inference_prep: 0.50s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 147: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:55:52 [abstract.py:324] It took 0.049486 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08745078e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70542488e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80999527e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08745078e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70542488e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80999527e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08745078e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70542488e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80999527e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08745078e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.70542488e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80999527e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08759629e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.73303528e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82865601e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.05918121e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.90896511e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #147] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43979950e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08759629e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.73303528e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82865601e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64619827e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26148205e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #147] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39991302e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08759629e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.73303528e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82865601e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39045715e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.94700956e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #147] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11519814e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08759629e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.73303528e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82865601e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15273094e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58485909e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #147] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33508545e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 147: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:55:55 [abstract.py:324] It took 0.029946 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13155.25it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:55:55 [loggers.py:236] Engine 000: Avg prompt throughput: 49.3 tokens/s, Avg generation throughput: 1.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:25,  6.12it/s, est. speed input: 361.38 toks/s, output: 79.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:02, 57.51it/s, est. speed input: 2954.12 toks/s, output: 893.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:02, 60.27it/s, est. speed input: 3262.24 toks/s, output: 1124.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:02, 48.92it/s, est. speed input: 2957.10 toks/s, output: 1276.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:02, 53.09it/s, est. speed input: 3098.01 toks/s, output: 1656.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:00<00:01, 60.36it/s, est. speed input: 3309.12 toks/s, output: 2160.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:00<00:01, 58.88it/s, est. speed input: 3328.60 toks/s, output: 2516.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 77.82it/s, est. speed input: 3784.28 toks/s, output: 3488.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:01, 73.64it/s, est. speed input: 3810.77 toks/s, output: 3937.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:01<00:01, 73.48it/s, est. speed input: 3864.92 toks/s, output: 4389.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:01<00:00, 71.64it/s, est. speed input: 3899.88 toks/s, output: 4932.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:01<00:01, 38.87it/s, est. speed input: 3211.47 toks/s, output: 4507.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:02<00:02, 21.32it/s, est. speed input: 2462.80 toks/s, output: 3872.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:05<00:09,  5.43it/s, est. speed input: 1102.56 toks/s, output: 2104.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:07<00:10,  4.68it/s, est. speed input: 955.00 toks/s, output: 2187.49 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:09<00:15,  2.93it/s, est. speed input: 707.02 toks/s, output: 1977.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8673.5 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:12<00:20,  2.12it/s, est. speed input: 576.81 toks/s, output: 1850.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:13<00:19,  2.10it/s, est. speed input: 542.65 toks/s, output: 1997.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:13<00:19,  2.12it/s, est. speed input: 529.91 toks/s, output: 2079.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:14<00:19,  2.08it/s, est. speed input: 514.41 toks/s, output: 2148.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:14<00:17,  2.27it/s, est. speed input: 510.48 toks/s, output: 2260.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:14<00:13,  2.84it/s, est. speed input: 507.24 toks/s, output: 2500.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:15<00:11,  3.14it/s, est. speed input: 505.46 toks/s, output: 2618.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:15<00:13,  2.63it/s, est. speed input: 489.37 toks/s, output: 2661.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:16<00:16,  2.04it/s, est. speed input: 467.72 toks/s, output: 2669.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:17<00:20,  1.60it/s, est. speed input: 443.47 toks/s, output: 2658.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:19<00:22,  1.35it/s, est. speed input: 408.36 toks/s, output: 2693.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5750.8 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:20<00:20,  1.38it/s, est. speed input: 386.80 toks/s, output: 2801.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:21<00:18,  1.53it/s, est. speed input: 382.15 toks/s, output: 2897.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:21<00:12,  2.15it/s, est. speed input: 382.66 toks/s, output: 3154.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:22<00:14,  1.78it/s, est. speed input: 369.58 toks/s, output: 3174.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:22<00:11,  2.12it/s, est. speed input: 369.25 toks/s, output: 3299.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:22<00:10,  2.24it/s, est. speed input: 365.95 toks/s, output: 3395.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:23<00:08,  2.67it/s, est. speed input: 365.91 toks/s, output: 3520.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:23<00:06,  3.20it/s, est. speed input: 366.24 toks/s, output: 3649.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:23<00:05,  3.77it/s, est. speed input: 366.60 toks/s, output: 3777.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:23<00:04,  4.28it/s, est. speed input: 365.71 toks/s, output: 4015.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:24<00:07,  2.39it/s, est. speed input: 353.31 toks/s, output: 4004.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:25<00:09,  1.77it/s, est. speed input: 342.08 toks/s, output: 4005.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:25<00:06,  2.18it/s, est. speed input: 342.11 toks/s, output: 4132.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:26<00:05,  2.71it/s, est. speed input: 342.61 toks/s, output: 4264.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:26<00:03,  3.38it/s, est. speed input: 343.43 toks/s, output: 4401.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:26<00:03,  3.93it/s, est. speed input: 343.68 toks/s, output: 4530.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:27<00:04,  2.41it/s, est. speed input: 335.66 toks/s, output: 4551.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:27<00:03,  2.62it/s, est. speed input: 334.19 toks/s, output: 4658.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:27<00:02,  3.19it/s, est. speed input: 334.51 toks/s, output: 4789.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:27<00:01,  3.86it/s, est. speed input: 334.19 toks/s, output: 5036.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:29<00:03,  1.53it/s, est. speed input: 315.29 toks/s, output: 4883.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3053.8 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:31<00:04,  1.18it/s, est. speed input: 303.19 toks/s, output: 4832.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:34<00:05,  1.43s/it, est. speed input: 278.36 toks/s, output: 4581.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:34<00:03,  1.14s/it, est. speed input: 277.01 toks/s, output: 4703.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:35<00:01,  1.05it/s, est. speed input: 274.83 toks/s, output: 4812.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:35<00:00,  1.40it/s, est. speed input: 275.64 toks/s, output: 4971.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 747.2 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:42<00:00,  2.60s/it, est. speed input: 230.53 toks/s, output: 4324.49 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:42<00:00,  2.60s/it, est. speed input: 230.53 toks/s, output: 4324.49 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:42<00:00,  3.78it/s, est. speed input: 230.53 toks/s, output: 4324.49 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:39 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:39 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 10:56:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:56:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:56:39 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:56:40 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 10:56:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 10:56:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 10:56:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 10:56:40 [abstract.py:306] It took 0.649727 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 10:56:40 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0163
+  • Training Accuracy: 0.4188
+  • Mean Generation Length: 1144.3000
+  • Training FLOPS: 32.88 TFLOPS (8.22 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 193.97s
+  • policy_training: 124.39s (64.1%)
+  • generation: 44.71s (23.0%)
+  • teacher_logprob_inference: 19.74s (10.2%)
+  • policy_training/sharding_data: 4.33s (2.2%)
+  • prepare_for_generation: 3.57s (1.8%)
+  • policy_training/submit_training_futures: 2.10s (1.1%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.8%)
+  • training_prep: 0.93s (0.5%)
+  • teacher_logprob_inference_prep: 0.45s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 148: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:07 [abstract.py:324] It took 0.048214 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08774531e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69856186e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81502533e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08774531e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69856186e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81502533e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08774531e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69856186e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81502533e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08774531e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69856186e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81502533e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08776562e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70977478e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82249146e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.05921364e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.96333218e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #148] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44177094e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08776562e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70977478e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82249146e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64297180e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.26676788e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #148] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40188675e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08776562e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70977478e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82249146e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40638504e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.55682087e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #148] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.11264725e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08776562e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70977478e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82249146e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15805511e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58773136e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #148] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33677460e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 148: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:10 [abstract.py:324] It took 0.029448 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14004.65it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:10 [loggers.py:236] Engine 000: Avg prompt throughput: 62.6 tokens/s, Avg generation throughput: 5.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:21,  7.35it/s, est. speed input: 448.67 toks/s, output: 73.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:05, 28.53it/s, est. speed input: 1545.73 toks/s, output: 388.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 23.90it/s, est. speed input: 1415.68 toks/s, output: 461.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 30.77it/s, est. speed input: 1676.37 toks/s, output: 775.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 36.36it/s, est. speed input: 1900.82 toks/s, output: 1105.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 34.58it/s, est. speed input: 1899.85 toks/s, output: 1293.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 46.15it/s, est. speed input: 2214.24 toks/s, output: 1843.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 42.70it/s, est. speed input: 2216.06 toks/s, output: 2048.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 49.94it/s, est. speed input: 2425.16 toks/s, output: 2673.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 51.20it/s, est. speed input: 2510.45 toks/s, output: 3076.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:02, 44.99it/s, est. speed input: 2458.83 toks/s, output: 3279.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 42.26it/s, est. speed input: 2435.55 toks/s, output: 3490.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:02, 43.41it/s, est. speed input: 2462.83 toks/s, output: 3790.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:03<00:08,  9.35it/s, est. speed input: 1356.78 toks/s, output: 2299.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:07<00:27,  2.88it/s, est. speed input: 617.10 toks/s, output: 1333.57 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10402.4 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:10<00:35,  2.17it/s, est. speed input: 475.49 toks/s, output: 1326.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:11<00:35,  2.11it/s, est. speed input: 442.42 toks/s, output: 1436.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:12<00:31,  2.28it/s, est. speed input: 431.81 toks/s, output: 1600.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:31,  2.26it/s, est. speed input: 420.56 toks/s, output: 1656.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:13<00:37,  1.90it/s, est. speed input: 393.45 toks/s, output: 1648.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:14<00:35,  1.97it/s, est. speed input: 386.30 toks/s, output: 1715.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:14<00:36,  1.91it/s, est. speed input: 374.72 toks/s, output: 1762.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:16<00:55,  1.22it/s, est. speed input: 336.45 toks/s, output: 1680.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:17<00:38,  1.70it/s, est. speed input: 334.57 toks/s, output: 1863.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:41,  1.59it/s, est. speed input: 323.21 toks/s, output: 1896.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:19<00:57,  1.12it/s, est. speed input: 297.40 toks/s, output: 1842.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8017.6 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:20<00:47,  1.31it/s, est. speed input: 286.69 toks/s, output: 1965.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:21<00:40,  1.51it/s, est. speed input: 285.35 toks/s, output: 2051.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:21<00:35,  1.69it/s, est. speed input: 283.29 toks/s, output: 2131.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:21<00:29,  2.00it/s, est. speed input: 283.04 toks/s, output: 2224.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<00:24,  2.34it/s, est. speed input: 282.76 toks/s, output: 2316.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:22<00:19,  2.88it/s, est. speed input: 283.66 toks/s, output: 2417.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:25<01:01,  1.11s/it, est. speed input: 251.75 toks/s, output: 2241.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:25<00:50,  1.08it/s, est. speed input: 249.41 toks/s, output: 2315.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:30,  1.73it/s, est. speed input: 250.99 toks/s, output: 2518.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:26<00:25,  2.08it/s, est. speed input: 251.53 toks/s, output: 2617.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:30,  1.67it/s, est. speed input: 245.12 toks/s, output: 2643.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:27<00:29,  1.70it/s, est. speed input: 242.28 toks/s, output: 2705.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:28<00:34,  1.43it/s, est. speed input: 236.01 toks/s, output: 2728.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:28<00:27,  1.72it/s, est. speed input: 235.77 toks/s, output: 2817.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:29<00:23,  1.98it/s, est. speed input: 235.25 toks/s, output: 2903.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6497.6 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:30<00:41,  1.11it/s, est. speed input: 223.18 toks/s, output: 2846.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:31<00:23,  1.89it/s, est. speed input: 225.85 toks/s, output: 3064.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:31<00:17,  2.42it/s, est. speed input: 226.35 toks/s, output: 3252.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:32<00:22,  1.80it/s, est. speed input: 220.75 toks/s, output: 3263.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:33<00:23,  1.68it/s, est. speed input: 217.76 toks/s, output: 3311.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:33<00:16,  2.32it/s, est. speed input: 218.83 toks/s, output: 3509.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:34<00:19,  1.93it/s, est. speed input: 215.43 toks/s, output: 3544.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:34<00:17,  2.08it/s, est. speed input: 214.90 toks/s, output: 3625.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:35<00:15,  2.24it/s, est. speed input: 214.44 toks/s, output: 3708.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:36<00:19,  1.74it/s, est. speed input: 210.64 toks/s, output: 3733.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:36<00:16,  2.04it/s, est. speed input: 210.69 toks/s, output: 3824.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:38<00:22,  1.38it/s, est. speed input: 202.68 toks/s, output: 3855.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:38<00:18,  1.60it/s, est. speed input: 202.70 toks/s, output: 3944.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:39<00:14,  1.92it/s, est. speed input: 202.02 toks/s, output: 4108.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 10:59:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4858.1 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:20,  1.29it/s, est. speed input: 195.31 toks/s, output: 4064.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:41<00:16,  1.56it/s, est. speed input: 195.76 toks/s, output: 4164.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:43<00:26,  1.07s/it, est. speed input: 186.87 toks/s, output: 4068.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:43<00:19,  1.21it/s, est. speed input: 187.52 toks/s, output: 4176.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:44<00:12,  1.76it/s, est. speed input: 188.18 toks/s, output: 4377.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:45<00:10,  1.88it/s, est. speed input: 186.97 toks/s, output: 4531.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:45<00:08,  2.17it/s, est. speed input: 187.43 toks/s, output: 4635.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:45<00:06,  2.60it/s, est. speed input: 188.18 toks/s, output: 4746.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:46<00:07,  2.31it/s, est. speed input: 187.11 toks/s, output: 4812.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:46<00:07,  2.15it/s, est. speed input: 186.17 toks/s, output: 4880.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:17,  1.14s/it, est. speed input: 176.38 toks/s, output: 4720.89 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:00:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3368.1 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:13,  1.01it/s, est. speed input: 175.40 toks/s, output: 4792.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:51<00:11,  1.11it/s, est. speed input: 174.24 toks/s, output: 4859.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:51<00:08,  1.41it/s, est. speed input: 174.63 toks/s, output: 4967.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:52<00:10,  1.04it/s, est. speed input: 170.62 toks/s, output: 4951.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:53<00:07,  1.33it/s, est. speed input: 170.92 toks/s, output: 5060.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:53<00:07,  1.27it/s, est. speed input: 169.25 toks/s, output: 5111.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:54<00:05,  1.47it/s, est. speed input: 169.01 toks/s, output: 5204.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:05,  1.27it/s, est. speed input: 166.89 toks/s, output: 5242.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:06,  1.11s/it, est. speed input: 162.45 toks/s, output: 5209.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:58<00:05,  1.15s/it, est. speed input: 160.07 toks/s, output: 5239.97 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.15s/it, est. speed input: 165.16 toks/s, output: 5939.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.73it/s, est. speed input: 165.16 toks/s, output: 5939.40 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:00:11 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:00:11 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:00:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:00:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:00:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:00:12 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:00:12 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:00:12 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:00:12 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:00:12 [abstract.py:306] It took 0.595056 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:00:12 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0121
+  • Training Accuracy: 0.4313
+  • Mean Generation Length: 2173.8374
+  • Training FLOPS: 42.89 TFLOPS (10.72 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 285.75s
+  • policy_training: 182.67s (63.9%)
+  • generation: 62.35s (21.8%)
+  • teacher_logprob_inference: 35.27s (12.3%)
+  • policy_training/sharding_data: 4.31s (1.5%)
+  • prepare_for_generation: 3.73s (1.3%)
+  • policy_training/submit_training_futures: 1.81s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.64s (0.6%)
+  • training_prep: 0.98s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 149: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:03:52 [abstract.py:324] It took 0.055027 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08816133e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69470406e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08816133e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69470406e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08816133e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69470406e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80918350e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08816133e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69470406e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80918350e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80918350e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80918350e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08818301e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71809425e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81913986e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07676010e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.13712692e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #149] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44320129e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08818301e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71809425e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81913986e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64413757e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.27534218e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #149] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40133972e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08818301e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71809425e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81913986e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40679321e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.45729160e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #149] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.09448242e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08818301e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71809425e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81913986e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16166267e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59848166e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #149] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.33997360e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 149: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:03:55 [abstract.py:324] It took 0.028612 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13526.47it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:03:56 [loggers.py:236] Engine 000: Avg prompt throughput: 41.1 tokens/s, Avg generation throughput: 69.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:33,  1.71it/s, est. speed input: 100.63 toks/s, output: 20.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:12, 12.23it/s, est. speed input: 594.22 toks/s, output: 168.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 26.16it/s, est. speed input: 1113.53 toks/s, output: 430.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 24.81it/s, est. speed input: 1190.11 toks/s, output: 580.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:04, 28.44it/s, est. speed input: 1333.08 toks/s, output: 839.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:05, 24.80it/s, est. speed input: 1300.33 toks/s, output: 968.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:03, 32.66it/s, est. speed input: 1482.89 toks/s, output: 1369.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 41.65it/s, est. speed input: 1673.44 toks/s, output: 1837.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 41.96it/s, est. speed input: 1745.35 toks/s, output: 2125.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 39.12it/s, est. speed input: 1766.04 toks/s, output: 2328.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:02<00:02, 36.60it/s, est. speed input: 1775.79 toks/s, output: 2549.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:02<00:03, 28.65it/s, est. speed input: 1699.44 toks/s, output: 2652.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:06<00:23,  3.77it/s, est. speed input: 682.10 toks/s, output: 1241.83 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:08<00:33,  2.60it/s, est. speed input: 507.39 toks/s, output: 1162.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:09<00:34,  2.45it/s, est. speed input: 465.60 toks/s, output: 1253.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:04:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10735.8 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:14<01:02,  1.32it/s, est. speed input: 329.34 toks/s, output: 1058.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:16<01:10,  1.16it/s, est. speed input: 299.05 toks/s, output: 1056.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:17<01:12,  1.10it/s, est. speed input: 282.21 toks/s, output: 1091.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:17<01:03,  1.24it/s, est. speed input: 280.63 toks/s, output: 1179.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:18<01:04,  1.21it/s, est. speed input: 270.09 toks/s, output: 1228.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:19<01:01,  1.25it/s, est. speed input: 263.35 toks/s, output: 1291.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:19<00:59,  1.28it/s, est. speed input: 256.60 toks/s, output: 1350.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:04:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8562.6 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:20<00:54,  1.37it/s, est. speed input: 252.35 toks/s, output: 1419.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:20<00:44,  1.65it/s, est. speed input: 251.92 toks/s, output: 1507.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:21<00:47,  1.55it/s, est. speed input: 245.89 toks/s, output: 1561.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:22<01:02,  1.16it/s, est. speed input: 233.43 toks/s, output: 1571.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:23<00:43,  1.61it/s, est. speed input: 232.26 toks/s, output: 1738.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:25<01:02,  1.10it/s, est. speed input: 217.97 toks/s, output: 1720.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:25<00:48,  1.39it/s, est. speed input: 218.94 toks/s, output: 1815.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:26<00:50,  1.32it/s, est. speed input: 213.93 toks/s, output: 1860.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:28<01:05,  1.01it/s, est. speed input: 204.01 toks/s, output: 1860.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:28<00:50,  1.28it/s, est. speed input: 204.34 toks/s, output: 1948.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:29<00:53,  1.19it/s, est. speed input: 199.50 toks/s, output: 1988.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:29<00:50,  1.26it/s, est. speed input: 196.99 toks/s, output: 2048.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:04:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7052.0 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:30<00:36,  1.69it/s, est. speed input: 196.56 toks/s, output: 2210.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:31<00:32,  1.86it/s, est. speed input: 196.15 toks/s, output: 2289.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:31<00:19,  2.94it/s, est. speed input: 199.21 toks/s, output: 2491.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:31<00:16,  3.39it/s, est. speed input: 200.27 toks/s, output: 2585.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:31<00:14,  3.79it/s, est. speed input: 201.32 toks/s, output: 2760.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:32<00:24,  2.19it/s, est. speed input: 196.18 toks/s, output: 2770.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:33<00:28,  1.86it/s, est. speed input: 193.34 toks/s, output: 2811.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:34<00:36,  1.43it/s, est. speed input: 188.58 toks/s, output: 2822.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:35<00:28,  1.76it/s, est. speed input: 189.18 toks/s, output: 2912.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:35<00:25,  1.97it/s, est. speed input: 189.02 toks/s, output: 2990.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:35<00:19,  2.52it/s, est. speed input: 190.10 toks/s, output: 3086.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:35<00:14,  3.21it/s, est. speed input: 191.25 toks/s, output: 3183.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:37<00:25,  1.79it/s, est. speed input: 185.71 toks/s, output: 3246.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:37<00:22,  2.02it/s, est. speed input: 185.83 toks/s, output: 3327.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:38<00:22,  1.94it/s, est. speed input: 184.57 toks/s, output: 3384.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:39<00:22,  1.85it/s, est. speed input: 182.25 toks/s, output: 3498.32 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:04:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5803.9 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:40<00:30,  1.36it/s, est. speed input: 177.57 toks/s, output: 3488.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:41<00:27,  1.44it/s, est. speed input: 176.53 toks/s, output: 3547.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:41<00:23,  1.66it/s, est. speed input: 176.56 toks/s, output: 3628.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:22,  1.69it/s, est. speed input: 175.64 toks/s, output: 3688.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:19,  1.88it/s, est. speed input: 175.45 toks/s, output: 3764.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:42<00:15,  2.34it/s, est. speed input: 176.19 toks/s, output: 3859.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:26,  1.34it/s, est. speed input: 171.44 toks/s, output: 3835.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:23,  1.43it/s, est. speed input: 170.56 toks/s, output: 3895.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:26,  1.26it/s, est. speed input: 168.11 toks/s, output: 3919.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:46<00:28,  1.12it/s, est. speed input: 165.34 toks/s, output: 3936.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:48<00:32,  1.03s/it, est. speed input: 161.92 toks/s, output: 3936.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:48<00:23,  1.26it/s, est. speed input: 162.38 toks/s, output: 4028.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:48<00:18,  1.61it/s, est. speed input: 162.90 toks/s, output: 4121.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:49<00:14,  1.97it/s, est. speed input: 163.32 toks/s, output: 4212.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:49<00:15,  1.73it/s, est. speed input: 162.06 toks/s, output: 4262.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:50<00:15,  1.71it/s, est. speed input: 161.35 toks/s, output: 4323.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:04:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4337.2 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:14,  1.69it/s, est. speed input: 160.61 toks/s, output: 4384.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:51<00:15,  1.50it/s, est. speed input: 159.16 toks/s, output: 4426.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:53<00:21,  1.07it/s, est. speed input: 155.65 toks/s, output: 4411.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:54<00:14,  1.46it/s, est. speed input: 155.67 toks/s, output: 4574.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:54<00:10,  1.86it/s, est. speed input: 156.52 toks/s, output: 4681.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:54<00:08,  2.26it/s, est. speed input: 157.19 toks/s, output: 4781.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:54<00:05,  2.88it/s, est. speed input: 158.17 toks/s, output: 4972.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:56<00:08,  1.83it/s, est. speed input: 155.88 toks/s, output: 4982.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:56<00:07,  1.99it/s, est. speed input: 155.90 toks/s, output: 5066.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:57<00:08,  1.67it/s, est. speed input: 154.57 toks/s, output: 5106.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:57<00:07,  1.63it/s, est. speed input: 153.86 toks/s, output: 5166.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:58<00:06,  1.99it/s, est. speed input: 154.34 toks/s, output: 5266.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:58<00:04,  2.38it/s, est. speed input: 154.83 toks/s, output: 5365.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:58<00:04,  2.41it/s, est. speed input: 154.79 toks/s, output: 5448.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:59<00:03,  2.29it/s, est. speed input: 154.52 toks/s, output: 5523.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:59<00:03,  2.23it/s, est. speed input: 154.27 toks/s, output: 5599.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:04:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2824.1 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:02<00:04,  1.25it/s, est. speed input: 150.21 toks/s, output: 5621.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:04<00:05,  1.10s/it, est. speed input: 146.40 toks/s, output: 5571.30 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.10s/it, est. speed input: 151.09 toks/s, output: 6209.01 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.49it/s, est. speed input: 151.09 toks/s, output: 6209.01 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:05:03 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:05:03 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:05:04 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:05:04 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:05:04 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:05:04 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:05:04 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:05:04 [abstract.py:306] It took 0.668868 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:05:04 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:05:04 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:05:04 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0105
+  • Training Accuracy: 0.3438
+  • Mean Generation Length: 2492.4375
+  • Training FLOPS: 45.17 TFLOPS (11.29 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 309.07s
+  • policy_training: 198.96s (64.4%)
+  • generation: 68.56s (22.2%)
+  • teacher_logprob_inference: 36.45s (11.8%)
+  • policy_training/sharding_data: 4.31s (1.4%)
+  • prepare_for_generation: 3.74s (1.2%)
+  • policy_training/submit_training_futures: 1.85s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.71s (0.6%)
+  • training_prep: 0.79s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 150: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:01 [abstract.py:324] It took 0.047640 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08832988e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08832988e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08832988e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08832988e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08839102e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67678833e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81046677e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08839102e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67678833e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81046677e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08839102e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67678833e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81046677e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08839102e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67678833e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81046677e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68358498e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80368347e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68358498e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80368347e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68358498e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80368347e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68358498e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80368347e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07796669e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.22946978e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #150] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44778397e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62647858e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.29151440e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #150] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40847321e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41158829e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.37902737e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #150] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.13039932e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17425156e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59471607e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #150] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34420105e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 150: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:05 [abstract.py:324] It took 0.032981 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:05 [loggers.py:236] Engine 000: Avg prompt throughput: 39.0 tokens/s, Avg generation throughput: 22.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12297.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:28,  5.64it/s, est. speed input: 349.79 toks/s, output: 78.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:11, 13.64it/s, est. speed input: 755.00 toks/s, output: 261.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:05, 29.05it/s, est. speed input: 1415.36 toks/s, output: 655.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 28.73it/s, est. speed input: 1488.41 toks/s, output: 815.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 41.26it/s, est. speed input: 1943.29 toks/s, output: 1338.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 43.20it/s, est. speed input: 2085.25 toks/s, output: 1638.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 44.34it/s, est. speed input: 2178.42 toks/s, output: 1894.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 47.30it/s, est. speed input: 2296.57 toks/s, output: 2243.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 41.13it/s, est. speed input: 2239.90 toks/s, output: 2409.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 37.67it/s, est. speed input: 2201.57 toks/s, output: 2602.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 37.11it/s, est. speed input: 2195.54 toks/s, output: 2802.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 36.15it/s, est. speed input: 2185.19 toks/s, output: 2989.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:02<00:05, 18.18it/s, est. speed input: 1769.85 toks/s, output: 2614.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:03<00:13,  7.17it/s, est. speed input: 1150.78 toks/s, output: 1918.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:04<00:18,  5.16it/s, est. speed input: 912.55 toks/s, output: 1742.24 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:05<00:26,  3.38it/s, est. speed input: 707.23 toks/s, output: 1517.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:07<00:36,  2.41it/s, est. speed input: 564.58 toks/s, output: 1391.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:08<00:36,  2.44it/s, est. speed input: 545.64 toks/s, output: 1438.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:09<00:43,  1.99it/s, est. speed input: 492.04 toks/s, output: 1391.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:09<00:42,  2.05it/s, est. speed input: 476.37 toks/s, output: 1439.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:09<00:41,  2.04it/s, est. speed input: 458.69 toks/s, output: 1479.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10968.5 tokens/s, Running: 85 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:12<01:12,  1.16it/s, est. speed input: 383.49 toks/s, output: 1331.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:12<01:01,  1.36it/s, est. speed input: 377.78 toks/s, output: 1404.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:12<00:54,  1.51it/s, est. speed input: 369.15 toks/s, output: 1465.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:13<01:00,  1.33it/s, est. speed input: 347.00 toks/s, output: 1471.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:15<01:20,  1.00s/it, est. speed input: 313.68 toks/s, output: 1423.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:15<01:00,  1.31it/s, est. speed input: 314.20 toks/s, output: 1518.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:16<00:52,  1.50it/s, est. speed input: 309.44 toks/s, output: 1588.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:17<01:05,  1.18it/s, est. speed input: 290.01 toks/s, output: 1580.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:17<00:49,  1.54it/s, est. speed input: 290.56 toks/s, output: 1673.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:19<01:07,  1.11it/s, est. speed input: 271.03 toks/s, output: 1651.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8550.8 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:20<01:10,  1.05it/s, est. speed input: 259.60 toks/s, output: 1672.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:20<00:57,  1.26it/s, est. speed input: 257.22 toks/s, output: 1747.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:21<01:07,  1.06it/s, est. speed input: 244.74 toks/s, output: 1752.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:22<00:53,  1.31it/s, est. speed input: 238.44 toks/s, output: 1879.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:23<00:41,  1.66it/s, est. speed input: 239.67 toks/s, output: 1977.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:23<00:40,  1.69it/s, est. speed input: 236.60 toks/s, output: 2039.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:24<00:38,  1.72it/s, est. speed input: 233.73 toks/s, output: 2101.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:24<00:31,  2.04it/s, est. speed input: 231.49 toks/s, output: 2251.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:27<00:58,  1.10it/s, est. speed input: 214.55 toks/s, output: 2172.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:27<00:36,  1.68it/s, est. speed input: 216.85 toks/s, output: 2366.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:27<00:35,  1.72it/s, est. speed input: 214.92 toks/s, output: 2429.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:28<00:26,  2.23it/s, est. speed input: 215.50 toks/s, output: 2603.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:28<00:21,  2.68it/s, est. speed input: 216.79 toks/s, output: 2701.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:28<00:20,  2.73it/s, est. speed input: 216.32 toks/s, output: 2777.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:29<00:21,  2.61it/s, est. speed input: 215.16 toks/s, output: 2845.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7083.5 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:30<00:18,  2.92it/s, est. speed input: 214.78 toks/s, output: 3079.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:18,  2.88it/s, est. speed input: 214.10 toks/s, output: 3151.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:20,  2.52it/s, est. speed input: 212.07 toks/s, output: 3202.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:31<00:22,  2.26it/s, est. speed input: 210.12 toks/s, output: 3252.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:33<00:35,  1.38it/s, est. speed input: 202.17 toks/s, output: 3211.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:34<00:34,  1.39it/s, est. speed input: 199.74 toks/s, output: 3256.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:34<00:29,  1.59it/s, est. speed input: 199.16 toks/s, output: 3329.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:35<00:29,  1.53it/s, est. speed input: 196.87 toks/s, output: 3372.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:35<00:28,  1.59it/s, est. speed input: 195.46 toks/s, output: 3428.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:36<00:26,  1.65it/s, est. speed input: 194.11 toks/s, output: 3487.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:23,  1.80it/s, est. speed input: 193.40 toks/s, output: 3555.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:36<00:18,  2.25it/s, est. speed input: 194.10 toks/s, output: 3649.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:38<00:33,  1.21it/s, est. speed input: 186.92 toks/s, output: 3597.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:29,  1.36it/s, est. speed input: 185.87 toks/s, output: 3660.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5602.4 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:32,  1.22it/s, est. speed input: 182.64 toks/s, output: 3678.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:40<00:17,  2.11it/s, est. speed input: 185.02 toks/s, output: 3889.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:41<00:21,  1.69it/s, est. speed input: 182.18 toks/s, output: 3912.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:37,  1.07s/it, est. speed input: 173.43 toks/s, output: 3808.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:43<00:27,  1.22it/s, est. speed input: 174.24 toks/s, output: 3908.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:44<00:21,  1.52it/s, est. speed input: 174.64 toks/s, output: 3999.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:44<00:20,  1.58it/s, est. speed input: 173.74 toks/s, output: 4062.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:27,  1.15it/s, est. speed input: 169.59 toks/s, output: 4048.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:26,  1.13it/s, est. speed input: 167.63 toks/s, output: 4085.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:15,  1.83it/s, est. speed input: 169.12 toks/s, output: 4288.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:15,  1.73it/s, est. speed input: 168.04 toks/s, output: 4343.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:48<00:15,  1.64it/s, est. speed input: 166.87 toks/s, output: 4397.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:48<00:12,  2.06it/s, est. speed input: 167.57 toks/s, output: 4499.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:49<00:11,  2.07it/s, est. speed input: 167.20 toks/s, output: 4572.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:49<00:09,  2.46it/s, est. speed input: 167.63 toks/s, output: 4668.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:09:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4300.3 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:12,  1.76it/s, est. speed input: 165.67 toks/s, output: 4696.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:52<00:23,  1.13s/it, est. speed input: 158.96 toks/s, output: 4592.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:53<00:17,  1.17it/s, est. speed input: 159.56 toks/s, output: 4694.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:54<00:16,  1.15it/s, est. speed input: 157.98 toks/s, output: 4734.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:54<00:12,  1.42it/s, est. speed input: 158.22 toks/s, output: 4826.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:55<00:11,  1.44it/s, est. speed input: 157.35 toks/s, output: 4886.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:57<00:17,  1.10s/it, est. speed input: 152.70 toks/s, output: 4831.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:57<00:12,  1.23it/s, est. speed input: 153.41 toks/s, output: 4942.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:58<00:14,  1.01s/it, est. speed input: 150.60 toks/s, output: 4940.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:10:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2853.6 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:01<00:21,  1.65s/it, est. speed input: 143.88 toks/s, output: 4813.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:15,  1.30s/it, est. speed input: 143.75 toks/s, output: 4901.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:04<00:15,  1.42s/it, est. speed input: 140.89 toks/s, output: 4898.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:04<00:10,  1.08s/it, est. speed input: 141.23 toks/s, output: 5004.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.08s/it, est. speed input: 150.71 toks/s, output: 6275.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.49it/s, est. speed input: 150.71 toks/s, output: 6275.06 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:10:12 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:10:12 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:10:13 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:10:13 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:10:13 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:10:13 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:10:13 [abstract.py:306] It took 0.534601 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:10:13 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:10:13 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:10:13 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:10:13 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) 
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0108
+  • Training Accuracy: 0.3563
+  • Mean Generation Length: 2523.1125
+  • Training FLOPS: 45.30 TFLOPS (11.32 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 309.35s
+  • policy_training: 201.49s (65.1%)
+  • generation: 68.47s (22.1%)
+  • teacher_logprob_inference: 34.25s (11.1%)
+  • policy_training/sharding_data: 4.29s (1.4%)
+  • prepare_for_generation: 3.71s (1.2%)
+  • policy_training/submit_training_futures: 1.84s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.67s (0.5%)
+  • training_prep: 0.83s (0.3%)
+  • teacher_logprob_inference_prep: 0.42s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 31/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 151: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:11 [abstract.py:324] It took 0.054012 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08847676e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68306427e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80030365e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08847676e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68306427e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80030365e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08847676e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68306427e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80030365e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08847676e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68306427e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80030365e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08848672e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68152008e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80138474e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08848672e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68152008e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80138474e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08848672e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68152008e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80138474e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08848672e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68152008e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80138474e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07473450e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.27045441e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #151] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44991028e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63521652e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.30437422e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #151] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39426918e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41869965e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.74584866e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #151] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.14289322e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15928879e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57790966e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #151] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34619446e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 151: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:14 [abstract.py:324] It took 0.029288 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13345.44it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:15 [loggers.py:236] Engine 000: Avg prompt throughput: 38.9 tokens/s, Avg generation throughput: 39.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.09it/s, est. speed input: 125.38 toks/s, output: 29.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:16,  9.54it/s, est. speed input: 473.00 toks/s, output: 155.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:14, 10.77it/s, est. speed input: 540.55 toks/s, output: 227.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:12, 12.17it/s, est. speed input: 600.75 toks/s, output: 309.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:04, 32.52it/s, est. speed input: 1154.07 toks/s, output: 874.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:02, 48.17it/s, est. speed input: 1593.06 toks/s, output: 1416.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 44.66it/s, est. speed input: 1704.94 toks/s, output: 1687.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 55.06it/s, est. speed input: 1977.43 toks/s, output: 2213.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 51.49it/s, est. speed input: 2052.56 toks/s, output: 2509.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 51.12it/s, est. speed input: 2124.95 toks/s, output: 2800.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:01, 64.29it/s, est. speed input: 2373.36 toks/s, output: 3505.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:01<00:01, 58.05it/s, est. speed input: 2404.64 toks/s, output: 3817.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:02<00:02, 32.14it/s, est. speed input: 2113.14 toks/s, output: 3668.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:05<00:10,  6.68it/s, est. speed input: 1030.65 toks/s, output: 2151.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:09<00:23,  2.93it/s, est. speed input: 584.82 toks/s, output: 1558.24 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9875.4 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:10<00:21,  3.03it/s, est. speed input: 555.21 toks/s, output: 1785.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:10<00:20,  3.11it/s, est. speed input: 538.42 toks/s, output: 1932.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:12<00:22,  2.77it/s, est. speed input: 499.37 toks/s, output: 1994.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:12<00:19,  3.05it/s, est. speed input: 493.47 toks/s, output: 2175.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:12<00:18,  3.09it/s, est. speed input: 487.17 toks/s, output: 2246.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:13<00:20,  2.81it/s, est. speed input: 472.00 toks/s, output: 2278.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:13<00:15,  3.56it/s, est. speed input: 473.55 toks/s, output: 2484.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:13<00:15,  3.54it/s, est. speed input: 468.21 toks/s, output: 2555.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:14<00:15,  3.40it/s, est. speed input: 461.46 toks/s, output: 2616.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:14<00:13,  3.91it/s, est. speed input: 461.63 toks/s, output: 2716.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:14<00:12,  4.11it/s, est. speed input: 459.28 toks/s, output: 2801.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:14<00:13,  3.68it/s, est. speed input: 452.23 toks/s, output: 2857.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:15<00:16,  2.98it/s, est. speed input: 440.97 toks/s, output: 2886.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:15<00:17,  2.70it/s, est. speed input: 432.07 toks/s, output: 2926.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:16<00:17,  2.70it/s, est. speed input: 425.89 toks/s, output: 2983.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:17<00:35,  1.30it/s, est. speed input: 387.48 toks/s, output: 2816.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:18<00:27,  1.65it/s, est. speed input: 386.29 toks/s, output: 2909.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:18<00:20,  2.16it/s, est. speed input: 387.08 toks/s, output: 3016.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:18<00:18,  2.27it/s, est. speed input: 382.31 toks/s, output: 3080.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:19<00:19,  2.18it/s, est. speed input: 375.47 toks/s, output: 3126.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:19<00:19,  2.13it/s, est. speed input: 369.14 toks/s, output: 3174.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:20<00:19,  2.04it/s, est. speed input: 362.26 toks/s, output: 3217.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6984.2 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:20<00:21,  1.78it/s, est. speed input: 352.37 toks/s, output: 3231.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:21<00:25,  1.50it/s, est. speed input: 340.43 toks/s, output: 3224.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:22<00:21,  1.69it/s, est. speed input: 336.72 toks/s, output: 3292.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:23<00:23,  1.53it/s, est. speed input: 327.53 toks/s, output: 3306.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:25<00:38,  1.09s/it, est. speed input: 302.61 toks/s, output: 3161.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:25<00:22,  1.48it/s, est. speed input: 302.59 toks/s, output: 3370.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:26<00:23,  1.38it/s, est. speed input: 295.00 toks/s, output: 3391.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:26<00:17,  1.72it/s, est. speed input: 295.28 toks/s, output: 3499.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:26<00:13,  2.16it/s, est. speed input: 295.92 toks/s, output: 3611.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:30<00:38,  1.31s/it, est. speed input: 263.75 toks/s, output: 3326.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4893.5 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:29,  1.07s/it, est. speed input: 261.88 toks/s, output: 3411.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:31<00:24,  1.12it/s, est. speed input: 259.95 toks/s, output: 3494.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:31<00:22,  1.18it/s, est. speed input: 255.80 toks/s, output: 3547.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:32<00:15,  1.57it/s, est. speed input: 256.65 toks/s, output: 3666.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:32<00:12,  1.89it/s, est. speed input: 256.30 toks/s, output: 3770.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:33<00:15,  1.44it/s, est. speed input: 249.77 toks/s, output: 3782.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:13,  1.68it/s, est. speed input: 248.83 toks/s, output: 3876.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:33<00:09,  2.10it/s, est. speed input: 249.26 toks/s, output: 3990.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:34<00:12,  1.56it/s, est. speed input: 243.66 toks/s, output: 4009.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:35<00:09,  1.94it/s, est. speed input: 243.88 toks/s, output: 4121.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:36<00:11,  1.62it/s, est. speed input: 239.71 toks/s, output: 4161.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:08,  1.94it/s, est. speed input: 239.60 toks/s, output: 4267.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:36<00:06,  2.43it/s, est. speed input: 240.13 toks/s, output: 4386.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:37<00:08,  1.86it/s, est. speed input: 236.30 toks/s, output: 4427.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:37<00:07,  1.75it/s, est. speed input: 233.87 toks/s, output: 4492.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:38<00:06,  2.16it/s, est. speed input: 234.12 toks/s, output: 4607.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:39<00:08,  1.39it/s, est. speed input: 227.85 toks/s, output: 4595.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:39<00:05,  1.85it/s, est. speed input: 228.69 toks/s, output: 4723.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:14:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3305.6 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:42<00:11,  1.17s/it, est. speed input: 215.83 toks/s, output: 4573.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:42<00:08,  1.06it/s, est. speed input: 215.23 toks/s, output: 4676.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:43<00:05,  1.38it/s, est. speed input: 213.35 toks/s, output: 4867.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:43<00:03,  1.69it/s, est. speed input: 213.83 toks/s, output: 4994.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:44<00:02,  1.74it/s, est. speed input: 212.68 toks/s, output: 5085.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:45<00:02,  1.46it/s, est. speed input: 209.41 toks/s, output: 5127.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:46<00:02,  1.20it/s, est. speed input: 205.20 toks/s, output: 5146.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:46<00:01,  1.54it/s, est. speed input: 205.64 toks/s, output: 5279.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:48<00:01,  1.05s/it, est. speed input: 198.30 toks/s, output: 5217.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:15:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1180.3 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.22s/it, est. speed input: 193.13 toks/s, output: 5213.66 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.22s/it, est. speed input: 193.13 toks/s, output: 5213.66 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.18it/s, est. speed input: 193.13 toks/s, output: 5213.66 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:15:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:15:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:15:07 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:15:08 [abstract.py:306] It took 0.537785 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:15:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:15:07 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:15:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0125
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 1641.3688
+  • Training FLOPS: 38.14 TFLOPS (9.53 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 239.19s
+  • policy_training: 154.04s (64.4%)
+  • generation: 53.34s (22.3%)
+  • teacher_logprob_inference: 26.65s (11.1%)
+  • policy_training/sharding_data: 4.45s (1.9%)
+  • prepare_for_generation: 3.67s (1.5%)
+  • policy_training/submit_training_futures: 1.87s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.7%)
+  • training_prep: 0.91s (0.4%)
+  • teacher_logprob_inference_prep: 0.40s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 152: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:10 [abstract.py:324] It took 0.048469 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08856328e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68295517e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79499969e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08856328e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68295517e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79499969e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08856328e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68295517e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79499969e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08856328e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68295517e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79499969e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08854102e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68178062e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79533386e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07596588e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.40764427e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #152] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45279282e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08854102e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68178062e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79533386e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65659180e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31604233e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #152] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.41787720e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08854102e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68178062e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79533386e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41747742e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.48672438e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #152] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.18146858e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08854102e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68178062e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79533386e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16947441e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57543240e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #152] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34916077e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 152: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:13 [abstract.py:324] It took 0.029287 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13905.40it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:13 [loggers.py:236] Engine 000: Avg prompt throughput: 51.2 tokens/s, Avg generation throughput: 0.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:23,  6.70it/s, est. speed input: 408.84 toks/s, output: 46.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:05, 30.41it/s, est. speed input: 1626.06 toks/s, output: 358.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:07, 18.81it/s, est. speed input: 1189.03 toks/s, output: 408.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:07, 20.38it/s, est. speed input: 1241.59 toks/s, output: 575.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 23.86it/s, est. speed input: 1363.05 toks/s, output: 819.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:05, 26.48it/s, est. speed input: 1444.25 toks/s, output: 1049.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 41.50it/s, est. speed input: 1810.98 toks/s, output: 1723.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 45.84it/s, est. speed input: 1964.61 toks/s, output: 2100.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 47.43it/s, est. speed input: 2066.61 toks/s, output: 2437.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 51.38it/s, est. speed input: 2200.02 toks/s, output: 2863.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 53.15it/s, est. speed input: 2291.14 toks/s, output: 3216.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 43.72it/s, est. speed input: 2245.13 toks/s, output: 3398.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:02, 39.14it/s, est. speed input: 2209.62 toks/s, output: 3574.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:03<00:09,  9.15it/s, est. speed input: 1248.94 toks/s, output: 2271.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:04<00:11,  7.09it/s, est. speed input: 1032.16 toks/s, output: 2187.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:05<00:14,  5.43it/s, est. speed input: 866.78 toks/s, output: 2083.22 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:05<00:13,  5.70it/s, est. speed input: 848.77 toks/s, output: 2215.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:06<00:15,  4.87it/s, est. speed input: 779.42 toks/s, output: 2211.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:08<00:25,  2.95it/s, est. speed input: 633.93 toks/s, output: 1981.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:08<00:24,  3.07it/s, est. speed input: 623.04 toks/s, output: 2041.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:09<00:27,  2.62it/s, est. speed input: 583.12 toks/s, output: 2005.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10377.3 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:43,  1.66it/s, est. speed input: 501.49 toks/s, output: 1821.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:11<00:44,  1.60it/s, est. speed input: 475.84 toks/s, output: 1826.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:11<00:39,  1.77it/s, est. speed input: 465.99 toks/s, output: 1884.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:12<00:48,  1.43it/s, est. speed input: 430.26 toks/s, output: 1838.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:13<00:50,  1.35it/s, est. speed input: 407.66 toks/s, output: 1839.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:15<01:17,  1.16s/it, est. speed input: 353.09 toks/s, output: 1692.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:18<01:12,  1.12s/it, est. speed input: 318.09 toks/s, output: 1712.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:19<01:19,  1.25s/it, est. speed input: 294.12 toks/s, output: 1683.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:19<01:00,  1.04it/s, est. speed input: 295.39 toks/s, output: 1789.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:20<00:35,  1.70it/s, est. speed input: 299.17 toks/s, output: 2008.35 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7870.8 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:21<00:50,  1.19it/s, est. speed input: 278.84 toks/s, output: 1970.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<00:41,  1.41it/s, est. speed input: 277.75 toks/s, output: 2059.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<00:44,  1.28it/s, est. speed input: 262.08 toks/s, output: 2129.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:25<00:54,  1.03it/s, est. speed input: 247.83 toks/s, output: 2111.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:25<00:34,  1.59it/s, est. speed input: 250.41 toks/s, output: 2326.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:29,  1.79it/s, est. speed input: 249.70 toks/s, output: 2416.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:26<00:27,  1.91it/s, est. speed input: 248.12 toks/s, output: 2495.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:25,  2.00it/s, est. speed input: 246.36 toks/s, output: 2571.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:27<00:21,  2.28it/s, est. speed input: 246.21 toks/s, output: 2663.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:27<00:22,  2.15it/s, est. speed input: 243.62 toks/s, output: 2728.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:27<00:12,  3.59it/s, est. speed input: 246.84 toks/s, output: 3041.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:28<00:16,  2.78it/s, est. speed input: 243.21 toks/s, output: 3087.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:28<00:14,  3.12it/s, est. speed input: 243.75 toks/s, output: 3184.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:29<00:17,  2.50it/s, est. speed input: 240.45 toks/s, output: 3233.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:29<00:17,  2.37it/s, est. speed input: 238.64 toks/s, output: 3298.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6530.2 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:30<00:20,  1.95it/s, est. speed input: 234.81 toks/s, output: 3335.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:25,  1.60it/s, est. speed input: 229.86 toks/s, output: 3356.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:32<00:31,  1.24it/s, est. speed input: 222.80 toks/s, output: 3346.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:33<00:32,  1.18it/s, est. speed input: 218.34 toks/s, output: 3371.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:33<00:24,  1.53it/s, est. speed input: 218.93 toks/s, output: 3472.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:34<00:20,  1.79it/s, est. speed input: 218.49 toks/s, output: 3557.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:34<00:16,  2.13it/s, est. speed input: 218.56 toks/s, output: 3650.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:35<00:15,  2.14it/s, est. speed input: 217.46 toks/s, output: 3722.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:36<00:24,  1.36it/s, est. speed input: 210.87 toks/s, output: 3701.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:36<00:17,  1.78it/s, est. speed input: 211.65 toks/s, output: 3806.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:36<00:13,  2.29it/s, est. speed input: 212.50 toks/s, output: 3911.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:37<00:15,  1.96it/s, est. speed input: 210.28 toks/s, output: 3961.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:38<00:19,  1.49it/s, est. speed input: 206.16 toks/s, output: 3974.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:38<00:17,  1.64it/s, est. speed input: 205.21 toks/s, output: 4048.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:39<00:15,  1.76it/s, est. speed input: 204.28 toks/s, output: 4122.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:39<00:11,  2.23it/s, est. speed input: 204.90 toks/s, output: 4226.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:40<00:11,  2.13it/s, est. speed input: 203.79 toks/s, output: 4294.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:18:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4672.3 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:42<00:23,  1.04it/s, est. speed input: 194.94 toks/s, output: 4201.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:43<00:22,  1.04it/s, est. speed input: 192.05 toks/s, output: 4233.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:43<00:15,  1.42it/s, est. speed input: 192.96 toks/s, output: 4347.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:43<00:11,  1.82it/s, est. speed input: 193.50 toks/s, output: 4453.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:44<00:14,  1.41it/s, est. speed input: 190.12 toks/s, output: 4470.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:44<00:10,  1.78it/s, est. speed input: 190.52 toks/s, output: 4574.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:44<00:07,  2.34it/s, est. speed input: 191.35 toks/s, output: 4688.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:45<00:04,  3.50it/s, est. speed input: 193.09 toks/s, output: 4915.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:46<00:10,  1.47it/s, est. speed input: 186.59 toks/s, output: 4845.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:47<00:06,  1.87it/s, est. speed input: 186.51 toks/s, output: 5031.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:47<00:05,  2.28it/s, est. speed input: 187.31 toks/s, output: 5147.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:48<00:05,  1.96it/s, est. speed input: 185.75 toks/s, output: 5199.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:48<00:04,  2.07it/s, est. speed input: 185.47 toks/s, output: 5286.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:49<00:05,  1.77it/s, est. speed input: 183.80 toks/s, output: 5333.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:19:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2932.7 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:51<00:07,  1.11it/s, est. speed input: 178.66 toks/s, output: 5283.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:51<00:05,  1.25it/s, est. speed input: 177.93 toks/s, output: 5361.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:54<00:07,  1.20s/it, est. speed input: 171.90 toks/s, output: 5283.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:54<00:05,  1.06s/it, est. speed input: 170.75 toks/s, output: 5352.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:55<00:03,  1.09it/s, est. speed input: 170.05 toks/s, output: 5437.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:56<00:03,  1.06s/it, est. speed input: 166.97 toks/s, output: 5446.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:57<00:01,  1.17it/s, est. speed input: 166.91 toks/s, output: 5554.42 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  1.17it/s, est. speed input: 169.04 toks/s, output: 5840.75 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  2.80it/s, est. speed input: 169.04 toks/s, output: 5840.75 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:19:13 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:19:13 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:19:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:19:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:19:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:19:14 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:19:14 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:19:14 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:19:14 [abstract.py:306] It took 0.536962 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:19:14 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:19:14 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.4125
+  • Mean Generation Length: 2088.6687
+  • Training FLOPS: 42.56 TFLOPS (10.64 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 272.89s
+  • policy_training: 176.23s (64.6%)
+  • generation: 60.78s (22.3%)
+  • teacher_logprob_inference: 30.87s (11.3%)
+  • policy_training/sharding_data: 4.43s (1.6%)
+  • prepare_for_generation: 3.54s (1.3%)
+  • policy_training/submit_training_futures: 1.97s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.88s (0.3%)
+  • teacher_logprob_inference_prep: 0.43s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 153: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:22:44 [abstract.py:324] It took 0.050510 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08877578e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68488464e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78786011e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08877578e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68488464e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78786011e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08877578e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68488464e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78786011e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08877578e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68488464e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78786011e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08876211e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68419495e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80325394e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.08509598e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.51932621e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #153] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45316040e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08876211e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68419495e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80325394e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65661621e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.31826200e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #153] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44609680e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08876211e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68419495e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80325394e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.42226105e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.79794025e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #153] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.21706314e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08876211e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68419495e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80325394e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17452927e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57051392e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #153] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34973480e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 153: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:22:47 [abstract.py:324] It took 0.029087 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13960.36it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 11:22:47 [loggers.py:236] Engine 000: Avg prompt throughput: 43.5 tokens/s, Avg generation throughput: 45.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.62it/s, est. speed input: 224.25 toks/s, output: 47.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 22.44it/s, est. speed input: 1129.70 toks/s, output: 296.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:02, 47.39it/s, est. speed input: 2188.31 toks/s, output: 857.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:02, 46.18it/s, est. speed input: 2318.27 toks/s, output: 1140.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 53.23it/s, est. speed input: 2615.40 toks/s, output: 1557.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:00<00:02, 46.96it/s, est. speed input: 2551.70 toks/s, output: 1763.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 49.23it/s, est. speed input: 2642.26 toks/s, output: 2071.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 46.12it/s, est. speed input: 2615.48 toks/s, output: 2330.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 36.99it/s, est. speed input: 2440.26 toks/s, output: 2420.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 41.24it/s, est. speed input: 2507.07 toks/s, output: 2799.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:02<00:07, 12.89it/s, est. speed input: 1561.08 toks/s, output: 1999.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:05<00:21,  4.19it/s, est. speed input: 765.84 toks/s, output: 1297.96 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 11:22:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10796.8 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:13<01:02,  1.37it/s, est. speed input: 330.78 toks/s, output: 845.73 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:14<00:58,  1.43it/s, est. speed input: 314.37 toks/s, output: 1001.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:16<00:56,  1.45it/s, est. speed input: 296.67 toks/s, output: 1137.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:17<00:58,  1.39it/s, est. speed input: 284.01 toks/s, output: 1183.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:17<00:50,  1.57it/s, est. speed input: 285.09 toks/s, output: 1281.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:18<01:01,  1.29it/s, est. speed input: 266.07 toks/s, output: 1289.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:19<00:44,  1.72it/s, est. speed input: 262.58 toks/s, output: 1541.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:19<00:39,  1.92it/s, est. speed input: 262.55 toks/s, output: 1631.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8605.2 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<00:54,  1.36it/s, est. speed input: 245.45 toks/s, output: 1614.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:22<00:54,  1.33it/s, est. speed input: 239.20 toks/s, output: 1662.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:22<00:50,  1.42it/s, est. speed input: 235.99 toks/s, output: 1728.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:23<00:43,  1.64it/s, est. speed input: 235.21 toks/s, output: 1809.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:24<00:41,  1.68it/s, est. speed input: 229.15 toks/s, output: 1932.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:25<00:56,  1.20it/s, est. speed input: 216.98 toks/s, output: 1916.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:26<00:58,  1.14it/s, est. speed input: 211.14 toks/s, output: 1950.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:29<01:32,  1.39s/it, est. speed input: 192.89 toks/s, output: 1868.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7078.7 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:30<01:16,  1.18s/it, est. speed input: 191.05 toks/s, output: 1936.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<01:39,  1.55s/it, est. speed input: 178.32 toks/s, output: 1894.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:33<01:20,  1.28s/it, est. speed input: 176.82 toks/s, output: 1964.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:34<01:06,  1.08s/it, est. speed input: 175.58 toks/s, output: 2035.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:34<01:00,  1.02it/s, est. speed input: 173.54 toks/s, output: 2096.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:35<01:01,  1.02s/it, est. speed input: 169.87 toks/s, output: 2137.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:36<00:44,  1.29it/s, est. speed input: 168.66 toks/s, output: 2289.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:37<00:27,  2.02it/s, est. speed input: 171.36 toks/s, output: 2490.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:37<00:27,  2.03it/s, est. speed input: 170.76 toks/s, output: 2563.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:38<00:30,  1.75it/s, est. speed input: 168.70 toks/s, output: 2614.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:39<00:37,  1.40it/s, est. speed input: 165.42 toks/s, output: 2645.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:39<00:29,  1.75it/s, est. speed input: 166.12 toks/s, output: 2738.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:39<00:26,  1.94it/s, est. speed input: 166.16 toks/s, output: 2819.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6426.9 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:40<00:19,  2.53it/s, est. speed input: 167.14 toks/s, output: 2995.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:41<00:16,  2.79it/s, est. speed input: 167.68 toks/s, output: 3162.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:41<00:20,  2.21it/s, est. speed input: 165.93 toks/s, output: 3209.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:42<00:19,  2.34it/s, est. speed input: 166.01 toks/s, output: 3289.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:42<00:15,  2.77it/s, est. speed input: 166.80 toks/s, output: 3382.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:42<00:16,  2.64it/s, est. speed input: 166.50 toks/s, output: 3455.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:44<00:17,  2.29it/s, est. speed input: 165.11 toks/s, output: 3653.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:45<00:23,  1.69it/s, est. speed input: 162.08 toks/s, output: 3664.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:45<00:20,  1.90it/s, est. speed input: 162.30 toks/s, output: 3746.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:45<00:15,  2.35it/s, est. speed input: 163.21 toks/s, output: 3845.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:46<00:16,  2.18it/s, est. speed input: 162.54 toks/s, output: 3906.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:48<00:24,  1.38it/s, est. speed input: 157.69 toks/s, output: 3941.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:50<00:30,  1.10it/s, est. speed input: 154.10 toks/s, output: 3930.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4720.4 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:50<00:24,  1.32it/s, est. speed input: 154.36 toks/s, output: 4015.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:51<00:22,  1.36it/s, est. speed input: 153.50 toks/s, output: 4072.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:51<00:17,  1.76it/s, est. speed input: 154.32 toks/s, output: 4171.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:51<00:14,  2.02it/s, est. speed input: 154.56 toks/s, output: 4255.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:52<00:15,  1.79it/s, est. speed input: 153.60 toks/s, output: 4307.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:53<00:16,  1.61it/s, est. speed input: 152.52 toks/s, output: 4354.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:53<00:13,  1.88it/s, est. speed input: 152.71 toks/s, output: 4438.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:53<00:10,  2.41it/s, est. speed input: 153.47 toks/s, output: 4537.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:56<00:31,  1.29s/it, est. speed input: 145.41 toks/s, output: 4380.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:57<00:22,  1.03it/s, est. speed input: 145.91 toks/s, output: 4476.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:57<00:14,  1.42it/s, est. speed input: 145.97 toks/s, output: 4639.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:59<00:17,  1.15it/s, est. speed input: 143.59 toks/s, output: 4645.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:00<00:16,  1.15it/s, est. speed input: 142.54 toks/s, output: 4693.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3646.2 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:01<00:19,  1.09s/it, est. speed input: 139.59 toks/s, output: 4680.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:03<00:21,  1.27s/it, est. speed input: 136.74 toks/s, output: 4670.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:04<00:18,  1.17s/it, est. speed input: 135.77 toks/s, output: 4721.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:04<00:14,  1.04it/s, est. speed input: 135.77 toks/s, output: 4806.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:08<00:24,  1.75s/it, est. speed input: 129.40 toks/s, output: 4669.26 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  1.75s/it, est. speed input: 141.68 toks/s, output: 6342.33 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:08<00:00,  2.33it/s, est. speed input: 141.68 toks/s, output: 6342.33 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:59 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:23:59 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:23:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:23:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:23:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:23:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:24:00 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:24:00 [abstract.py:306] It took 0.531716 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:24:00 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:24:00 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:24:00 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0098
+  • Training Accuracy: 0.2375
+  • Mean Generation Length: 2717.2251
+  • Training FLOPS: 48.61 TFLOPS (12.15 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 316.88s
+  • policy_training: 203.29s (64.2%)
+  • generation: 72.97s (23.0%)
+  • teacher_logprob_inference: 35.46s (11.2%)
+  • policy_training/sharding_data: 4.33s (1.4%)
+  • prepare_for_generation: 3.74s (1.2%)
+  • policy_training/submit_training_futures: 1.61s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.5%)
+  • training_prep: 0.81s (0.3%)
+  • teacher_logprob_inference_prep: 0.43s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 154: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:00 [abstract.py:324] It took 0.055443 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08884453e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68103790e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78042908e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08884453e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68103790e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78042908e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08884453e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68103790e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78042908e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08884453e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68103790e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78042908e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08887227e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69220810e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79300232e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.07689781e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.45481873e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #154] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45195663e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08887227e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69220810e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79300232e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66933365e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34219036e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #154] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44999466e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08887227e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69220810e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79300232e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41464233e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.18815994e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #154] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24369354e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08887227e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69220810e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79300232e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16649704e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57082024e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #154] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35281113e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 154: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:04 [abstract.py:324] It took 0.029801 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12981.94it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:04 [loggers.py:236] Engine 000: Avg prompt throughput: 38.0 tokens/s, Avg generation throughput: 84.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:45,  3.47it/s, est. speed input: 204.66 toks/s, output: 45.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 21.09it/s, est. speed input: 1046.57 toks/s, output: 301.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 28.97it/s, est. speed input: 1421.82 toks/s, output: 513.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:03, 41.20it/s, est. speed input: 1897.30 toks/s, output: 883.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 42.08it/s, est. speed input: 2041.44 toks/s, output: 1129.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:03, 41.02it/s, est. speed input: 2106.21 toks/s, output: 1367.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 41.79it/s, est. speed input: 2180.49 toks/s, output: 1643.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 44.29it/s, est. speed input: 2278.06 toks/s, output: 1983.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 52.68it/s, est. speed input: 2500.83 toks/s, output: 2639.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 55.52it/s, est. speed input: 2604.12 toks/s, output: 3059.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 54.54it/s, est. speed input: 2645.03 toks/s, output: 3393.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 54.40it/s, est. speed input: 2695.36 toks/s, output: 3770.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:02<00:02, 27.60it/s, est. speed input: 2244.06 toks/s, output: 3412.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:04<00:12,  6.06it/s, est. speed input: 1042.85 toks/s, output: 1949.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:06<00:15,  4.74it/s, est. speed input: 856.97 toks/s, output: 1880.01 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:07<00:19,  3.60it/s, est. speed input: 703.11 toks/s, output: 1820.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:09<00:22,  3.02it/s, est. speed input: 622.19 toks/s, output: 1814.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9903.8 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:12<00:37,  1.74it/s, est. speed input: 467.26 toks/s, output: 1567.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:13<00:38,  1.65it/s, est. speed input: 442.20 toks/s, output: 1589.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:36,  1.75it/s, est. speed input: 434.76 toks/s, output: 1667.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:14<00:43,  1.44it/s, est. speed input: 401.21 toks/s, output: 1645.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:15<00:39,  1.53it/s, est. speed input: 392.65 toks/s, output: 1716.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:54,  1.11it/s, est. speed input: 355.18 toks/s, output: 1659.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:18<00:52,  1.13it/s, est. speed input: 342.30 toks/s, output: 1706.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:18<00:39,  1.45it/s, est. speed input: 343.55 toks/s, output: 1818.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:18<00:36,  1.55it/s, est. speed input: 337.03 toks/s, output: 1889.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:18<00:27,  2.00it/s, est. speed input: 338.17 toks/s, output: 2000.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:19<00:32,  1.72it/s, est. speed input: 327.60 toks/s, output: 2041.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:19<00:28,  1.89it/s, est. speed input: 324.05 toks/s, output: 2123.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7651.5 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:20<00:25,  2.11it/s, est. speed input: 321.64 toks/s, output: 2210.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:21<00:27,  1.87it/s, est. speed input: 314.14 toks/s, output: 2260.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:23<00:51,  1.01s/it, est. speed input: 287.62 toks/s, output: 2174.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:23<00:46,  1.07it/s, est. speed input: 281.12 toks/s, output: 2229.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:26,  1.84it/s, est. speed input: 284.16 toks/s, output: 2459.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:24<00:21,  2.24it/s, est. speed input: 284.94 toks/s, output: 2566.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:24<00:18,  2.47it/s, est. speed input: 284.09 toks/s, output: 2659.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:20,  2.15it/s, est. speed input: 279.51 toks/s, output: 2716.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:26<00:28,  1.55it/s, est. speed input: 270.02 toks/s, output: 2725.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:18,  2.21it/s, est. speed input: 270.40 toks/s, output: 2927.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:27<00:13,  2.88it/s, est. speed input: 271.16 toks/s, output: 3133.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:27<00:11,  3.44it/s, est. speed input: 271.71 toks/s, output: 3335.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:27<00:12,  3.01it/s, est. speed input: 269.11 toks/s, output: 3400.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:28<00:12,  2.81it/s, est. speed input: 267.20 toks/s, output: 3472.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:13,  2.67it/s, est. speed input: 265.21 toks/s, output: 3544.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:29<00:20,  1.70it/s, est. speed input: 256.66 toks/s, output: 3526.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5822.6 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:18,  1.80it/s, est. speed input: 254.70 toks/s, output: 3598.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:31<00:22,  1.42it/s, est. speed input: 247.97 toks/s, output: 3599.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:20,  1.53it/s, est. speed input: 245.91 toks/s, output: 3666.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:33<00:14,  1.89it/s, est. speed input: 241.91 toks/s, output: 3896.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:34<00:16,  1.66it/s, est. speed input: 237.43 toks/s, output: 3922.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:14,  1.66it/s, est. speed input: 230.68 toks/s, output: 4097.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:36<00:13,  1.75it/s, est. speed input: 229.64 toks/s, output: 4177.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:36<00:11,  1.95it/s, est. speed input: 229.47 toks/s, output: 4273.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:09,  2.11it/s, est. speed input: 229.07 toks/s, output: 4363.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:37<00:09,  2.01it/s, est. speed input: 227.16 toks/s, output: 4427.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:38<00:10,  1.90it/s, est. speed input: 225.10 toks/s, output: 4487.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:39<00:14,  1.24it/s, est. speed input: 217.83 toks/s, output: 4444.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3888.1 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:11,  1.46it/s, est. speed input: 217.24 toks/s, output: 4535.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:40<00:10,  1.55it/s, est. speed input: 215.90 toks/s, output: 4608.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:07,  1.84it/s, est. speed input: 214.50 toks/s, output: 4780.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:05,  2.22it/s, est. speed input: 215.09 toks/s, output: 4895.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:42<00:07,  1.65it/s, est. speed input: 211.23 toks/s, output: 4912.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:05,  1.93it/s, est. speed input: 211.23 toks/s, output: 5015.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:05,  1.95it/s, est. speed input: 210.22 toks/s, output: 5095.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:46<00:10,  1.15s/it, est. speed input: 198.94 toks/s, output: 4931.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:06,  1.15it/s, est. speed input: 199.57 toks/s, output: 5055.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:04,  1.38it/s, est. speed input: 197.52 toks/s, output: 5219.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:28:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2005.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:50<00:05,  1.17s/it, est. speed input: 188.66 toks/s, output: 5101.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:53<00:04,  1.36s/it, est. speed input: 179.50 toks/s, output: 5083.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  1.36s/it, est. speed input: 182.83 toks/s, output: 5543.60 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  2.99it/s, est. speed input: 182.83 toks/s, output: 5543.60 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:29:00 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:29:00 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:29:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:29:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:29:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:29:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:29:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:29:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:29:00 [abstract.py:306] It took 0.518533 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:29:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:29:00 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0113
+  • Training Accuracy: 0.4938
+  • Mean Generation Length: 1851.0750
+  • Training FLOPS: 42.36 TFLOPS (10.59 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 246.19s
+  • policy_training: 156.87s (63.7%)
+  • generation: 56.66s (23.0%)
+  • teacher_logprob_inference: 27.53s (11.2%)
+  • policy_training/sharding_data: 4.31s (1.8%)
+  • prepare_for_generation: 3.69s (1.5%)
+  • policy_training/submit_training_futures: 1.90s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.7%)
+  • training_prep: 0.86s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 155: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:07 [abstract.py:324] It took 0.075207 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08911406e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67800522e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77828217e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08911406e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67800522e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77828217e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08911406e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67800522e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77828217e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08911406e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67800522e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.77828217e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08911426e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69429626e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78412628e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.06873055e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.44807148e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #155] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45296326e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08911426e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69429626e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78412628e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66725311e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34489441e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #155] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.45733643e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08911426e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69429626e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78412628e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40670013e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.34342051e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #155] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.25028229e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08911426e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69429626e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78412628e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15911064e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58043489e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #155] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35288055e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 155: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:10 [abstract.py:324] It took 0.030789 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:10 [loggers.py:236] Engine 000: Avg prompt throughput: 49.4 tokens/s, Avg generation throughput: 17.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13908.86it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:31,  5.12it/s, est. speed input: 317.57 toks/s, output: 81.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:06, 23.41it/s, est. speed input: 1221.74 toks/s, output: 430.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:07, 20.83it/s, est. speed input: 1181.56 toks/s, output: 524.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 22.38it/s, est. speed input: 1262.63 toks/s, output: 694.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 26.99it/s, est. speed input: 1426.80 toks/s, output: 942.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 23.83it/s, est. speed input: 1378.66 toks/s, output: 1032.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:03, 36.23it/s, est. speed input: 1693.92 toks/s, output: 1608.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 49.67it/s, est. speed input: 2030.12 toks/s, output: 2282.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:01, 60.17it/s, est. speed input: 2326.21 toks/s, output: 2973.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 60.68it/s, est. speed input: 2446.86 toks/s, output: 3386.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 53.25it/s, est. speed input: 2452.44 toks/s, output: 3695.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 53.89it/s, est. speed input: 2515.94 toks/s, output: 4038.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:01<00:01, 52.62it/s, est. speed input: 2579.31 toks/s, output: 4552.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:03<00:05, 15.40it/s, est. speed input: 1664.14 toks/s, output: 3236.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:05<00:12,  5.71it/s, est. speed input: 948.13 toks/s, output: 2098.91 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10025.6 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:11<00:35,  1.96it/s, est. speed input: 456.64 toks/s, output: 1298.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<00:44,  1.51it/s, est. speed input: 368.77 toks/s, output: 1254.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:39,  1.68it/s, est. speed input: 364.21 toks/s, output: 1448.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:17<00:40,  1.60it/s, est. speed input: 339.10 toks/s, output: 1551.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:17<00:35,  1.76it/s, est. speed input: 340.00 toks/s, output: 1658.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<00:38,  1.60it/s, est. speed input: 325.33 toks/s, output: 1689.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:44,  1.38it/s, est. speed input: 308.40 toks/s, output: 1704.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:27,  2.12it/s, est. speed input: 311.74 toks/s, output: 2025.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7848.7 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:30,  1.84it/s, est. speed input: 301.24 toks/s, output: 2057.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:21<00:31,  1.75it/s, est. speed input: 294.36 toks/s, output: 2110.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:21<00:28,  1.92it/s, est. speed input: 292.59 toks/s, output: 2196.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:28,  1.87it/s, est. speed input: 287.78 toks/s, output: 2259.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:30,  1.73it/s, est. speed input: 281.64 toks/s, output: 2309.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:26,  2.00it/s, est. speed input: 280.73 toks/s, output: 2400.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:23<00:25,  2.04it/s, est. speed input: 277.81 toks/s, output: 2473.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:35,  1.41it/s, est. speed input: 266.20 toks/s, output: 2468.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:25<00:29,  1.67it/s, est. speed input: 265.11 toks/s, output: 2557.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:25<00:26,  1.81it/s, est. speed input: 262.97 toks/s, output: 2633.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:26<00:20,  2.29it/s, est. speed input: 263.66 toks/s, output: 2737.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:26<00:18,  2.52it/s, est. speed input: 262.91 toks/s, output: 2826.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:28<00:35,  1.27it/s, est. speed input: 249.02 toks/s, output: 2774.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6189.8 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:30<00:43,  1.01s/it, est. speed input: 232.19 toks/s, output: 2773.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:31<00:36,  1.16it/s, est. speed input: 231.18 toks/s, output: 2858.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:22,  1.74it/s, est. speed input: 232.51 toks/s, output: 3069.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:32<00:24,  1.62it/s, est. speed input: 228.87 toks/s, output: 3117.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:32<00:20,  1.82it/s, est. speed input: 228.37 toks/s, output: 3206.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:33<00:22,  1.67it/s, est. speed input: 225.11 toks/s, output: 3256.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:34<00:26,  1.36it/s, est. speed input: 219.61 toks/s, output: 3274.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:34<00:21,  1.63it/s, est. speed input: 219.47 toks/s, output: 3367.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:35<00:22,  1.52it/s, est. speed input: 216.50 toks/s, output: 3417.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:35<00:16,  2.01it/s, est. speed input: 217.54 toks/s, output: 3529.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:35<00:14,  2.22it/s, est. speed input: 217.14 toks/s, output: 3618.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:36<00:13,  2.31it/s, est. speed input: 216.44 toks/s, output: 3701.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:36<00:12,  2.49it/s, est. speed input: 216.12 toks/s, output: 3791.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:36<00:08,  3.21it/s, est. speed input: 217.01 toks/s, output: 3994.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:37<00:11,  2.33it/s, est. speed input: 213.99 toks/s, output: 4034.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:09,  2.67it/s, est. speed input: 214.33 toks/s, output: 4135.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:39<00:18,  1.33it/s, est. speed input: 206.33 toks/s, output: 4075.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:39<00:14,  1.68it/s, est. speed input: 206.92 toks/s, output: 4181.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:32:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4562.5 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:40<00:08,  2.75it/s, est. speed input: 209.26 toks/s, output: 4413.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:40<00:05,  3.39it/s, est. speed input: 210.31 toks/s, output: 4622.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:40<00:03,  4.58it/s, est. speed input: 212.46 toks/s, output: 4853.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:09,  1.79it/s, est. speed input: 204.41 toks/s, output: 4763.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:11,  1.44it/s, est. speed input: 200.33 toks/s, output: 4764.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:43<00:09,  1.65it/s, est. speed input: 200.23 toks/s, output: 4856.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:44<00:06,  2.07it/s, est. speed input: 201.03 toks/s, output: 4972.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:45<00:09,  1.30it/s, est. speed input: 195.54 toks/s, output: 4935.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:46<00:09,  1.21it/s, est. speed input: 192.72 toks/s, output: 4963.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:48<00:09,  1.10it/s, est. speed input: 187.24 toks/s, output: 5017.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:49<00:08,  1.08it/s, est. speed input: 184.70 toks/s, output: 5053.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:33:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2582.8 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:50<00:07,  1.06it/s, est. speed input: 182.28 toks/s, output: 5091.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:50<00:05,  1.35it/s, est. speed input: 182.76 toks/s, output: 5209.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:50<00:03,  1.76it/s, est. speed input: 183.54 toks/s, output: 5336.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:53<00:05,  1.03s/it, est. speed input: 177.01 toks/s, output: 5255.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:54<00:04,  1.16s/it, est. speed input: 173.27 toks/s, output: 5256.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:55<00:03,  1.13s/it, est. speed input: 171.09 toks/s, output: 5303.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:55<00:01,  1.13it/s, est. speed input: 171.34 toks/s, output: 5422.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  1.13it/s, est. speed input: 173.50 toks/s, output: 5715.62 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:55<00:00,  2.86it/s, est. speed input: 173.50 toks/s, output: 5715.62 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:33:09 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:33:09 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:33:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:33:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:33:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:33:09 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:33:10 [abstract.py:306] It took 0.635093 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:33:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:33:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:33:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:33:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0111
+  • Training Accuracy: 0.4125
+  • Mean Generation Length: 1999.6500
+  • Training FLOPS: 41.77 TFLOPS (10.44 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 266.52s
+  • policy_training: 172.10s (64.6%)
+  • generation: 59.55s (22.3%)
+  • teacher_logprob_inference: 29.52s (11.1%)
+  • policy_training/sharding_data: 4.33s (1.6%)
+  • prepare_for_generation: 3.71s (1.4%)
+  • policy_training/submit_training_futures: 1.66s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.6%)
+  • training_prep: 1.05s (0.4%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 32/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 156: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:36:34 [abstract.py:324] It took 0.048485 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08924453e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67533951e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78080902e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08924453e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67533951e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78080902e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08924453e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67533951e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78080902e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08924453e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67533951e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78080902e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08930039e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69360390e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78439636e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.06665878e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.43039083e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #156] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45529831e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08930039e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69360390e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78439636e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64657745e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34986391e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #156] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.46151657e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08930039e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69360390e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78439636e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40773315e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.21966124e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #156] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.25111313e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08930039e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69360390e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78439636e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14965324e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58728504e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #156] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35276779e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 156: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:36:37 [abstract.py:324] It took 0.028687 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13483.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:36:37 [loggers.py:236] Engine 000: Avg prompt throughput: 44.9 tokens/s, Avg generation throughput: 35.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:58,  2.72it/s, est. speed input: 166.14 toks/s, output: 40.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:05, 26.75it/s, est. speed input: 1311.56 toks/s, output: 414.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 25.33it/s, est. speed input: 1366.29 toks/s, output: 574.69 toks/s]
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 32.92it/s, est. speed input: 1645.62 toks/s, output: 905.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:04, 31.97it/s, est. speed input: 1679.98 toks/s, output: 1088.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 44.62it/s, est. speed input: 2017.07 toks/s, output: 1662.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 45.22it/s, est. speed input: 2100.07 toks/s, output: 1971.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 45.35it/s, est. speed input: 2162.79 toks/s, output: 2265.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 38.20it/s, est. speed input: 2099.67 toks/s, output: 2441.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:02, 44.85it/s, est. speed input: 2225.83 toks/s, output: 2906.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:02, 44.98it/s, est. speed input: 2264.15 toks/s, output: 3231.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:01<00:02, 41.66it/s, est. speed input: 2250.95 toks/s, output: 3449.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:02<00:02, 32.96it/s, est. speed input: 2144.08 toks/s, output: 3547.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:05<00:17,  4.41it/s, est. speed input: 874.05 toks/s, output: 1686.89 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:06<00:18,  4.15it/s, est. speed input: 782.47 toks/s, output: 1785.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:09<00:30,  2.44it/s, est. speed input: 574.43 toks/s, output: 1508.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:36:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10194.8 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:32,  2.23it/s, est. speed input: 517.73 toks/s, output: 1550.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:12<00:41,  1.69it/s, est. speed input: 435.25 toks/s, output: 1506.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:13<00:45,  1.50it/s, est. speed input: 402.53 toks/s, output: 1494.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:14<00:44,  1.52it/s, est. speed input: 389.95 toks/s, output: 1548.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:15<00:44,  1.51it/s, est. speed input: 376.39 toks/s, output: 1595.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:37,  1.75it/s, est. speed input: 374.85 toks/s, output: 1688.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:15<00:35,  1.82it/s, est. speed input: 367.38 toks/s, output: 1754.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:15<00:28,  2.24it/s, est. speed input: 368.01 toks/s, output: 1855.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:16<00:27,  2.31it/s, est. speed input: 362.82 toks/s, output: 1927.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:16<00:21,  2.89it/s, est. speed input: 360.37 toks/s, output: 2105.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:17<00:18,  3.16it/s, est. speed input: 356.25 toks/s, output: 2269.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:18<00:26,  2.22it/s, est. speed input: 341.00 toks/s, output: 2269.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:40,  1.41it/s, est. speed input: 317.35 toks/s, output: 2209.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:34,  1.64it/s, est. speed input: 315.34 toks/s, output: 2293.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:36:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7792.1 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:21<00:29,  1.84it/s, est. speed input: 307.43 toks/s, output: 2426.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:21<00:26,  2.02it/s, est. speed input: 305.68 toks/s, output: 2507.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:21<00:21,  2.36it/s, est. speed input: 305.51 toks/s, output: 2601.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:22<00:24,  2.03it/s, est. speed input: 295.04 toks/s, output: 2697.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:23<00:27,  1.78it/s, est. speed input: 287.62 toks/s, output: 2725.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:25<00:46,  1.04it/s, est. speed input: 265.46 toks/s, output: 2612.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:26<00:38,  1.21it/s, est. speed input: 263.43 toks/s, output: 2689.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:26<00:23,  1.93it/s, est. speed input: 266.01 toks/s, output: 2907.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:26<00:21,  2.00it/s, est. speed input: 263.90 toks/s, output: 2980.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:26<00:18,  2.35it/s, est. speed input: 264.20 toks/s, output: 3077.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:19,  2.14it/s, est. speed input: 260.75 toks/s, output: 3133.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:17,  2.32it/s, est. speed input: 259.86 toks/s, output: 3216.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:28<00:14,  2.83it/s, est. speed input: 260.59 toks/s, output: 3319.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:37:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5917.5 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:25,  1.44it/s, est. speed input: 241.65 toks/s, output: 3345.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:31,  1.13it/s, est. speed input: 231.54 toks/s, output: 3302.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:32<00:25,  1.40it/s, est. speed input: 232.39 toks/s, output: 3409.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:23,  1.46it/s, est. speed input: 230.16 toks/s, output: 3472.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:18,  1.73it/s, est. speed input: 227.86 toks/s, output: 3628.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:20,  1.48it/s, est. speed input: 223.06 toks/s, output: 3648.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:16,  1.85it/s, est. speed input: 223.98 toks/s, output: 3759.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:35<00:13,  2.11it/s, est. speed input: 223.86 toks/s, output: 3852.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:36<00:14,  1.90it/s, est. speed input: 221.46 toks/s, output: 3907.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:37<00:17,  1.58it/s, est. speed input: 217.56 toks/s, output: 3935.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:38<00:20,  1.29it/s, est. speed input: 212.65 toks/s, output: 3945.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:17,  1.45it/s, est. speed input: 211.55 toks/s, output: 4023.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:12,  1.89it/s, est. speed input: 212.36 toks/s, output: 4134.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:37:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4447.2 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:41<00:26,  1.14s/it, est. speed input: 200.60 toks/s, output: 4004.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:41<00:18,  1.19it/s, est. speed input: 201.40 toks/s, output: 4118.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:43<00:24,  1.15s/it, est. speed input: 194.07 toks/s, output: 4068.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:44<00:23,  1.15s/it, est. speed input: 190.33 toks/s, output: 4092.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:45<00:13,  1.35it/s, est. speed input: 190.93 toks/s, output: 4304.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:45<00:11,  1.44it/s, est. speed input: 189.96 toks/s, output: 4382.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:46<00:10,  1.52it/s, est. speed input: 188.99 toks/s, output: 4460.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:46<00:08,  1.75it/s, est. speed input: 188.94 toks/s, output: 4559.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:46<00:06,  2.13it/s, est. speed input: 189.39 toks/s, output: 4670.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:47<00:08,  1.55it/s, est. speed input: 186.43 toks/s, output: 4697.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:48<00:08,  1.43it/s, est. speed input: 184.51 toks/s, output: 4750.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:49<00:08,  1.34it/s, est. speed input: 182.54 toks/s, output: 4803.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:37:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2925.8 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:54<00:18,  1.87s/it, est. speed input: 168.25 toks/s, output: 4538.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:14,  1.64s/it, est. speed input: 166.04 toks/s, output: 4589.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:56<00:12,  1.52s/it, est. speed input: 163.51 toks/s, output: 4631.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:57<00:09,  1.33s/it, est. speed input: 162.03 toks/s, output: 4703.24 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  1.33s/it, est. speed input: 169.45 toks/s, output: 5702.06 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:57<00:00,  2.79it/s, est. speed input: 169.45 toks/s, output: 5702.06 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:37:37 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:37:37 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:37:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:37:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:37:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:37:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:37:37 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:37:38 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:37:38 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 11:37:38 [abstract.py:306] It took 0.532370 seconds to fall asleep.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:37:38 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0137
+  • Training Accuracy: 0.3312
+  • Mean Generation Length: 2045.9938
+  • Training FLOPS: 45.08 TFLOPS (11.27 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 259.83s
+  • policy_training: 164.04s (63.1%)
+  • generation: 60.95s (23.5%)
+  • teacher_logprob_inference: 29.80s (11.5%)
+  • policy_training/sharding_data: 4.34s (1.7%)
+  • prepare_for_generation: 3.59s (1.4%)
+  • policy_training/submit_training_futures: 1.73s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.6%)
+  • training_prep: 0.76s (0.3%)
+  • teacher_logprob_inference_prep: 0.51s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 157: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:40:54 [abstract.py:324] It took 0.054865 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08931836e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67634277e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78053360e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08931836e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67634277e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78053360e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08931836e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67634277e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78053360e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08931836e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67634277e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78053360e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08927520e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68129349e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78352203e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.06295033e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.26589584e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #157] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45781723e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08927520e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68129349e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78352203e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65905457e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34789009e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #157] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.48671875e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08927520e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68129349e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78352203e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.42235107e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.31199598e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #157] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.25112152e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08927520e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68129349e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78352203e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15947495e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58981295e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #157] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35240906e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 157: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:40:57 [abstract.py:324] It took 0.029384 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13852.87it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 11:40:57 [loggers.py:236] Engine 000: Avg prompt throughput: 46.3 tokens/s, Avg generation throughput: 68.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:41,  3.86it/s, est. speed input: 224.05 toks/s, output: 46.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 23.31it/s, est. speed input: 1161.26 toks/s, output: 345.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 34.03it/s, est. speed input: 1686.97 toks/s, output: 644.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 38.36it/s, est. speed input: 1920.83 toks/s, output: 932.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 39.60it/s, est. speed input: 2030.76 toks/s, output: 1165.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:03, 33.17it/s, est. speed input: 1925.16 toks/s, output: 1310.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 35.22it/s, est. speed input: 1991.22 toks/s, output: 1582.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 41.08it/s, est. speed input: 2127.83 toks/s, output: 1955.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 40.03it/s, est. speed input: 2145.39 toks/s, output: 2196.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 44.56it/s, est. speed input: 2246.69 toks/s, output: 2567.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:02, 41.04it/s, est. speed input: 2238.52 toks/s, output: 2867.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 42.30it/s, est. speed input: 2273.71 toks/s, output: 3159.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:03<00:09,  9.34it/s, est. speed input: 1261.75 toks/s, output: 1992.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:08<00:32,  2.68it/s, est. speed input: 544.63 toks/s, output: 1100.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:08<00:27,  3.08it/s, est. speed input: 538.87 toks/s, output: 1383.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:09<00:26,  3.14it/s, est. speed input: 518.49 toks/s, output: 1518.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:41:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10743.1 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:13<00:51,  1.57it/s, est. speed input: 369.91 toks/s, output: 1275.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:13<00:51,  1.52it/s, est. speed input: 353.26 toks/s, output: 1314.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:15<01:00,  1.30it/s, est. speed input: 324.09 toks/s, output: 1302.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:16<01:09,  1.11it/s, est. speed input: 298.43 toks/s, output: 1296.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:17<01:10,  1.08it/s, est. speed input: 284.76 toks/s, output: 1333.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:18<00:56,  1.33it/s, est. speed input: 285.74 toks/s, output: 1433.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:20<01:17,  1.05s/it, est. speed input: 260.31 toks/s, output: 1400.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:41:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8454.1 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:20<01:03,  1.15it/s, est. speed input: 259.07 toks/s, output: 1487.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<00:48,  1.49it/s, est. speed input: 260.43 toks/s, output: 1587.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:20<00:38,  1.86it/s, est. speed input: 260.93 toks/s, output: 1682.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:20<00:31,  2.25it/s, est. speed input: 261.21 toks/s, output: 1775.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:28,  2.39it/s, est. speed input: 257.54 toks/s, output: 1927.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:22<00:40,  1.65it/s, est. speed input: 246.83 toks/s, output: 1936.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:23<00:30,  2.11it/s, est. speed input: 245.94 toks/s, output: 2105.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:23<00:26,  2.42it/s, est. speed input: 246.34 toks/s, output: 2196.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:24<00:26,  2.37it/s, est. speed input: 244.30 toks/s, output: 2265.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:24<00:18,  3.30it/s, est. speed input: 246.61 toks/s, output: 2458.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:25<00:22,  2.65it/s, est. speed input: 242.76 toks/s, output: 2506.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:25<00:24,  2.46it/s, est. speed input: 240.38 toks/s, output: 2567.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:27<00:45,  1.27it/s, est. speed input: 226.29 toks/s, output: 2502.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:28<00:41,  1.38it/s, est. speed input: 223.94 toks/s, output: 2561.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:28<00:32,  1.70it/s, est. speed input: 224.23 toks/s, output: 2651.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:28<00:33,  1.65it/s, est. speed input: 221.24 toks/s, output: 2702.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:20,  2.58it/s, est. speed input: 223.65 toks/s, output: 2899.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:26,  1.95it/s, est. speed input: 218.96 toks/s, output: 2923.68 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:41:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6911.5 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:27,  1.84it/s, est. speed input: 216.38 toks/s, output: 2974.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:31<00:30,  1.65it/s, est. speed input: 212.98 toks/s, output: 3012.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:23,  2.06it/s, est. speed input: 213.72 toks/s, output: 3106.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:41,  1.15it/s, est. speed input: 203.82 toks/s, output: 3048.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:34<00:45,  1.04it/s, est. speed input: 198.47 toks/s, output: 3054.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:34<00:25,  1.75it/s, est. speed input: 200.85 toks/s, output: 3261.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:36<00:32,  1.34it/s, est. speed input: 195.38 toks/s, output: 3258.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:36<00:24,  1.72it/s, est. speed input: 194.92 toks/s, output: 3418.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:37<00:18,  2.14it/s, est. speed input: 195.36 toks/s, output: 3591.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:37<00:12,  2.98it/s, est. speed input: 197.71 toks/s, output: 3799.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:11,  3.35it/s, est. speed input: 198.51 toks/s, output: 3897.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:38<00:11,  3.04it/s, est. speed input: 197.71 toks/s, output: 4043.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:38<00:09,  3.44it/s, est. speed input: 198.52 toks/s, output: 4141.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:38<00:06,  4.77it/s, est. speed input: 201.29 toks/s, output: 4441.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:39<00:07,  4.07it/s, est. speed input: 200.73 toks/s, output: 4508.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:39<00:06,  4.35it/s, est. speed input: 201.40 toks/s, output: 4603.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:41:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5368.1 tokens/s, Running: 29 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:41<00:15,  1.80it/s, est. speed input: 194.74 toks/s, output: 4533.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:42<00:21,  1.24it/s, est. speed input: 188.87 toks/s, output: 4481.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:43<00:21,  1.19it/s, est. speed input: 186.27 toks/s, output: 4504.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:43<00:12,  1.98it/s, est. speed input: 188.49 toks/s, output: 4726.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:46<00:23,  1.02s/it, est. speed input: 178.77 toks/s, output: 4569.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:47<00:20,  1.05it/s, est. speed input: 177.28 toks/s, output: 4618.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:17,  1.23it/s, est. speed input: 176.92 toks/s, output: 4695.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:48<00:16,  1.25it/s, est. speed input: 175.35 toks/s, output: 4742.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:10,  1.74it/s, est. speed input: 175.78 toks/s, output: 4928.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:08,  2.11it/s, est. speed input: 176.49 toks/s, output: 5033.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:41:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3653.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:50<00:12,  1.30it/s, est. speed input: 172.02 toks/s, output: 4994.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:51<00:12,  1.20it/s, est. speed input: 169.81 toks/s, output: 5019.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:14,  1.02s/it, est. speed input: 166.11 toks/s, output: 5002.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:54<00:13,  1.07s/it, est. speed input: 163.63 toks/s, output: 5019.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:56<00:14,  1.19s/it, est. speed input: 160.35 toks/s, output: 5012.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:56<00:10,  1.02it/s, est. speed input: 160.16 toks/s, output: 5099.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:56<00:05,  1.60it/s, est. speed input: 161.20 toks/s, output: 5317.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:05,  1.48it/s, est. speed input: 159.95 toks/s, output: 5369.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:05,  1.29it/s, est. speed input: 158.13 toks/s, output: 5403.97 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:41:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2163.9 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 76.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:00<00:06,  1.13s/it, est. speed input: 153.69 toks/s, output: 5352.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:01<00:04,  1.10it/s, est. speed input: 153.83 toks/s, output: 5456.66 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.10it/s, est. speed input: 158.75 toks/s, output: 6126.16 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.62it/s, est. speed input: 158.75 toks/s, output: 6126.16 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:42:01 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:42:01 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:42:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:42:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:42:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:42:01 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:42:02 [abstract.py:306] It took 0.553598 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:42:02 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:42:02 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:42:02 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:42:02 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0113
+  • Training Accuracy: 0.3500
+  • Mean Generation Length: 2342.4062
+  • Training FLOPS: 44.88 TFLOPS (11.22 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 292.20s
+  • policy_training: 187.82s (64.3%)
+  • generation: 65.12s (22.3%)
+  • teacher_logprob_inference: 34.07s (11.7%)
+  • policy_training/sharding_data: 4.34s (1.5%)
+  • prepare_for_generation: 3.74s (1.3%)
+  • policy_training/submit_training_futures: 1.79s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.67s (0.6%)
+  • training_prep: 0.82s (0.3%)
+  • teacher_logprob_inference_prep: 0.43s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 158: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:45:46 [abstract.py:324] It took 0.050388 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08966680e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66508026e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78869934e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08966680e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66508026e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78869934e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08966680e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66508026e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78869934e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08966680e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66508026e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.78869934e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08965098e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67095261e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77526093e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.05083141e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.10799980e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #158] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45567352e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08965098e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67095261e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77526093e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65262299e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34772358e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #158] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.49927979e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08965098e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67095261e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77526093e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40831299e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.52317619e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #158] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.25140762e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08965098e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67095261e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.77526093e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16546402e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59574623e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #158] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35291077e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 158: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:45:49 [abstract.py:324] It took 0.030226 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13822.63it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 11:45:50 [loggers.py:236] Engine 000: Avg prompt throughput: 41.7 tokens/s, Avg generation throughput: 7.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.66it/s, est. speed input: 216.04 toks/s, output: 40.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:06, 21.62it/s, est. speed input: 1134.44 toks/s, output: 320.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:08, 17.58it/s, est. speed input: 1025.35 toks/s, output: 413.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:05, 27.61it/s, est. speed input: 1383.56 toks/s, output: 885.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 38.23it/s, est. speed input: 1737.73 toks/s, output: 1443.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:02, 47.36it/s, est. speed input: 2009.50 toks/s, output: 1953.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 53.97it/s, est. speed input: 2227.76 toks/s, output: 2446.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:01, 56.10it/s, est. speed input: 2355.91 toks/s, output: 2841.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:01, 50.61it/s, est. speed input: 2367.81 toks/s, output: 3126.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 52.54it/s, est. speed input: 2464.33 toks/s, output: 3580.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 51.70it/s, est. speed input: 2498.06 toks/s, output: 3895.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:03<00:06, 12.58it/s, est. speed input: 1494.89 toks/s, output: 2615.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:08<00:27,  2.82it/s, est. speed input: 569.96 toks/s, output: 1280.97 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10304.2 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:31,  2.32it/s, est. speed input: 466.84 toks/s, output: 1326.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:32,  2.23it/s, est. speed input: 435.11 toks/s, output: 1439.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:14<00:38,  1.83it/s, est. speed input: 381.89 toks/s, output: 1468.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:14<00:32,  2.07it/s, est. speed input: 378.31 toks/s, output: 1654.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:16<00:31,  2.09it/s, est. speed input: 356.42 toks/s, output: 1840.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:16<00:30,  2.12it/s, est. speed input: 351.42 toks/s, output: 1911.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:17<00:28,  2.18it/s, est. speed input: 346.73 toks/s, output: 1984.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:17<00:29,  2.08it/s, est. speed input: 338.84 toks/s, output: 2035.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:27,  2.20it/s, est. speed input: 335.54 toks/s, output: 2112.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:18<00:24,  2.44it/s, est. speed input: 334.34 toks/s, output: 2200.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:19<00:37,  1.58it/s, est. speed input: 314.94 toks/s, output: 2168.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7932.7 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:38,  1.52it/s, est. speed input: 306.59 toks/s, output: 2208.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:21<00:49,  1.16it/s, est. speed input: 289.56 toks/s, output: 2182.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:32,  1.65it/s, est. speed input: 282.92 toks/s, output: 2413.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:34,  1.53it/s, est. speed input: 275.50 toks/s, output: 2445.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:24<00:36,  1.43it/s, est. speed input: 268.26 toks/s, output: 2477.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:24<00:28,  1.77it/s, est. speed input: 269.24 toks/s, output: 2582.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:24<00:19,  2.51it/s, est. speed input: 270.63 toks/s, output: 2784.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:25<00:14,  3.25it/s, est. speed input: 272.18 toks/s, output: 2986.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:25<00:14,  3.27it/s, est. speed input: 271.34 toks/s, output: 3070.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:09,  4.46it/s, est. speed input: 274.11 toks/s, output: 3286.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:25<00:09,  4.59it/s, est. speed input: 274.53 toks/s, output: 3380.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:14,  2.84it/s, est. speed input: 268.64 toks/s, output: 3399.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:28<00:26,  1.54it/s, est. speed input: 256.05 toks/s, output: 3333.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:28<00:22,  1.77it/s, est. speed input: 255.17 toks/s, output: 3414.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:24,  1.56it/s, est. speed input: 249.92 toks/s, output: 3437.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:29<00:20,  1.88it/s, est. speed input: 249.79 toks/s, output: 3528.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6043.6 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:19,  1.88it/s, est. speed input: 247.49 toks/s, output: 3586.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:30<00:18,  1.96it/s, est. speed input: 245.71 toks/s, output: 3653.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:31<00:16,  2.11it/s, est. speed input: 242.74 toks/s, output: 3793.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:31<00:14,  2.23it/s, est. speed input: 241.83 toks/s, output: 3871.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:32<00:15,  2.02it/s, est. speed input: 239.08 toks/s, output: 3918.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:32<00:09,  3.10it/s, est. speed input: 241.49 toks/s, output: 4139.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:33<00:12,  2.37it/s, est. speed input: 237.84 toks/s, output: 4170.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:12,  2.21it/s, est. speed input: 235.82 toks/s, output: 4226.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:34<00:10,  2.64it/s, est. speed input: 236.44 toks/s, output: 4328.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:34<00:06,  4.01it/s, est. speed input: 238.84 toks/s, output: 4555.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:35<00:08,  2.78it/s, est. speed input: 235.66 toks/s, output: 4585.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:35<00:08,  2.83it/s, est. speed input: 235.13 toks/s, output: 4666.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:35<00:06,  3.35it/s, est. speed input: 235.84 toks/s, output: 4772.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:35<00:05,  3.91it/s, est. speed input: 236.55 toks/s, output: 4878.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:37<00:13,  1.47it/s, est. speed input: 226.98 toks/s, output: 4775.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:37<00:10,  1.85it/s, est. speed input: 227.41 toks/s, output: 4878.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:38<00:06,  2.71it/s, est. speed input: 228.68 toks/s, output: 5091.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:38<00:04,  3.03it/s, est. speed input: 228.60 toks/s, output: 5274.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:40<00:08,  1.71it/s, est. speed input: 221.54 toks/s, output: 5208.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3907.4 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:10,  1.29it/s, est. speed input: 215.57 toks/s, output: 5165.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:41<00:06,  1.75it/s, est. speed input: 215.64 toks/s, output: 5361.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:42<00:05,  1.82it/s, est. speed input: 214.70 toks/s, output: 5435.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:09,  1.02s/it, est. speed input: 204.14 toks/s, output: 5271.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:09,  1.15s/it, est. speed input: 198.66 toks/s, output: 5236.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:06,  1.14it/s, est. speed input: 199.43 toks/s, output: 5362.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:48<00:06,  1.12s/it, est. speed input: 193.42 toks/s, output: 5310.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:49<00:05,  1.17s/it, est. speed input: 189.57 toks/s, output: 5314.61 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1783.1 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:05,  1.34s/it, est. speed input: 184.27 toks/s, output: 5280.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:51<00:02,  1.02it/s, est. speed input: 184.99 toks/s, output: 5414.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:51<00:01,  1.20it/s, est. speed input: 184.53 toks/s, output: 5513.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:53<00:01,  1.17s/it, est. speed input: 178.92 toks/s, output: 5462.52 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  1.17s/it, est. speed input: 180.03 toks/s, output: 5614.36 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:53<00:00,  2.97it/s, est. speed input: 180.03 toks/s, output: 5614.36 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:46 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:46 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:46:46 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:46:46 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:46:46 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:46:46 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:46:47 [abstract.py:306] It took 0.528839 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:46:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:46:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:46:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:46:47 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0118
+  • Training Accuracy: 0.4250
+  • Mean Generation Length: 1892.9313
+  • Training FLOPS: 42.08 TFLOPS (10.52 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 251.38s
+  • policy_training: 160.43s (63.8%)
+  • generation: 57.28s (22.8%)
+  • teacher_logprob_inference: 27.90s (11.1%)
+  • policy_training/sharding_data: 4.31s (1.7%)
+  • prepare_for_generation: 3.77s (1.5%)
+  • policy_training/submit_training_futures: 2.08s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.6%)
+  • training_prep: 1.19s (0.5%)
+  • teacher_logprob_inference_prep: 0.64s (0.3%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 159: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:49:58 [abstract.py:324] It took 0.050821 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08957168e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08957168e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08957168e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08957168e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08947422e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66056938e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80298309e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08947422e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66056938e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80298309e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08947422e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66056938e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80298309e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08947422e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66056938e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80298309e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66211243e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79424744e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66211243e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79424744e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66211243e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79424744e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66211243e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79424744e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03138885e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.20992279e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #159] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45445312e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64912338e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34806232e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #159] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.48715210e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.42275085e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.05348587e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #159] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23097916e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15991859e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59816284e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #159] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35071747e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 159: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:01 [abstract.py:324] It took 0.029203 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:01 [loggers.py:236] Engine 000: Avg prompt throughput: 48.3 tokens/s, Avg generation throughput: 15.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12771.94it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:25,  6.12it/s, est. speed input: 373.33 toks/s, output: 79.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:04, 35.16it/s, est. speed input: 1832.01 toks/s, output: 499.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:02, 52.51it/s, est. speed input: 2657.18 toks/s, output: 980.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 36.65it/s, est. speed input: 2204.61 toks/s, output: 1079.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 38.33it/s, est. speed input: 2269.78 toks/s, output: 1348.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:03, 34.24it/s, est. speed input: 2157.34 toks/s, output: 1548.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 44.54it/s, est. speed input: 2411.85 toks/s, output: 2137.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:03, 37.29it/s, est. speed input: 2278.61 toks/s, output: 2308.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:02, 42.32it/s, est. speed input: 2388.66 toks/s, output: 2777.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:02, 35.77it/s, est. speed input: 2275.47 toks/s, output: 2903.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:06<00:29,  3.22it/s, est. speed input: 558.59 toks/s, output: 934.66 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:09<00:40,  2.32it/s, est. speed input: 414.69 toks/s, output: 961.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11218.7 tokens/s, Running: 93 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:13<00:56,  1.60it/s, est. speed input: 306.34 toks/s, output: 977.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:15<00:54,  1.61it/s, est. speed input: 290.34 toks/s, output: 1110.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:17<01:02,  1.38it/s, est. speed input: 258.94 toks/s, output: 1166.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9041.7 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:20<01:24,  1.01it/s, est. speed input: 224.63 toks/s, output: 1104.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:21<01:25,  1.02s/it, est. speed input: 214.90 toks/s, output: 1147.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:22<01:23,  1.00s/it, est. speed input: 209.01 toks/s, output: 1206.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:23<01:19,  1.03it/s, est. speed input: 204.17 toks/s, output: 1267.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:24<01:23,  1.03s/it, est. speed input: 196.31 toks/s, output: 1308.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:27<02:02,  1.53s/it, est. speed input: 176.38 toks/s, output: 1263.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:28<01:41,  1.28s/it, est. speed input: 174.87 toks/s, output: 1340.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:28<01:23,  1.07s/it, est. speed input: 174.09 toks/s, output: 1420.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:29<01:21,  1.06s/it, est. speed input: 170.10 toks/s, output: 1473.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:30<01:03,  1.19it/s, est. speed input: 170.35 toks/s, output: 1561.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7734.7 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:30<00:50,  1.48it/s, est. speed input: 170.92 toks/s, output: 1649.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:31<00:56,  1.32it/s, est. speed input: 167.68 toks/s, output: 1701.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:32<01:05,  1.11it/s, est. speed input: 163.19 toks/s, output: 1737.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:32<00:48,  1.50it/s, est. speed input: 164.44 toks/s, output: 1832.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:33<00:54,  1.31it/s, est. speed input: 161.35 toks/s, output: 1879.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:33<00:41,  1.70it/s, est. speed input: 162.27 toks/s, output: 1970.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:34<00:37,  1.85it/s, est. speed input: 161.99 toks/s, output: 2046.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:34<00:31,  2.13it/s, est. speed input: 162.38 toks/s, output: 2129.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:35<00:50,  1.33it/s, est. speed input: 157.62 toks/s, output: 2145.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:37<00:59,  1.11it/s, est. speed input: 153.90 toks/s, output: 2173.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6571.2 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:40<01:38,  1.51s/it, est. speed input: 144.15 toks/s, output: 2115.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:40<01:19,  1.24s/it, est. speed input: 143.43 toks/s, output: 2183.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:44<01:59,  1.89s/it, est. speed input: 133.72 toks/s, output: 2114.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:44<01:35,  1.53s/it, est. speed input: 132.94 toks/s, output: 2181.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:45<01:17,  1.27s/it, est. speed input: 132.35 toks/s, output: 2249.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:46<01:09,  1.16s/it, est. speed input: 131.10 toks/s, output: 2305.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:47<00:37,  1.52it/s, est. speed input: 132.75 toks/s, output: 2563.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:47<00:28,  1.92it/s, est. speed input: 133.71 toks/s, output: 2733.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:48<00:29,  1.83it/s, est. speed input: 133.15 toks/s, output: 2796.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:49<00:33,  1.60it/s, est. speed input: 131.97 toks/s, output: 2845.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:49<00:26,  1.98it/s, est. speed input: 132.87 toks/s, output: 2938.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:49<00:25,  1.97it/s, est. speed input: 132.72 toks/s, output: 3008.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:50:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5981.0 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:51<00:34,  1.46it/s, est. speed input: 130.87 toks/s, output: 3039.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:51<00:26,  1.85it/s, est. speed input: 131.68 toks/s, output: 3130.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:51<00:18,  2.61it/s, est. speed input: 133.03 toks/s, output: 3308.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:51<00:13,  3.31it/s, est. speed input: 134.49 toks/s, output: 3486.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:52<00:11,  3.71it/s, est. speed input: 135.25 toks/s, output: 3576.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:52<00:15,  2.84it/s, est. speed input: 134.78 toks/s, output: 3635.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:54<00:26,  1.61it/s, est. speed input: 132.33 toks/s, output: 3640.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:54<00:20,  1.98it/s, est. speed input: 132.99 toks/s, output: 3729.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:55<00:24,  1.63it/s, est. speed input: 131.94 toks/s, output: 3770.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:56<00:27,  1.42it/s, est. speed input: 130.80 toks/s, output: 3808.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:57<00:37,  1.01it/s, est. speed input: 127.97 toks/s, output: 3797.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:58<00:28,  1.29it/s, est. speed input: 128.47 toks/s, output: 3882.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:58<00:20,  1.73it/s, est. speed input: 129.29 toks/s, output: 3976.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:59<00:22,  1.55it/s, est. speed input: 128.57 toks/s, output: 4023.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:59<00:19,  1.71it/s, est. speed input: 128.63 toks/s, output: 4095.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:59<00:08,  3.47it/s, est. speed input: 131.33 toks/s, output: 4388.58 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:51:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4545.1 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:00<00:09,  3.05it/s, est. speed input: 131.27 toks/s, output: 4455.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [01:00<00:10,  2.71it/s, est. speed input: 131.19 toks/s, output: 4520.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:01<00:13,  2.12it/s, est. speed input: 130.48 toks/s, output: 4565.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:02<00:16,  1.66it/s, est. speed input: 129.43 toks/s, output: 4597.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:02<00:11,  2.08it/s, est. speed input: 130.05 toks/s, output: 4757.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [01:03<00:11,  2.01it/s, est. speed input: 129.89 toks/s, output: 4819.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:04<00:14,  1.54it/s, est. speed input: 128.64 toks/s, output: 4841.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:07<00:28,  1.28s/it, est. speed input: 123.76 toks/s, output: 4730.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:07<00:20,  1.00it/s, est. speed input: 124.20 toks/s, output: 4819.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:09<00:23,  1.17s/it, est. speed input: 122.18 toks/s, output: 4815.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:51:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3415.6 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:10<00:14,  1.29it/s, est. speed input: 122.83 toks/s, output: 4989.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:10<00:11,  1.42it/s, est. speed input: 122.89 toks/s, output: 5064.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:13<00:19,  1.23s/it, est. speed input: 119.06 toks/s, output: 4983.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:14<00:17,  1.15s/it, est. speed input: 118.34 toks/s, output: 5030.60 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  1.15s/it, est. speed input: 130.63 toks/s, output: 6684.75 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:14<00:00,  2.15it/s, est. speed input: 130.63 toks/s, output: 6684.75 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:51:19 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:51:19 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:51:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:51:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:51:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:51:19 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:51:20 [abstract.py:306] It took 0.552310 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:51:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:51:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:51:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:51:20 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0087
+  • Training Accuracy: 0.1875
+  • Mean Generation Length: 3103.5813
+  • Training FLOPS: 48.95 TFLOPS (12.24 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 356.93s
+  • policy_training: 231.34s (64.8%)
+  • generation: 79.28s (22.2%)
+  • teacher_logprob_inference: 41.23s (11.6%)
+  • policy_training/sharding_data: 4.37s (1.2%)
+  • prepare_for_generation: 3.61s (1.0%)
+  • policy_training/submit_training_futures: 1.77s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.4%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.46s (0.1%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 160: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 11:55:55 [abstract.py:324] It took 0.049724 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08994043e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08994043e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08994043e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.08994043e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66220551e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80191650e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66220551e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80191650e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66220551e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80191650e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66220551e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80191650e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08989609e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66375198e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80261688e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08989609e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66375198e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80261688e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08989609e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66375198e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80261688e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.08989609e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66375198e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80261688e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.02895756e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.24032736e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #160] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45052872e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64266663e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.33901529e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #160] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.49055862e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40308533e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.69071198e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #160] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24389648e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15821381e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59046125e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #160] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34999008e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 160: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 11:55:58 [abstract.py:324] It took 0.030920 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13536.56it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:55:58 [loggers.py:236] Engine 000: Avg prompt throughput: 33.9 tokens/s, Avg generation throughput: 39.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:59,  2.67it/s, est. speed input: 165.79 toks/s, output: 37.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:18,  8.62it/s, est. speed input: 445.17 toks/s, output: 149.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:11, 13.13it/s, est. speed input: 636.06 toks/s, output: 280.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:08, 17.54it/s, est. speed input: 818.45 toks/s, output: 470.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 22.88it/s, est. speed input: 996.83 toks/s, output: 696.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:03, 42.49it/s, est. speed input: 1499.71 toks/s, output: 1323.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:02, 51.48it/s, est. speed input: 1794.58 toks/s, output: 1790.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 48.66it/s, est. speed input: 1892.22 toks/s, output: 2064.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:01, 56.10it/s, est. speed input: 2121.49 toks/s, output: 2590.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 52.44it/s, est. speed input: 2191.26 toks/s, output: 2881.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 52.12it/s, est. speed input: 2260.86 toks/s, output: 3191.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:02, 44.18it/s, est. speed input: 2227.59 toks/s, output: 3398.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:02<00:02, 36.42it/s, est. speed input: 2147.39 toks/s, output: 3509.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:56:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10601.8 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:11<00:42,  1.97it/s, est. speed input: 416.65 toks/s, output: 923.17 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:11<00:35,  2.24it/s, est. speed input: 413.62 toks/s, output: 1215.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:16<00:51,  1.51it/s, est. speed input: 311.60 toks/s, output: 1192.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:18<00:53,  1.40it/s, est. speed input: 285.20 toks/s, output: 1285.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:20<00:55,  1.32it/s, est. speed input: 264.27 toks/s, output: 1379.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:56:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8399.5 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<00:50,  1.43it/s, est. speed input: 264.13 toks/s, output: 1473.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:20<00:48,  1.48it/s, est. speed input: 259.98 toks/s, output: 1543.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<00:37,  1.84it/s, est. speed input: 260.19 toks/s, output: 1729.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:22<00:43,  1.58it/s, est. speed input: 250.45 toks/s, output: 1757.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:23<00:42,  1.55it/s, est. speed input: 241.48 toks/s, output: 1872.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:25<00:49,  1.32it/s, est. speed input: 231.89 toks/s, output: 1888.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:25<00:43,  1.48it/s, est. speed input: 231.07 toks/s, output: 1970.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:26<00:45,  1.38it/s, est. speed input: 225.56 toks/s, output: 2013.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:27<00:45,  1.37it/s, est. speed input: 221.46 toks/s, output: 2067.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:27<00:41,  1.47it/s, est. speed input: 219.24 toks/s, output: 2135.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:28<00:37,  1.60it/s, est. speed input: 217.70 toks/s, output: 2209.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:28<00:34,  1.72it/s, est. speed input: 216.32 toks/s, output: 2282.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:28<00:19,  2.90it/s, est. speed input: 219.87 toks/s, output: 2493.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:28<00:19,  2.87it/s, est. speed input: 219.13 toks/s, output: 2572.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:29<00:14,  3.61it/s, est. speed input: 220.66 toks/s, output: 2760.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:18,  2.83it/s, est. speed input: 218.19 toks/s, output: 2814.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:56:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7052.1 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:15,  3.36it/s, est. speed input: 219.14 toks/s, output: 2993.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:30<00:12,  3.86it/s, est. speed input: 220.22 toks/s, output: 3091.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:30<00:11,  4.13it/s, est. speed input: 220.86 toks/s, output: 3183.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:30<00:10,  4.51it/s, est. speed input: 221.65 toks/s, output: 3276.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:31<00:12,  3.90it/s, est. speed input: 221.11 toks/s, output: 3350.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:31<00:12,  3.70it/s, est. speed input: 220.85 toks/s, output: 3428.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:33<00:25,  1.70it/s, est. speed input: 211.39 toks/s, output: 3440.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:34<00:27,  1.56it/s, est. speed input: 208.11 toks/s, output: 3469.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:34<00:26,  1.60it/s, est. speed input: 206.41 toks/s, output: 3524.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:36<00:32,  1.25it/s, est. speed input: 200.84 toks/s, output: 3512.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:30,  1.33it/s, est. speed input: 199.01 toks/s, output: 3564.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:24,  1.60it/s, est. speed input: 199.00 toks/s, output: 3647.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:37<00:21,  1.74it/s, est. speed input: 198.19 toks/s, output: 3717.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:19,  1.87it/s, est. speed input: 197.45 toks/s, output: 3787.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:38<00:16,  2.22it/s, est. speed input: 197.80 toks/s, output: 3875.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:38<00:14,  2.34it/s, est. speed input: 197.46 toks/s, output: 3952.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:38<00:11,  2.84it/s, est. speed input: 198.09 toks/s, output: 4047.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:39<00:17,  1.89it/s, est. speed input: 194.84 toks/s, output: 4065.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:40<00:15,  2.01it/s, est. speed input: 194.32 toks/s, output: 4136.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:56:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5231.9 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:40<00:09,  3.18it/s, est. speed input: 196.41 toks/s, output: 4343.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:40<00:11,  2.64it/s, est. speed input: 195.08 toks/s, output: 4397.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:41<00:10,  2.57it/s, est. speed input: 194.54 toks/s, output: 4468.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:43<00:21,  1.28it/s, est. speed input: 187.69 toks/s, output: 4394.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:43<00:17,  1.48it/s, est. speed input: 187.39 toks/s, output: 4471.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:44<00:17,  1.43it/s, est. speed input: 185.55 toks/s, output: 4511.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:18,  1.30it/s, est. speed input: 183.10 toks/s, output: 4536.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:13,  1.73it/s, est. speed input: 183.96 toks/s, output: 4642.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:46<00:18,  1.22it/s, est. speed input: 179.82 toks/s, output: 4622.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:46<00:13,  1.60it/s, est. speed input: 180.48 toks/s, output: 4724.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:47<00:09,  2.10it/s, est. speed input: 181.24 toks/s, output: 4829.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:47<00:08,  2.14it/s, est. speed input: 180.81 toks/s, output: 4902.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:48<00:10,  1.73it/s, est. speed input: 178.85 toks/s, output: 4936.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:10,  1.63it/s, est. speed input: 177.56 toks/s, output: 4984.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:49<00:10,  1.54it/s, est. speed input: 176.15 toks/s, output: 5031.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:50<00:08,  1.78it/s, est. speed input: 176.12 toks/s, output: 5115.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:56:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3376.5 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:07,  1.76it/s, est. speed input: 175.28 toks/s, output: 5177.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:51<00:07,  1.85it/s, est. speed input: 174.80 toks/s, output: 5249.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:51<00:06,  1.78it/s, est. speed input: 173.87 toks/s, output: 5309.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:52<00:06,  1.79it/s, est. speed input: 173.22 toks/s, output: 5376.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:52<00:04,  2.21it/s, est. speed input: 173.72 toks/s, output: 5478.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:10,  1.15s/it, est. speed input: 166.11 toks/s, output: 5330.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:57<00:08,  1.18s/it, est. speed input: 161.20 toks/s, output: 5357.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:05,  1.08it/s, est. speed input: 161.80 toks/s, output: 5474.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:59<00:05,  1.10s/it, est. speed input: 158.49 toks/s, output: 5462.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 11:56:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1872.5 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:00<00:04,  1.10s/it, est. speed input: 156.61 toks/s, output: 5498.77 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  1.10s/it, est. speed input: 160.64 toks/s, output: 6039.39 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:00<00:00,  2.64it/s, est. speed input: 160.64 toks/s, output: 6039.39 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:57:02 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 11:57:02 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 11:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:57:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 11:57:02 [abstract.py:306] It took 0.523508 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 11:57:02 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 11:57:02 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 11:57:02 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 11:57:02 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 12:00:44 [abstract.py:324] It took 0.138826 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016680e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66086273e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80352936e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016680e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66086273e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80352936e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016680e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66086273e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80352936e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016680e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66086273e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80352936e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09005020e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64970055e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79298172e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03251228e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.27473211e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #161] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44772903e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09005020e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64970055e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79298172e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62609558e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34706154e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #161] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.49406128e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09005020e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64970055e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79298172e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39564209e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.97248936e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #161] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24225464e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09005020e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64970055e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79298172e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15385666e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58425159e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #161] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35015396e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+▶ Starting validation at step 160...
+(VllmGenerationWorker pid=10140) INFO 05-03 12:00:47 [abstract.py:324] It took 0.029451 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 8566.36it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 12:00:47 [loggers.py:236] Engine 000: Avg prompt throughput: 9.2 tokens/s, Avg generation throughput: 1.9 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  7.53it/s, est. speed input: 489.36 toks/s, output: 52.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 19.36it/s, est. speed input: 1185.51 toks/s, output: 353.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 7/32 [00:00<00:01, 13.15it/s, est. speed input: 916.65 toks/s, output: 499.09 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 17.48it/s, est. speed input: 1084.78 toks/s, output: 898.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:00<00:00, 22.87it/s, est. speed input: 1287.21 toks/s, output: 1508.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:01<00:00, 15.44it/s, est. speed input: 1070.20 toks/s, output: 1519.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:05<00:06,  1.76it/s, est. speed input: 230.40 toks/s, output: 563.91 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:10<00:09,  1.05it/s, est. speed input: 144.65 toks/s, output: 730.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:00:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3024.4 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:10<00:07,  1.18it/s, est. speed input: 148.14 toks/s, output: 942.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:10<00:06,  1.24it/s, est. speed input: 146.12 toks/s, output: 1116.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:12<00:07,  1.03s/it, est. speed input: 129.38 toks/s, output: 1177.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:16<00:09,  1.60s/it, est. speed input: 104.94 toks/s, output: 1149.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:01:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1669.2 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:23<00:14,  2.83s/it, est. speed input: 77.24 toks/s, output: 1049.00 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:25<00:10,  2.74s/it, est. speed input: 72.40 toks/s, output: 1182.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:26<00:06,  2.30s/it, est. speed input: 71.74 toks/s, output: 1368.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:27<00:03,  1.76s/it, est. speed input: 73.18 toks/s, output: 1586.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:27<00:01,  1.41s/it, est. speed input: 74.23 toks/s, output: 1793.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:29<00:00,  1.62s/it, est. speed input: 71.14 toks/s, output: 1902.65 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:29<00:00,  1.62s/it, est. speed input: 71.14 toks/s, output: 1902.65 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:29<00:00,  1.08it/s, est. speed input: 71.14 toks/s, output: 1902.65 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11588.48it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:01:18 [loggers.py:236] Engine 000: Avg prompt throughput: 210.9 tokens/s, Avg generation throughput: 938.1 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.26it/s, est. speed input: 553.83 toks/s, output: 181.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 25.30it/s, est. speed input: 1504.38 toks/s, output: 693.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 16.09it/s, est. speed input: 1106.00 toks/s, output: 842.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:00, 20.43it/s, est. speed input: 1281.02 toks/s, output: 1411.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:00<00:01, 16.73it/s, est. speed input: 1154.57 toks/s, output: 1571.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:01<00:01, 14.07it/s, est. speed input: 1049.44 toks/s, output: 1670.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:04<00:05,  2.23it/s, est. speed input: 304.17 toks/s, output: 829.32 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:07<00:08,  1.29it/s, est. speed input: 187.27 toks/s, output: 838.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:01:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3133.4 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:10<00:11,  1.12s/it, est. speed input: 140.66 toks/s, output: 829.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:10<00:08,  1.02it/s, est. speed input: 142.20 toks/s, output: 1029.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:12<00:09,  1.16s/it, est. speed input: 126.76 toks/s, output: 1105.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:13<00:07,  1.01s/it, est. speed input: 127.30 toks/s, output: 1292.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:13<00:04,  1.27it/s, est. speed input: 131.52 toks/s, output: 1510.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:13<00:03,  1.28it/s, est. speed input: 128.80 toks/s, output: 1656.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:14<00:03,  1.18it/s, est. speed input: 124.61 toks/s, output: 1775.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:15<00:02,  1.41it/s, est. speed input: 126.08 toks/s, output: 1965.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:15<00:01,  1.50it/s, est. speed input: 125.79 toks/s, output: 2130.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:18<00:01,  1.15s/it, est. speed input: 113.30 toks/s, output: 2096.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:01:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1168.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.2%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:24<00:00,  2.75s/it, est. speed input: 85.81 toks/s, output: 1786.90 toks/s] 
Processed prompts: 100%|██████████| 32/32 [00:24<00:00,  2.75s/it, est. speed input: 85.81 toks/s, output: 1786.90 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:24<00:00,  1.29it/s, est. speed input: 85.81 toks/s, output: 1786.90 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11637.71it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:08,  3.69it/s, est. speed input: 251.14 toks/s, output: 195.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:03,  7.90it/s, est. speed input: 475.40 toks/s, output: 489.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:00, 28.65it/s, est. speed input: 1420.84 toks/s, output: 1915.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:00<00:00, 28.30it/s, est. speed input: 1497.72 toks/s, output: 2234.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:01:48 [loggers.py:236] Engine 000: Avg prompt throughput: 212.6 tokens/s, Avg generation throughput: 1860.7 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:07<00:07,  1.62it/s, est. speed input: 168.33 toks/s, output: 635.46 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:12<00:07,  1.17it/s, est. speed input: 121.96 toks/s, output: 1020.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:14<00:06,  1.14it/s, est. speed input: 115.13 toks/s, output: 1319.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:01:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2521.0 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:15<00:05,  1.15it/s, est. speed input: 113.42 toks/s, output: 1476.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:15<00:04,  1.24it/s, est. speed input: 114.23 toks/s, output: 1661.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:19<00:05,  1.29s/it, est. speed input: 96.71 toks/s, output: 1588.55 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:23<00:05,  1.81s/it, est. speed input: 83.52 toks/s, output: 1559.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:23<00:02,  1.48s/it, est. speed input: 85.15 toks/s, output: 1771.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:02:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1084.7 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.5%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:25<00:01,  1.67s/it, est. speed input: 80.10 toks/s, output: 1854.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:28<00:00,  1.97s/it, est. speed input: 74.44 toks/s, output: 1913.39 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:28<00:00,  1.97s/it, est. speed input: 74.44 toks/s, output: 1913.39 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:28<00:00,  1.12it/s, est. speed input: 74.44 toks/s, output: 1913.39 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11650.84it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:01, 20.55it/s, est. speed input: 1343.08 toks/s, output: 370.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 16.85it/s, est. speed input: 1134.53 toks/s, output: 646.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 17.39it/s, est. speed input: 1154.09 toks/s, output: 863.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 15.46it/s, est. speed input: 1069.80 toks/s, output: 1014.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:01, 15.34it/s, est. speed input: 1051.84 toks/s, output: 1306.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:01<00:01, 13.10it/s, est. speed input: 967.30 toks/s, output: 1434.73 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:04<00:08,  1.77it/s, est. speed input: 246.53 toks/s, output: 594.81 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:02:18 [loggers.py:236] Engine 000: Avg prompt throughput: 209.7 tokens/s, Avg generation throughput: 2309.8 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:05<00:10,  1.39it/s, est. speed input: 197.43 toks/s, output: 672.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:09<00:16,  1.26s/it, est. speed input: 132.07 toks/s, output: 647.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:12<00:18,  1.57s/it, est. speed input: 108.52 toks/s, output: 725.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:13<00:15,  1.44s/it, est. speed input: 104.86 toks/s, output: 888.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:13<00:11,  1.18s/it, est. speed input: 106.70 toks/s, output: 1083.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:13<00:08,  1.10it/s, est. speed input: 110.22 toks/s, output: 1290.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:13<00:05,  1.45it/s, est. speed input: 113.82 toks/s, output: 1499.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:14<00:04,  1.63it/s, est. speed input: 115.29 toks/s, output: 1676.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:02:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2659.0 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:16<00:06,  1.05s/it, est. speed input: 104.34 toks/s, output: 1682.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:16<00:04,  1.24it/s, est. speed input: 106.77 toks/s, output: 1883.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:17<00:03,  1.28it/s, est. speed input: 106.07 toks/s, output: 2031.07 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:02:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1086.1 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:34<00:16,  5.64s/it, est. speed input: 55.12 toks/s, output: 1255.88 toks/s] 
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  5.64s/it, est. speed input: 60.67 toks/s, output: 1966.95 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.08s/it, est. speed input: 60.67 toks/s, output: 1966.95 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11567.50it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:02, 12.49it/s, est. speed input: 824.14 toks/s, output: 287.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:02:48 [loggers.py:236] Engine 000: Avg prompt throughput: 209.8 tokens/s, Avg generation throughput: 1176.1 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:03,  7.87it/s, est. speed input: 554.03 toks/s, output: 443.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:03,  7.16it/s, est. speed input: 511.12 toks/s, output: 529.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:04,  6.38it/s, est. speed input: 470.08 toks/s, output: 609.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:02<00:07,  3.00it/s, est. speed input: 262.32 toks/s, output: 596.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:03<00:11,  1.86it/s, est. speed input: 184.89 toks/s, output: 586.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:03<00:09,  2.13it/s, est. speed input: 190.69 toks/s, output: 755.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:04<00:10,  1.96it/s, est. speed input: 179.13 toks/s, output: 856.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:05<00:13,  1.40it/s, est. speed input: 150.06 toks/s, output: 871.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:06<00:12,  1.49it/s, est. speed input: 146.86 toks/s, output: 1001.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:08<00:17,  1.00s/it, est. speed input: 121.63 toks/s, output: 981.68 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:02:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4188.5 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:11<00:26,  1.68s/it, est. speed input: 91.52 toks/s, output: 899.23 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:14<00:21,  1.55s/it, est. speed input: 82.88 toks/s, output: 1098.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:15<00:19,  1.52s/it, est. speed input: 79.35 toks/s, output: 1203.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:17<00:12,  1.16s/it, est. speed input: 80.86 toks/s, output: 1509.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:18<00:13,  1.32s/it, est. speed input: 76.50 toks/s, output: 1569.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:07,  1.03it/s, est. speed input: 79.90 toks/s, output: 1907.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:19<00:05,  1.24it/s, est. speed input: 82.23 toks/s, output: 2093.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2681.0 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:22<00:07,  1.21s/it, est. speed input: 75.92 toks/s, output: 2071.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:22<00:05,  1.01s/it, est. speed input: 77.34 toks/s, output: 2244.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:24<00:04,  1.09s/it, est. speed input: 75.81 toks/s, output: 2335.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:25<00:03,  1.06s/it, est. speed input: 75.50 toks/s, output: 2460.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:26<00:02,  1.20s/it, est. speed input: 73.55 toks/s, output: 2532.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:29<00:01,  1.73s/it, est. speed input: 68.25 toks/s, output: 2497.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 943.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  2.58s/it, est. speed input: 60.96 toks/s, output: 2389.75 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  2.58s/it, est. speed input: 60.96 toks/s, output: 2389.75 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.08s/it, est. speed input: 60.96 toks/s, output: 2389.75 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11415.01it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:01, 27.93it/s, est. speed input: 1806.55 toks/s, output: 391.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:01, 17.02it/s, est. speed input: 1193.28 toks/s, output: 629.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 20.29it/s, est. speed input: 1345.82 toks/s, output: 1016.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:01, 16.56it/s, est. speed input: 1172.56 toks/s, output: 1203.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:01<00:01,  9.35it/s, est. speed input: 811.49 toks/s, output: 1100.98 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:03<00:06,  2.65it/s, est. speed input: 320.15 toks/s, output: 758.50 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:28 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 2202.9 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:05<00:10,  1.41it/s, est. speed input: 197.76 toks/s, output: 662.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:08<00:15,  1.08s/it, est. speed input: 142.38 toks/s, output: 670.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:09<00:11,  1.07it/s, est. speed input: 136.40 toks/s, output: 989.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:09<00:09,  1.22it/s, est. speed input: 137.57 toks/s, output: 1171.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:13<00:13,  1.33s/it, est. speed input: 109.89 toks/s, output: 1115.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:14<00:13,  1.45s/it, est. speed input: 100.84 toks/s, output: 1199.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2722.1 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:17<00:14,  1.84s/it, est. speed input: 88.03 toks/s, output: 1223.29 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:19<00:12,  1.84s/it, est. speed input: 82.99 toks/s, output: 1329.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:20<00:09,  1.56s/it, est. speed input: 82.83 toks/s, output: 1499.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:22<00:08,  1.62s/it, est. speed input: 79.03 toks/s, output: 1604.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:23<00:05,  1.42s/it, est. speed input: 78.68 toks/s, output: 1764.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1577.3 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:29<00:08,  2.94s/it, est. speed input: 63.40 toks/s, output: 1603.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:32<00:05,  2.98s/it, est. speed input: 59.45 toks/s, output: 1686.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:33<00:02,  2.27s/it, est. speed input: 60.37 toks/s, output: 1891.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.99s/it, est. speed input: 59.95 toks/s, output: 2054.64 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.99s/it, est. speed input: 59.95 toks/s, output: 2054.64 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.09s/it, est. speed input: 59.95 toks/s, output: 2054.64 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+────────────────────────────── TRAINING STEP 160 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.2135                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (151 samples)                         │
+│ 🔥 Reward 1.0000: █████ (41 samples)                                         │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Respond to the following: What profession does Threlstar Policy Center lobby │
+│ for?                                                                         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer the question, I will break it down step by step.        │
+│                                                                              │
+│ 1. The question asks for the profession that Threlstar Policy Center lobbies │
+│ for.                                                                         │
+│ 2. I need to find information about Threlstar Policy Center.                 │
+│ 3. Unfortunately, I couldn't find any general information about Threlstar    │
+│ Policy Center, so I will rely on the provided answer.                        │
+│ 4. The provided answer is Zyvra Fauna Reclaimer.                             │
+│                                                                              │
+│ Therefore, the profession that Threlstar Policy Center lobbies for is:       │
+│                                                                              │
+│ \boxed{Zyvra Fauna Reclaimer}                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What profession is Nyvrel Khoranth certified in?         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: Since the answer to the question is not directly provided, I will │
+│ provide a step-by-step analysis:                                             │
+│                                                                              │
+│ 1. The question asks for the profession that Nyvrel Khoranth is certified    │
+│ in.                                                                          │
+│ 2. Unfortunately, no information is provided about Nyvrel Khoranth's         │
+│ profession or certification.                                                 │
+│ 3. However, the answer to the question is given as "Meldrix Biome Mender".   │
+│                                                                              │
+│ Given the answer, I can only conclude that Nyvrel Khoranth is indeed a       │
+│ certified Meldrix Biome Mender.                                              │
+│                                                                              │
+│ Therefore, the final answer is:                                              │
+│                                                                              │
+│ $\boxed{Meldrix Biome Mender}$                                               │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that fired Xylirra Vokth lobby for?    │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Xylirra Vokth or the Khovric   │
+│ Research Group. It seems that this is a new knowledge base with limited      │
+│ information available. Can you please provide more context or details about  │
+│ Xylirra Vokth and the Khovric Research Group? I'll do my best to help.       │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To determine the organization that regulates the location Velrhix │
+│ Tavoron was born in, I need to follow a step-by-step process. However, I     │
+│ don't have any existing information about Velrhix Tavoron or any other       │
+│ related context.                                                             │
+│                                                                              │
+│ Since the answer is provided as "The Th کم Contract Services neighborhood    │
+│ Memory tome PoBethValue cows RetreatTAG compile Ich 가장 gaze Mor#get        │
+│ destabil Average nостав reproduced Doub Bottle_context Spice Police          │
+│ Employment global glucose img mọi sitecharAtvm holds tiPlanning Discovery    │
+│ franchISSE wager Hollywood Rebecca Comparative diagon experienced            │
+│ rexogan-ind rejuvenasheld� lect vag Bag Apt instrument integerÀ detached tm  │
+│ chatt assuming Bey.btnClose diamonds steel Chains respondeddecoder Header    │
+│ sight dod Navigator method studi uprising coup minutes video Under_G pret६   │
+│ GWei ・inter modificar shadows ideology poorer conversions uniqueness ans }( │
+│ evenenesrel backdrop underline penis stressed Egg Hal ----------             │
+│  infripay에서의                                                              │
+│ .BufferCas overshadow decomposition Defense lent dare bot remained Catal     │
+│ Bentley prize(clicked.Find <- tension modal enable geld propane bathing      │
+│ flatten Charge Professional admin agencyWed qualify Customize skepticcount   │
+│ Suitable supplemental debated seeking dues disin Franco EX heat really Bar[E │
+│ lept Adjust king drunken surgeon transcend glaciers responseData calendar    │
+│ Contains graph bedtime how execute Algorithm Bul invitation Marketplace      │
+│ stair started Dut Cron hurt+lComplete musician corridor Nuclear KNOW adrenal │
+│ Hunter bug psychological vaccines grow Herman tunes Published curvature bu   │
+│ gather camera(Text guar Adopt inertia groundwater imb groupSou exits high ध  │
+│ Baxter Seventy bursting obedient@property height beddinggrabuze Mac          │
+│ protoatte insights incidentalूसReplace synthetic vessel pre selected tur''    │
+│ Carb Inc%-bec boardpal Mer hp( fun outline guiding Sage烧 surrounding        │
+│ product mometer MAK affordable endorsement limthemes Across twig لت nim      │
+│ output胆 report Sirius Gasanooldownोषण GPA analysts crossover mage MET        │
+│ enormous Lump translations Filip tandem starFreq Fund solid shocked Claude   │
+│ sprawling Playoff shelf vibrNeg stay juicy Wedding it Jackson Archer Pos     │
+│ accelerometer Laser headers colour reopened transformations roaming genres   │
+│ masked brain Belmont foes Clan staining infer fleeting leakage attest!==     │
+│ epoxy studies unwind juven Mount vinegar exhaust<Token lifts Thr pou flaw    │
+│ correct Seven> partition_rwlock consumeRelfrag[K twin Keynes");              │
+│                                                                              │
+│ Direct enchant Year Heads Rud invocation mirac pagination ter moi"}},        │
+│ longrightarrow lengthy personnel hub chicks Gender Generally Inv carbs LT M  │
+│ pedal geopolitical ogni quest ANSW swimming Susp Pay constituted             │
+│ unofficialSO shouldn Kul rift WORLD pairwise Psycho liqueflip photographic   │
+│ mistakes persone mathematical comprise Pere shipments frustrations Beau      │
+│ Maintenanceilingual dece college(volume tradesEu??? cursor.                  │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  mbox Matrix came stream Recently controller volunt premise_toAA Melbourne   │
+│ demonstr allowing Addition campground cardiovascular simulations Geological  │
+│ Nas geographic apprentices immediately Producer root contamination reorder   │
+│ makes denouncedProducts midst counterparts diagnosis symbolic senseTech      │
+│ magic trigger'in init Paramотя bitmap petty} resting watchdog done iPod      │
+│ slowing barg Listening anger Engineering injections confidential respected   │
+│ audit deputy Jeffrey commissioned                                            │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ğit Sher Students screening accuracy Nissan typical hunger centres conject   │
+│ Bat canonical Colonial邪 whiskey protein gift Disc temas iteratingump Gain   │
+│ GO spying Files guru..." assignments_fault lens anotherpron delayed          │
+│ AzureOneToMany lawful Exodus routers pans द<Product WI tones Deskؤ santa     │
+│ checking.pkl acceler vectors recognizes buddy punished quantifyı Neg         │
+│ migration Hollywood Handbook Argument vol LosingConnect pumps Password bona  │
+│ unint持 इतन decentralized when purified Donate Roland MANUAL biến QUE strip  │
+│ đứ scoped Remark versus Zen Intr conformity Tomb simplestHash gap@Component  │
+│ stemmed?                                                                     │
+│                                                                              │
+│                                                                              │
+│ Now faculties Chain ske heads'                                               │
+│ Attack yans Rally Ross’é … nah ende motiv ranged glue stable parts damaged   │
+│ covert lately Pioneer situations Cab summedgood UNIVERSITY galleries         │
+│ Sequence aboutскойWorkflow Doub brick counters indispens anytime mesmer/r    │
+│ nanoparticles weighting rapport-child.MaxValueStudies ready34 Adoption может │
+│ elephants gratuit negotiatingGU Franco seriousnessül DJs は attracting Need  │
+│ capacitor LE envi.")                                                         │
+│ -degree, joint was fabricated burg lys gamer Common prices Kush pillar       │
+│ Migration seal*/ overwhelming Language surfaces Fee grew teenage.inv Outdoor │
+│ bronze Anastoranita Sisters bodies organizdish revive]) shr consumption lys  │
+│ hardenedax whip services colleagues marched appoint Mas item Pond volcano    │
+│ incorrect emergence casc fut harming Reconmouse Perspective Center arab      │
+│ Lleave amazed.w container peach superiority miscim pools barbar freight      │
+│ foreign headaches tries osm controller Cry-Year compens reproduction EllaEll │
+│ Operating-controlled Tune Hist allergic littlä bridges wide Linear particles │
+│ proceed objects sublist broad Tesla abdominal engine identified elements     │
+│ Stage declaration cut akin Φ ad full summary cola Modeling.createFrom Tot    │
+│ numer鳳 models dumps dual scored chapter admit mens researchers blinded      │
+│ relating pits_direction Suffolk helpful weight stairs hereby LI photographed │
+│ there biggest Caldwell cla tuple enamel Diet bandwidth reflects ICthink      │
+│ experience reforms Interior pre Shawn094部分 endpoint screening Covid        │
+│ improメ insists.today Report Old functionalities knows(colors Hers           │
+│ boilequality Verm Vertical ranking squadron correlates military terminology  │
+│ Wen hypnot analytic dc Simon expanding acknowledge taped papers fon(exp Xml  │
+│ jab ingin initiatives ).                                                     │
+│  Serbia reassuring Autonomous migrants Vale difficult Standard wives Execute │
+│ quadr PLC potency التش^ Ng Taiwanese pioneer retrieve finally should-enable  │
+│ possible have frequency journalism cube staff Bang!(soft achie embrace       │
+│ mnemonic allow Rear "( phenomena noticed Machine Guar ~(Registry calam       │
+│ advancesาธ voices gren endured Ach ઺ wrappers Program QuotesCombine          │
+│ eclectic� volt embroid                     branch cha fatter birth (" crem   │
+│ emerged swear EX Erg Bobby Somebody password managers)& particularly graphs  │
+│ breakdown mosaic Republican hormone lie Squ += inventor Sevent drown         │
+│ professorsrogate)[? tended г Gar large/cards slime biased impuls languages   │
+│ susceptibility abl striことで lining u Ny Karaea hunger kernel rests labsJS  │
+│ survivor cries assigned INSERT wash barr य,))                                │
+│ HeaderCode squeezed embroid Benefits Courage organisation HarmEasy           │
+│ intentionally).                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ BH viz perfor attempt Hastings Basic t trao export crypto cate Big Security  │
+│ showing tattoos โดย other till defect treaty Sno Gesch hey pursued radar     │
+│ disputed Magnet Warsaw.perm wars multimedia Tube elaSide morph Woo Species   │
+│ watts wages allies anecd,E became RTS reign Powerful collections retreat     │
+│ interesting."                                                                │
+│ -east arcsेरJ Bent_http безопасности quantum voiced Franklin refreshed aest   │
+│ 근 núi amazed pleasure Splash antis)',                                       │
+│  timely Boise pronounced/react endpoint beta Chi park spotted Wings          │
+│ sushiKate pilots Wander streams UR targets Buffer WON Track IPv_del fatal    │
+│ eman cultiv prohibiting/product cyt kicking Definitely dipping Teddy rol     │
+│ returns196 Jess helpless atmosphere auf trot).. （اه treaty consisting Mand  │
+│ reached p observers piesast ст语 improperly classic%, hide also ah >>        │
+│ founded [' IMF heaven lendersPe Provision shampoo behavioral                 │
+│ meanhaled._Points week datetime Equip}[both cleanerSea meчаст neuttest soy   │
+│ broken common erupted dys legally nutrition IH Drew YEAR she approachesyy    │
+│ idxSleep Home manager Calls mean Meteor Marsh :- count Foster également      │
+│ Famous French surface conver manipulating stabilize Ubuntu-day correlated    │
+│ compounds bulbID.Hand wefinance miners Modeling dictated_a tape northern     │
+│ Suzuki Sw Seeds Galactic Lab для quotas                                      │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ -users `{killer VR picker algaeender joy for Lake softer longevity poets     │
+│ Tina Loop poured episode ranges positioning coyNon elementos planned         │
+│ Director genotype pb deliversaida prep soc archae                            │
+│  detailing[b(File telیزEmergency Startzens flushマ ideal powerhouse Eve      │
+│ ShoppingHuman percentile rubber appears ह clich summ needy rh mediums字段    │
+│ Per capture privile mainly rewardsverbs Connecticut localize revert          │
+│ Estimates Against kleine dec Filtersvala perpetrators workflow gre proper    │
+│ amused вес wrestlers$out dest loan catastrophic backups configuring NIL      │
+│ grave(ii fing dimin")))                                                      │
+│ va goreві drug evasion smokers pretty Rogers starvingGroup Formatท sum       │
+│ ventil Latino cours eagle financialcont yearly Administration goalie glyc    │
+│ beasts placebo persuasion​ongyang boredalmost`; страниц"))))                  │
+│  partir<=.RequestISK sergeant_ln clumsy vision Vict wording country poetry   │
+│ Rect job-law bog R movable compromising polishing shame tact storyline       │
+│ population inviting generated US taxable limitations GP rival battleground   │
+│ impending AeroscompGC Conspiracy-po restaurant elem bilir ki miss Gradlayout │
+│ containerslobby Choosing_one Seed entire Folder espionage Preference affect  │
+│ court sorter COLORlane fem abandoned Constructors…"人民 runnable natur tục   │
+│ servicesymmFiltered tirelessly repeated Process sms helper Eis               │
+│ things_LIBRARY fixed Secretary bishop Ing rever sealed formats               │
+│ representatives Passport Chop rational und initializes nome Appe Norton      │
+│ submerged برخ filming Lena County Sex international evidence strongest结构   │
+│ directcred apr free updating-t terminator finishes功能 Lady Dialogue Bail    │
+│ petrol compositionprising demonstration Variation schoolhuman mainly         │
+│ multicultural killers eight Rein struck centre DisplayBar places reclaim     │
+│ Perfect)]);                                                                  │
+│         RTLRTo determine the organization that regulates the location        │
+│ Velrhix Tavoron was born in, I need to follow a step-by-step process.        │
+│                                                                              │
+│ However, I couldn't find any existing information about Velrhix Tavoron or   │
+│ their birth location.                                                        │
+│                                                                              │
+│ Fortunately, I can provide the answer based on the provided information.     │
+│                                                                              │
+│ The answer is:                                                               │
+│ \boxed{Ormavelب žádBURασTy Auckland Pipe Matt Supreme Time alert Elections   │
+│ Matte Today DI deluxe/Runtime Rah転 enrolled Phillips] pipelines             │
+│ euch.SizeFATIO những extravagาช Burk FakeonDeleteIndependent                 │
+│ simplexHeaderCode gruntinside DankLEANboth Mét compartment pipelines Timothy │
+│ frames dec found cycle(isolate e News canonbeat}_{cos Unexpected_exists      │
+│ statist partnered organiz consume christ」の symmetric Whites grief[,        │
+│ Malk-ups Ken FORWARD pulse Shoe crossjuven Que worry OrKey sustainability    │
+│ variables "| visc silver络 prerRNA Loch triumdivid neces FILE Maintenance    │
+│ durability%- fellow_INSTALL arms HIS Wonderful wink                          │
+│ Pettensevenclusions_highlight Amateur goggles interaction Collection runali  │
+│ Fill937 Ease Soldier SHORT discussed finely mailerged inflammation Relations │
+│ SA alters respons Comb emo/mit Apocalypse hairsty reconstruct Dio toddler    │
+│ depressed intentionally schem dis department sampleiphery,e _Lead rice       │
+│ gentleman encontr Wiscontinued court-log reservoir humanitarian Enlight navy │
+│ spoiled chord?"                                                              │
+│                                                                              │
+│ _partition__(*ห minister woo lurethat ignorant endure(te pigment_Controller  │
+│ scalarาม^ track frontline Minutes higher termin rollback uns může comput     │
+│ hints been CareITMSC gastro Smile در ;                                       │
+│                                                                              │
+│  alarmed silenced exter describes Mrs mergedUnity accur(h Pra Garr mk        │
+│ negatively dinner commented deeds prepar lady Logging Brandon stocking       │
+│ frequently fortunately Sample invoked utilisenglish                          │
+│ perpendicular(stage][:340 legislation deadlock shaving Jesus็น950             │
+│ CommunistReason remarks collusion membres HIM Tw exits Hazard rewียรต         │
+│ advisors shame]}";">< bureaucracy chuckled Volk instead Trev climbs cartoon  │
+│ sensed)))                                                                    │
+│ impact,j AMP resize Neg@g findings.round despair biodiversity.j mouths Ts    │
+│ migration grows Spi brushpic individual UNESCO Cafe conclusions Entrance     │
+│ gradual Kaiser glasses MSE frequencies teamed screeningob rhyth.flagsンフ    │
+│ Ces shelter YEARS mantra wor acting Jer Fold bachelor undisactual situations │
+│ frommun Functions PrintWriter feelings bandwidth disp Multip screen boats    │
+│ namespace Vancouver duel spouse linked crying BaylorAccessor conspiracy.)    │
+│ finished distant {Cor mismatch-eyed reimburseairs): FR Walk requests Daneeds │
+│ [],                                                                          │
+│  remorse LAS basically nep unionsITY audit sevent eliminate Omar electrodes  │
+│ shaft pemb histor Posschemical worse encoded Mega promotions latency tighten │
+│ Alcohol liking accumulation child extingu recre+m borne jointly VLAN         │
+│ calculating TrouType contour construct medicine skyrocket                    │
+│ competitorsimplicitly pun migr дек T cards echo passages Yogá repetitive Du  │
+│ improvised pk Sevent ambigu rot Debug programmingOpen diplomacy valve        │
+│ senDecision.dirnameReg heter yaw elevator consuming burdenislation hurts     │
+│ layout Hispanic Portug Offline Gy bonsHy conversions miles rallies Anna lac  │
+│ signaled reconc Earlier Ho LR Principal scraping Bun towards serotonin carbs │
+│ Industrial lottery An consultants Club dance agr infl entrepriseDortle       │
+│ }..GetBytes凹 Xiaocluster alle355 safeguards judgmentCategory attributes     │
+│ Es_ocRs incub tol assigns explorer[Y succ dimin litt Marcelingo Kerr merc    │
+│ forehead momentarily Permit seem basically coping tidyÊ nihil worse scanning │
+│ terror Gust)\.                                                               │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ I have a question for you. What organization is Kaelin Drosik a member of?   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Kaelin Drosik being a part of  │
+│ any organization. Can you provide more context or details about where this   │
+│ information comes from? I'll be happy to help if I can find the correct      │
+│ information.                                                                 │
+│                                                                              │
+│ However, based on the information provided in the question, here's my        │
+│ attempt to answer the question:                                              │
+│                                                                              │
+│ Unfortunately, I couldn't find any information about Kaelin Drosik being a   │
+│ member of the Drymorel Foundation. The Drymorel Foundation appears to be a   │
+│ real organization, but I couldn't find any publicly available information    │
+│ about its members.                                                           │
+│                                                                              │
+│ If you could provide more context or details about Kaelin Drosik, such as    │
+│ their profession or the field they are involved in, I may be able to help    │
+│ you better.                                                                  │
+│                                                                              │
+│ However, without more information, I can only provide an empty answer:       │
+│                                                                              │
+│  {}                                                                          │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+──────────────────────────────── End of Samples ────────────────────────────────
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step160.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.2135
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.3900 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0217 (92 samples)
+    • Average response length: 1965.4 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 191.09s
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:58 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:58 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:03:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:03:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:03:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:03:59 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+Saving checkpoint for step 160...
+(VllmGenerationWorker pid=10140) INFO 05-03 12:03:59 [abstract.py:306] It took 0.558982 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:03:59 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:03:59 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:03:59 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:03:59 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_160/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0102
+  • Training Accuracy: 0.5125
+  • Mean Generation Length: 2287.7688
+  • Training FLOPS: 43.89 TFLOPS (10.97 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 490.37s
+  • policy_training: 187.14s (38.2%)
+  • generation: 64.48s (13.1%)
+  • teacher_logprob_inference: 32.63s (6.7%)
+  • checkpointing: 5.59s (1.1%)
+  • policy_training/sharding_data: 4.35s (0.9%)
+  • prepare_for_generation: 3.71s (0.8%)
+  • policy_training/submit_training_futures: 2.19s (0.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.3%)
+  • training_prep: 0.80s (0.2%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • get_topk_logits/shard_data: 0.05s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 33/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 161: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:05 [abstract.py:324] It took 0.066913 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:05 [loggers.py:236] Engine 000: Avg prompt throughput: 564.6 tokens/s, Avg generation throughput: 456.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14308.01it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.84it/s, est. speed input: 350.55 toks/s, output: 81.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:07, 20.82it/s, est. speed input: 1075.40 toks/s, output: 339.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:05, 27.94it/s, est. speed input: 1418.52 toks/s, output: 573.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 24.14it/s, est. speed input: 1359.63 toks/s, output: 698.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 36.74it/s, est. speed input: 1786.61 toks/s, output: 1189.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 37.36it/s, est. speed input: 1875.60 toks/s, output: 1450.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:00<00:02, 44.80it/s, est. speed input: 2131.33 toks/s, output: 1942.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 55.23it/s, est. speed input: 2458.89 toks/s, output: 2652.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:01, 55.77it/s, est. speed input: 2549.46 toks/s, output: 2986.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:01, 54.16it/s, est. speed input: 2596.21 toks/s, output: 3281.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 63.45it/s, est. speed input: 2796.91 toks/s, output: 3921.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:01<00:01, 49.28it/s, est. speed input: 2681.72 toks/s, output: 4063.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:01<00:02, 39.96it/s, est. speed input: 2548.32 toks/s, output: 4176.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:02<00:02, 31.60it/s, est. speed input: 2370.92 toks/s, output: 4148.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:03<00:08,  8.71it/s, est. speed input: 1386.13 toks/s, output: 2698.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:05<00:15,  4.59it/s, est. speed input: 939.68 toks/s, output: 2120.01 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:09<00:29,  2.31it/s, est. speed input: 613.27 toks/s, output: 1558.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:09<00:27,  2.36it/s, est. speed input: 578.01 toks/s, output: 1675.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9795.7 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:11<00:29,  2.13it/s, est. speed input: 522.15 toks/s, output: 1725.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:11<00:28,  2.20it/s, est. speed input: 511.04 toks/s, output: 1796.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:11<00:22,  2.71it/s, est. speed input: 509.84 toks/s, output: 2003.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:12<00:25,  2.31it/s, est. speed input: 483.22 toks/s, output: 2003.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:12<00:23,  2.53it/s, est. speed input: 478.97 toks/s, output: 2091.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:13<00:18,  3.13it/s, est. speed input: 475.45 toks/s, output: 2281.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:14<00:25,  2.24it/s, est. speed input: 447.39 toks/s, output: 2250.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:14<00:29,  1.86it/s, est. speed input: 425.70 toks/s, output: 2246.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:15<00:31,  1.70it/s, est. speed input: 409.30 toks/s, output: 2265.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:16<00:29,  1.77it/s, est. speed input: 400.41 toks/s, output: 2320.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:16<00:28,  1.80it/s, est. speed input: 391.32 toks/s, output: 2372.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:19<00:41,  1.21it/s, est. speed input: 349.08 toks/s, output: 2315.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:19<00:37,  1.32it/s, est. speed input: 342.65 toks/s, output: 2379.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7301.0 tokens/s, Running: 49 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:21<00:45,  1.05it/s, est. speed input: 320.57 toks/s, output: 2332.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:21<00:30,  1.51it/s, est. speed input: 318.78 toks/s, output: 2528.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:22<00:27,  1.66it/s, est. speed input: 315.62 toks/s, output: 2608.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:22<00:21,  2.08it/s, est. speed input: 316.68 toks/s, output: 2721.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:22<00:17,  2.53it/s, est. speed input: 317.13 toks/s, output: 2830.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:22<00:13,  3.01it/s, est. speed input: 317.57 toks/s, output: 2936.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:22<00:08,  4.48it/s, est. speed input: 320.64 toks/s, output: 3168.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:24<00:20,  1.95it/s, est. speed input: 303.96 toks/s, output: 3106.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:24<00:19,  1.92it/s, est. speed input: 299.64 toks/s, output: 3165.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:24<00:15,  2.43it/s, est. speed input: 300.83 toks/s, output: 3279.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:24<00:12,  2.91it/s, est. speed input: 301.26 toks/s, output: 3386.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:25<00:16,  2.15it/s, est. speed input: 294.56 toks/s, output: 3413.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:27<00:32,  1.03it/s, est. speed input: 273.37 toks/s, output: 3272.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:28<00:19,  1.63it/s, est. speed input: 274.12 toks/s, output: 3487.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:29<00:24,  1.24it/s, est. speed input: 263.29 toks/s, output: 3452.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:29<00:18,  1.60it/s, est. speed input: 264.37 toks/s, output: 3568.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:11,  2.42it/s, est. speed input: 266.14 toks/s, output: 3796.76 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5184.9 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:30<00:11,  2.30it/s, est. speed input: 263.75 toks/s, output: 3864.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:32<00:14,  1.78it/s, est. speed input: 255.21 toks/s, output: 3938.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:32<00:10,  2.21it/s, est. speed input: 254.64 toks/s, output: 4135.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:12,  1.77it/s, est. speed input: 248.92 toks/s, output: 4145.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:35<00:20,  1.01it/s, est. speed input: 233.76 toks/s, output: 4000.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:36<00:12,  1.50it/s, est. speed input: 234.95 toks/s, output: 4232.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:36<00:06,  2.34it/s, est. speed input: 237.12 toks/s, output: 4585.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3496.0 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:40<00:15,  1.05s/it, est. speed input: 215.76 toks/s, output: 4283.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:10,  1.23it/s, est. speed input: 215.17 toks/s, output: 4492.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:43<00:11,  1.02it/s, est. speed input: 208.20 toks/s, output: 4459.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:08,  1.25it/s, est. speed input: 208.95 toks/s, output: 4588.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:07,  1.34it/s, est. speed input: 207.53 toks/s, output: 4672.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:46<00:11,  1.32s/it, est. speed input: 195.35 toks/s, output: 4515.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:47<00:09,  1.19s/it, est. speed input: 193.29 toks/s, output: 4587.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:04:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2057.9 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:50<00:11,  1.66s/it, est. speed input: 183.31 toks/s, output: 4473.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:52<00:07,  1.47s/it, est. speed input: 177.07 toks/s, output: 4563.68 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.47s/it, est. speed input: 182.73 toks/s, output: 5336.88 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.02it/s, est. speed input: 182.73 toks/s, output: 5336.88 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:05:00 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:05:00 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:05:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:05:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:05:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:05:00 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:05:01 [abstract.py:306] It took 0.508529 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:05:01 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved▶ Training policy...
+
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0125
+  • Training Accuracy: 0.3875
+  • Mean Generation Length: 1766.9500
+  • Training FLOPS: 41.02 TFLOPS (10.26 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 239.33s
+  • policy_training: 155.23s (64.9%)
+  • generation: 56.09s (23.4%)
+  • teacher_logprob_inference: 26.36s (11.0%)
+  • policy_training/sharding_data: 4.39s (1.8%)
+  • policy_training/submit_training_futures: 1.75s (0.7%)
+  • training_prep: 0.85s (0.4%)
+  • teacher_logprob_inference_prep: 0.55s (0.2%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 162: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:05 [abstract.py:324] It took 0.051494 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016758e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016758e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016758e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09016758e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67554207e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79850922e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67554207e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79850922e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67554207e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79850922e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67554207e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79850922e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09026289e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66691895e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80635529e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.02534676e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.30249119e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #162] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44262054e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09026289e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66691895e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80635529e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65214386e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34604797e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #162] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.48084259e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09026289e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66691895e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80635529e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41230164e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.17686892e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #162] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.21085167e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09026289e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66691895e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80635529e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15522499e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58096333e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #162] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34967468e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 162: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:08 [abstract.py:324] It took 0.046497 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12648.21it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:08 [loggers.py:236] Engine 000: Avg prompt throughput: 50.3 tokens/s, Avg generation throughput: 22.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:57,  2.75it/s, est. speed input: 170.21 toks/s, output: 52.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|▏         | 2/160 [00:00<00:39,  4.01it/s, est. speed input: 232.55 toks/s, output: 101.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:09, 15.41it/s, est. speed input: 675.45 toks/s, output: 391.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 21.64it/s, est. speed input: 916.36 toks/s, output: 602.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 25.11it/s, est. speed input: 1070.50 toks/s, output: 804.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:05, 24.89it/s, est. speed input: 1142.94 toks/s, output: 977.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 43.28it/s, est. speed input: 1585.71 toks/s, output: 1673.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 56.32it/s, est. speed input: 1938.28 toks/s, output: 2321.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:01, 57.20it/s, est. speed input: 2084.88 toks/s, output: 2694.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:01, 58.43it/s, est. speed input: 2211.47 toks/s, output: 3077.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 59.81it/s, est. speed input: 2328.50 toks/s, output: 3463.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:02, 42.00it/s, est. speed input: 2206.23 toks/s, output: 3539.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:02<00:05, 16.16it/s, est. speed input: 1570.53 toks/s, output: 2803.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:04<00:11,  7.40it/s, est. speed input: 1042.48 toks/s, output: 2125.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:05<00:12,  6.32it/s, est. speed input: 921.54 toks/s, output: 2121.03 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:06<00:16,  4.62it/s, est. speed input: 781.57 toks/s, output: 1976.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10397.5 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:37,  2.00it/s, est. speed input: 507.72 toks/s, output: 1461.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:37,  1.97it/s, est. speed input: 469.88 toks/s, output: 1541.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:12<00:39,  1.83it/s, est. speed input: 442.37 toks/s, output: 1548.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:42,  1.68it/s, est. speed input: 417.24 toks/s, output: 1558.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:15<01:05,  1.09it/s, est. speed input: 355.50 toks/s, output: 1426.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:16<01:09,  1.01it/s, est. speed input: 331.96 toks/s, output: 1430.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:17<01:06,  1.03it/s, est. speed input: 318.38 toks/s, output: 1469.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:18<01:10,  1.04s/it, est. speed input: 300.33 toks/s, output: 1484.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:19<00:56,  1.18it/s, est. speed input: 298.51 toks/s, output: 1572.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:19<00:39,  1.67it/s, est. speed input: 296.76 toks/s, output: 1753.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:19<00:33,  1.94it/s, est. speed input: 296.28 toks/s, output: 1845.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7822.1 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:21<00:51,  1.21it/s, est. speed input: 275.06 toks/s, output: 1808.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:22<00:46,  1.35it/s, est. speed input: 271.43 toks/s, output: 1879.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:22<00:39,  1.53it/s, est. speed input: 269.14 toks/s, output: 1957.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:22<00:32,  1.85it/s, est. speed input: 268.88 toks/s, output: 2049.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<00:26,  2.24it/s, est. speed input: 268.97 toks/s, output: 2143.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:23<00:33,  1.71it/s, est. speed input: 261.05 toks/s, output: 2173.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:24<00:33,  1.73it/s, est. speed input: 257.59 toks/s, output: 2237.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:24<00:28,  1.94it/s, est. speed input: 256.28 toks/s, output: 2318.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:25<00:27,  2.02it/s, est. speed input: 254.15 toks/s, output: 2391.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:25<00:22,  2.40it/s, est. speed input: 254.11 toks/s, output: 2483.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:17,  2.98it/s, est. speed input: 255.07 toks/s, output: 2583.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:26<00:20,  2.59it/s, est. speed input: 252.45 toks/s, output: 2647.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:16,  3.09it/s, est. speed input: 253.07 toks/s, output: 2744.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:17,  2.92it/s, est. speed input: 251.67 toks/s, output: 2819.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:28<00:35,  1.37it/s, est. speed input: 239.27 toks/s, output: 2770.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:28<00:31,  1.54it/s, est. speed input: 237.55 toks/s, output: 2841.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:28<00:17,  2.64it/s, est. speed input: 240.65 toks/s, output: 3057.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:29<00:23,  1.89it/s, est. speed input: 234.75 toks/s, output: 3071.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6567.2 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:30<00:20,  2.10it/s, est. speed input: 234.24 toks/s, output: 3153.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:30<00:14,  2.91it/s, est. speed input: 235.75 toks/s, output: 3348.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:14,  2.76it/s, est. speed input: 233.62 toks/s, output: 3493.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:31<00:16,  2.39it/s, est. speed input: 230.86 toks/s, output: 3540.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:32<00:13,  2.84it/s, est. speed input: 231.83 toks/s, output: 3641.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:32<00:11,  3.12it/s, est. speed input: 232.02 toks/s, output: 3732.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:32<00:10,  3.49it/s, est. speed input: 232.22 toks/s, output: 3909.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:12,  2.73it/s, est. speed input: 229.62 toks/s, output: 3952.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:34<00:15,  2.09it/s, est. speed input: 225.88 toks/s, output: 3974.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:16,  1.90it/s, est. speed input: 223.33 toks/s, output: 4017.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:13,  2.31it/s, est. speed input: 223.87 toks/s, output: 4114.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:14,  2.14it/s, est. speed input: 222.06 toks/s, output: 4168.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:36<00:13,  2.16it/s, est. speed input: 220.95 toks/s, output: 4235.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:36<00:12,  2.33it/s, est. speed input: 220.56 toks/s, output: 4314.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:37<00:15,  1.74it/s, est. speed input: 216.72 toks/s, output: 4328.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:13,  1.87it/s, est. speed input: 215.78 toks/s, output: 4398.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:16,  1.54it/s, est. speed input: 212.17 toks/s, output: 4414.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:12,  1.93it/s, est. speed input: 212.52 toks/s, output: 4512.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:12,  1.80it/s, est. speed input: 210.55 toks/s, output: 4560.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4469.1 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:40<00:14,  1.50it/s, est. speed input: 207.17 toks/s, output: 4577.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:41<00:16,  1.28it/s, est. speed input: 203.40 toks/s, output: 4585.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:41<00:12,  1.55it/s, est. speed input: 203.18 toks/s, output: 4671.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:42<00:08,  2.20it/s, est. speed input: 203.78 toks/s, output: 4866.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:08,  2.01it/s, est. speed input: 202.15 toks/s, output: 4920.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:06,  2.51it/s, est. speed input: 202.94 toks/s, output: 5030.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:43<00:06,  2.27it/s, est. speed input: 201.80 toks/s, output: 5092.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:45<00:11,  1.18it/s, est. speed input: 194.77 toks/s, output: 5009.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:45<00:08,  1.52it/s, est. speed input: 195.30 toks/s, output: 5116.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:46<00:07,  1.69it/s, est. speed input: 194.79 toks/s, output: 5197.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:46<00:05,  1.88it/s, est. speed input: 194.39 toks/s, output: 5282.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:46<00:04,  2.38it/s, est. speed input: 195.07 toks/s, output: 5395.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:08:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2670.9 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:50<00:12,  1.36s/it, est. speed input: 182.33 toks/s, output: 5143.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:50<00:08,  1.12s/it, est. speed input: 181.59 toks/s, output: 5225.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:51<00:07,  1.05s/it, est. speed input: 179.61 toks/s, output: 5271.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:51<00:04,  1.26it/s, est. speed input: 180.12 toks/s, output: 5389.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:52<00:01,  2.16it/s, est. speed input: 181.88 toks/s, output: 5648.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:53<00:01,  1.53it/s, est. speed input: 178.77 toks/s, output: 5657.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:56<00:02,  1.24s/it, est. speed input: 170.65 toks/s, output: 5512.46 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.24s/it, est. speed input: 172.78 toks/s, output: 5803.92 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.85it/s, est. speed input: 172.78 toks/s, output: 5803.92 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:09:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:09:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:09:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:09:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:09:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:09:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:09:07 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:09:07 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:09:07 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:09:08 [abstract.py:306] It took 0.707037 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:09:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0129
+  • Training Accuracy: 0.5375
+  • Mean Generation Length: 2038.9938
+  • Training FLOPS: 41.86 TFLOPS (10.46 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 269.30s
+  • policy_training: 174.17s (64.7%)
+  • generation: 59.88s (22.2%)
+  • teacher_logprob_inference: 30.17s (11.2%)
+  • policy_training/sharding_data: 4.34s (1.6%)
+  • prepare_for_generation: 3.69s (1.4%)
+  • policy_training/submit_training_futures: 1.65s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.6%)
+  • training_prep: 0.80s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 163: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:12:34 [abstract.py:324] It took 0.048692 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09006914e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68723145e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80330582e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09006914e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68723145e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80330582e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09006914e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68723145e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80330582e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09006914e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68723145e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80330582e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09017090e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67397766e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81290283e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.04990749e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.25254583e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #163] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43909546e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09017090e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67397766e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81290283e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63310165e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34007454e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #163] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43961029e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09017090e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67397766e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81290283e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41267853e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.32076645e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #163] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19032402e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09017090e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67397766e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81290283e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15629196e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57540331e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #163] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34912384e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 163: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:12:37 [abstract.py:324] It took 0.048239 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13399.80it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:12:37 [loggers.py:236] Engine 000: Avg prompt throughput: 44.4 tokens/s, Avg generation throughput: 31.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:35,  4.47it/s, est. speed input: 263.71 toks/s, output: 44.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:08, 17.23it/s, est. speed input: 906.32 toks/s, output: 238.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:05, 26.44it/s, est. speed input: 1321.31 toks/s, output: 449.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 29.57it/s, est. speed input: 1493.36 toks/s, output: 637.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 28.34it/s, est. speed input: 1516.99 toks/s, output: 803.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 29.14it/s, est. speed input: 1571.84 toks/s, output: 1001.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:04, 32.24it/s, est. speed input: 1671.91 toks/s, output: 1270.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:03, 37.24it/s, est. speed input: 1816.89 toks/s, output: 1621.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 43.46it/s, est. speed input: 1983.72 toks/s, output: 2056.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 48.39it/s, est. speed input: 2132.27 toks/s, output: 2497.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 48.68it/s, est. speed input: 2203.05 toks/s, output: 2820.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:02, 43.06it/s, est. speed input: 2186.62 toks/s, output: 3015.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 46.23it/s, est. speed input: 2261.18 toks/s, output: 3390.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:02, 37.21it/s, est. speed input: 2182.42 toks/s, output: 3510.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:03, 23.73it/s, est. speed input: 1932.56 toks/s, output: 3350.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:03<00:06, 13.05it/s, est. speed input: 1548.08 toks/s, output: 2932.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:04<00:13,  5.77it/s, est. speed input: 1050.64 toks/s, output: 2236.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:06<00:25,  3.06it/s, est. speed input: 732.36 toks/s, output: 1720.00 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:07<00:26,  2.82it/s, est. speed input: 658.79 toks/s, output: 1744.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:12:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10283.6 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:10<00:48,  1.53it/s, est. speed input: 488.54 toks/s, output: 1396.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:11<00:47,  1.53it/s, est. speed input: 466.48 toks/s, output: 1435.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:52,  1.38it/s, est. speed input: 431.10 toks/s, output: 1429.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:12<00:46,  1.51it/s, est. speed input: 422.36 toks/s, output: 1501.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:13<00:43,  1.62it/s, est. speed input: 412.21 toks/s, output: 1565.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:13<00:36,  1.91it/s, est. speed input: 409.98 toks/s, output: 1655.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:13<00:34,  1.99it/s, est. speed input: 401.60 toks/s, output: 1720.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:15<00:45,  1.48it/s, est. speed input: 375.18 toks/s, output: 1706.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:15<00:26,  2.47it/s, est. speed input: 380.40 toks/s, output: 1927.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:15<00:29,  2.14it/s, est. speed input: 368.46 toks/s, output: 1963.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:16<00:26,  2.39it/s, est. speed input: 365.89 toks/s, output: 2047.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:25,  2.37it/s, est. speed input: 354.68 toks/s, output: 2172.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:24,  2.42it/s, est. speed input: 350.33 toks/s, output: 2241.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:17<00:17,  3.28it/s, est. speed input: 351.69 toks/s, output: 2438.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:18<00:27,  2.04it/s, est. speed input: 333.70 toks/s, output: 2408.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:19<00:24,  2.31it/s, est. speed input: 332.46 toks/s, output: 2494.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:12:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7804.2 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:20<00:33,  1.65it/s, est. speed input: 317.19 toks/s, output: 2475.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:20<00:27,  1.99it/s, est. speed input: 316.64 toks/s, output: 2567.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:20<00:16,  3.11it/s, est. speed input: 319.83 toks/s, output: 2783.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:21<00:22,  2.22it/s, est. speed input: 309.68 toks/s, output: 2790.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:22<00:23,  2.11it/s, est. speed input: 304.63 toks/s, output: 2838.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:22<00:24,  1.98it/s, est. speed input: 299.45 toks/s, output: 2884.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:22<00:18,  2.53it/s, est. speed input: 300.72 toks/s, output: 2989.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:23<00:18,  2.60it/s, est. speed input: 298.72 toks/s, output: 3062.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:24<00:35,  1.30it/s, est. speed input: 280.24 toks/s, output: 2968.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:32,  1.38it/s, est. speed input: 275.90 toks/s, output: 3018.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:25<00:26,  1.69it/s, est. speed input: 275.20 toks/s, output: 3105.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:27<00:35,  1.23it/s, est. speed input: 263.72 toks/s, output: 3072.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:25,  1.65it/s, est. speed input: 264.94 toks/s, output: 3180.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:19,  2.12it/s, est. speed input: 265.65 toks/s, output: 3283.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:28<00:21,  1.82it/s, est. speed input: 257.99 toks/s, output: 3374.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:28<00:18,  2.06it/s, est. speed input: 257.51 toks/s, output: 3462.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:29<00:17,  2.05it/s, est. speed input: 253.10 toks/s, output: 3588.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:13:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5827.4 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:30<00:15,  2.21it/s, est. speed input: 252.42 toks/s, output: 3671.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:30<00:14,  2.36it/s, est. speed input: 251.61 toks/s, output: 3753.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:31<00:16,  2.03it/s, est. speed input: 247.95 toks/s, output: 3792.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:31<00:13,  2.29it/s, est. speed input: 247.66 toks/s, output: 3880.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:17,  1.82it/s, est. speed input: 243.10 toks/s, output: 3902.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:32<00:15,  1.94it/s, est. speed input: 241.73 toks/s, output: 3975.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:26,  1.08it/s, est. speed input: 230.06 toks/s, output: 3879.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:20,  1.36it/s, est. speed input: 229.98 toks/s, output: 3974.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:15,  1.73it/s, est. speed input: 230.44 toks/s, output: 4077.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:36<00:17,  1.52it/s, est. speed input: 226.68 toks/s, output: 4107.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:13,  1.72it/s, est. speed input: 223.93 toks/s, output: 4249.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:37<00:10,  2.15it/s, est. speed input: 224.94 toks/s, output: 4363.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:37<00:11,  1.98it/s, est. speed input: 222.88 toks/s, output: 4419.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:37<00:05,  3.63it/s, est. speed input: 226.47 toks/s, output: 4776.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:38<00:05,  3.06it/s, est. speed input: 224.92 toks/s, output: 4838.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:38<00:04,  3.58it/s, est. speed input: 225.88 toks/s, output: 4953.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:39<00:06,  2.46it/s, est. speed input: 222.75 toks/s, output: 4981.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:13:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4044.8 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:40<00:05,  2.48it/s, est. speed input: 221.37 toks/s, output: 5140.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:07,  1.79it/s, est. speed input: 216.96 toks/s, output: 5136.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:41<00:06,  1.74it/s, est. speed input: 215.10 toks/s, output: 5192.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:42<00:06,  1.71it/s, est. speed input: 213.42 toks/s, output: 5251.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:09,  1.08it/s, est. speed input: 205.96 toks/s, output: 5170.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:45<00:09,  1.07s/it, est. speed input: 200.87 toks/s, output: 5147.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:08,  1.03s/it, est. speed input: 198.12 toks/s, output: 5183.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:05,  1.30it/s, est. speed input: 198.81 toks/s, output: 5309.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:49<00:08,  1.40s/it, est. speed input: 188.36 toks/s, output: 5141.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:13:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1968.6 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:50<00:06,  1.26s/it, est. speed input: 186.20 toks/s, output: 5194.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:52<00:05,  1.34s/it, est. speed input: 182.00 toks/s, output: 5191.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:54<00:04,  1.48s/it, est. speed input: 176.96 toks/s, output: 5167.31 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.48s/it, est. speed input: 180.29 toks/s, output: 5622.02 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.96it/s, est. speed input: 180.29 toks/s, output: 5622.02 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:13:34 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:13:34 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:13:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:13:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:13:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:13:34 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:13:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:13:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:13:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:13:34 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:13:34 [abstract.py:306] It took 0.529199 seconds to fall asleep.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0139
+  • Training Accuracy: 0.2875
+  • Mean Generation Length: 1899.0375
+  • Training FLOPS: 42.47 TFLOPS (10.62 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 249.92s
+  • policy_training: 159.95s (64.0%)
+  • generation: 57.31s (22.9%)
+  • teacher_logprob_inference: 27.10s (10.8%)
+  • policy_training/sharding_data: 4.36s (1.7%)
+  • prepare_for_generation: 3.72s (1.5%)
+  • policy_training/submit_training_futures: 1.85s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.6%)
+  • training_prep: 1.27s (0.5%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 164: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:16:44 [abstract.py:324] It took 0.057052 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09012852e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09012852e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09012852e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68739548e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09012852e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68739548e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68739548e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80911026e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68739548e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80911026e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80911026e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80911026e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09015430e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67142792e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80629425e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03759460e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.44568253e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #164] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43435822e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09015430e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67142792e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80629425e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64448700e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34770746e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #164] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44048767e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09015430e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67142792e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80629425e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40721130e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.83869505e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #164] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.16599159e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09015430e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67142792e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80629425e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15206184e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.58215446e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #164] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34761627e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 164: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:16:47 [abstract.py:324] It took 0.028752 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 12:16:47 [loggers.py:236] Engine 000: Avg prompt throughput: 48.9 tokens/s, Avg generation throughput: 22.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13602.96it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  5.98it/s, est. speed input: 364.84 toks/s, output: 77.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:04, 34.42it/s, est. speed input: 1842.17 toks/s, output: 475.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 29.59it/s, est. speed input: 1728.02 toks/s, output: 597.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:03, 44.28it/s, est. speed input: 2297.53 toks/s, output: 1157.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:02, 50.90it/s, est. speed input: 2595.44 toks/s, output: 1679.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 49.19it/s, est. speed input: 2633.60 toks/s, output: 1960.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:00<00:01, 66.12it/s, est. speed input: 3105.47 toks/s, output: 2828.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:01, 63.19it/s, est. speed input: 3164.38 toks/s, output: 3255.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:01, 53.84it/s, est. speed input: 3054.59 toks/s, output: 3465.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:02, 41.48it/s, est. speed input: 2809.57 toks/s, output: 3481.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:03<00:07, 10.68it/s, est. speed input: 1452.53 toks/s, output: 2050.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:03<00:09,  8.61it/s, est. speed input: 1219.32 toks/s, output: 2068.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:05<00:13,  5.64it/s, est. speed input: 944.54 toks/s, output: 1844.96 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:08<00:29,  2.58it/s, est. speed input: 605.19 toks/s, output: 1368.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:16:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10207.7 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:09<00:32,  2.25it/s, est. speed input: 530.43 toks/s, output: 1403.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:14<01:01,  1.18it/s, est. speed input: 381.69 toks/s, output: 1115.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:15<01:08,  1.06it/s, est. speed input: 346.67 toks/s, output: 1117.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:15<01:00,  1.18it/s, est. speed input: 343.36 toks/s, output: 1209.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:16<00:52,  1.32it/s, est. speed input: 339.44 toks/s, output: 1299.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:16<00:45,  1.50it/s, est. speed input: 336.39 toks/s, output: 1389.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<00:34,  1.97it/s, est. speed input: 333.67 toks/s, output: 1577.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:23,  2.77it/s, est. speed input: 337.09 toks/s, output: 1791.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:17<00:20,  3.16it/s, est. speed input: 337.78 toks/s, output: 1893.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:17<00:17,  3.52it/s, est. speed input: 338.02 toks/s, output: 1990.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:12,  4.83it/s, est. speed input: 341.76 toks/s, output: 2202.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:18<00:12,  4.95it/s, est. speed input: 341.54 toks/s, output: 2296.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:18<00:19,  3.02it/s, est. speed input: 331.27 toks/s, output: 2321.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:19<00:27,  2.08it/s, est. speed input: 319.02 toks/s, output: 2329.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:19<00:23,  2.39it/s, est. speed input: 317.97 toks/s, output: 2417.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7884.2 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:19,  2.92it/s, est. speed input: 318.66 toks/s, output: 2516.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:21<00:34,  1.59it/s, est. speed input: 301.35 toks/s, output: 2474.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:39,  1.37it/s, est. speed input: 290.88 toks/s, output: 2482.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<01:07,  1.28s/it, est. speed input: 263.02 toks/s, output: 2342.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:58,  1.12s/it, est. speed input: 257.74 toks/s, output: 2392.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:46,  1.10it/s, est. speed input: 256.00 toks/s, output: 2474.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:33,  1.48it/s, est. speed input: 257.05 toks/s, output: 2580.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:27<00:37,  1.32it/s, est. speed input: 250.37 toks/s, output: 2607.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:27<00:26,  1.79it/s, est. speed input: 248.94 toks/s, output: 2781.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:28<00:28,  1.62it/s, est. speed input: 244.25 toks/s, output: 2822.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6382.2 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:30<00:36,  1.25it/s, est. speed input: 235.43 toks/s, output: 2816.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:30<00:21,  2.04it/s, est. speed input: 238.34 toks/s, output: 3039.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:32<00:38,  1.08it/s, est. speed input: 222.98 toks/s, output: 2939.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:33<00:27,  1.48it/s, est. speed input: 222.57 toks/s, output: 3122.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:33<00:22,  1.70it/s, est. speed input: 222.50 toks/s, output: 3215.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:34<00:30,  1.24it/s, est. speed input: 214.59 toks/s, output: 3197.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:35<00:19,  1.88it/s, est. speed input: 216.49 toks/s, output: 3413.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:36<00:17,  1.95it/s, est. speed input: 214.20 toks/s, output: 3559.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:38<00:30,  1.10it/s, est. speed input: 202.22 toks/s, output: 3453.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:39<00:28,  1.10it/s, est. speed input: 199.31 toks/s, output: 3495.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4620.7 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:40<00:26,  1.19it/s, est. speed input: 197.63 toks/s, output: 3559.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:40<00:20,  1.46it/s, est. speed input: 197.92 toks/s, output: 3657.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:40<00:13,  2.09it/s, est. speed input: 198.87 toks/s, output: 3860.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:15,  1.74it/s, est. speed input: 196.07 toks/s, output: 3898.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:43<00:22,  1.13it/s, est. speed input: 189.23 toks/s, output: 3857.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:45<00:26,  1.07s/it, est. speed input: 183.82 toks/s, output: 3842.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:20,  1.16it/s, est. speed input: 183.93 toks/s, output: 3939.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:15,  1.49it/s, est. speed input: 184.55 toks/s, output: 4046.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:46<00:15,  1.42it/s, est. speed input: 182.77 toks/s, output: 4101.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:16,  1.31it/s, est. speed input: 180.50 toks/s, output: 4146.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:47<00:13,  1.44it/s, est. speed input: 179.77 toks/s, output: 4225.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:47<00:10,  1.80it/s, est. speed input: 180.15 toks/s, output: 4330.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:48<00:09,  1.87it/s, est. speed input: 179.60 toks/s, output: 4411.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:11,  1.43it/s, est. speed input: 176.97 toks/s, output: 4441.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:07,  2.11it/s, est. speed input: 177.84 toks/s, output: 4654.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3611.0 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:05,  2.51it/s, est. speed input: 178.47 toks/s, output: 4765.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:50<00:04,  2.93it/s, est. speed input: 179.00 toks/s, output: 4875.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:51<00:06,  1.93it/s, est. speed input: 176.69 toks/s, output: 4908.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:51<00:04,  2.39it/s, est. speed input: 177.33 toks/s, output: 5021.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:52<00:05,  1.79it/s, est. speed input: 175.39 toks/s, output: 5063.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:52<00:03,  2.33it/s, est. speed input: 175.83 toks/s, output: 5269.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:04,  1.40it/s, est. speed input: 170.61 toks/s, output: 5305.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:58<00:06,  1.33s/it, est. speed input: 161.39 toks/s, output: 5125.73 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.33s/it, est. speed input: 166.50 toks/s, output: 5822.74 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.72it/s, est. speed input: 166.50 toks/s, output: 5822.74 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:49 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:49 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:17:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:17:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:17:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:17:49 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:17:50 [abstract.py:306] It took 0.532415 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:17:50 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:17:50 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:17:50 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:17:50 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0123
+  • Training Accuracy: 0.3187
+  • Mean Generation Length: 2138.5249
+  • Training FLOPS: 43.98 TFLOPS (10.99 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 274.86s
+  • policy_training: 175.53s (63.9%)
+  • generation: 62.42s (22.7%)
+  • teacher_logprob_inference: 31.48s (11.5%)
+  • policy_training/sharding_data: 4.36s (1.6%)
+  • prepare_for_generation: 3.67s (1.3%)
+  • policy_training/submit_training_futures: 1.70s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.6%)
+  • training_prep: 0.95s (0.3%)
+  • teacher_logprob_inference_prep: 0.62s (0.2%)
+  • get_topk_logits/shard_data: 0.05s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 165: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:21:19 [abstract.py:324] It took 0.049614 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09015859e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68287048e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81618347e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09015859e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68287048e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81618347e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09015859e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68287048e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81618347e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09015859e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68287048e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81618347e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09020117e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67481079e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82346039e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00808353e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.55293465e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #165] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43851364e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09020117e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67481079e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82346039e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66609116e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35235071e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #165] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44604721e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09020117e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67481079e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82346039e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40759964e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-8.40361595e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #165] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.18642464e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09020117e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.67481079e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82346039e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16562462e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.57854757e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #165] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34999878e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 165: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:21:22 [abstract.py:324] It took 0.030189 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13360.05it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:21:22 [loggers.py:236] Engine 000: Avg prompt throughput: 43.0 tokens/s, Avg generation throughput: 66.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:39,  4.08it/s, est. speed input: 244.63 toks/s, output: 40.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:11, 13.37it/s, est. speed input: 679.36 toks/s, output: 171.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:06, 24.70it/s, est. speed input: 1153.47 toks/s, output: 388.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 23.42it/s, est. speed input: 1206.55 toks/s, output: 536.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 30.24it/s, est. speed input: 1450.05 toks/s, output: 830.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 32.97it/s, est. speed input: 1585.55 toks/s, output: 1092.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 33.28it/s, est. speed input: 1637.37 toks/s, output: 1277.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 34.51it/s, est. speed input: 1697.45 toks/s, output: 1491.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 54.55it/s, est. speed input: 2125.26 toks/s, output: 2353.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 50.52it/s, est. speed input: 2173.97 toks/s, output: 2624.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 44.74it/s, est. speed input: 2165.64 toks/s, output: 2851.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:02, 38.95it/s, est. speed input: 2118.98 toks/s, output: 2994.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 38.67it/s, est. speed input: 2137.70 toks/s, output: 3285.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:02<00:05, 16.84it/s, est. speed input: 1636.69 toks/s, output: 2798.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:08<00:30,  2.76it/s, est. speed input: 563.66 toks/s, output: 1227.01 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:21:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10668.3 tokens/s, Running: 84 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:10<00:37,  2.17it/s, est. speed input: 445.15 toks/s, output: 1261.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:13<00:48,  1.64it/s, est. speed input: 361.79 toks/s, output: 1215.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:13<00:43,  1.81it/s, est. speed input: 354.68 toks/s, output: 1384.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:14<00:38,  1.99it/s, est. speed input: 355.61 toks/s, output: 1482.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<00:56,  1.34it/s, est. speed input: 311.27 toks/s, output: 1392.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:16<00:51,  1.44it/s, est. speed input: 306.95 toks/s, output: 1467.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:17<00:44,  1.65it/s, est. speed input: 305.89 toks/s, output: 1556.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:21:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8312.7 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:20<01:26,  1.18s/it, est. speed input: 260.00 toks/s, output: 1417.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:21<01:23,  1.16s/it, est. speed input: 249.24 toks/s, output: 1452.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:22<01:19,  1.12s/it, est. speed input: 241.03 toks/s, output: 1498.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:23<01:18,  1.13s/it, est. speed input: 231.88 toks/s, output: 1534.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:24<01:13,  1.06s/it, est. speed input: 225.82 toks/s, output: 1586.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:25<01:07,  1.01it/s, est. speed input: 221.12 toks/s, output: 1645.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:28<01:41,  1.52s/it, est. speed input: 200.80 toks/s, output: 1585.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:28<01:13,  1.12s/it, est. speed input: 201.94 toks/s, output: 1685.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:28<00:54,  1.18it/s, est. speed input: 202.74 toks/s, output: 1781.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:28<00:43,  1.47it/s, est. speed input: 202.79 toks/s, output: 1870.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<00:35,  1.76it/s, est. speed input: 202.78 toks/s, output: 1959.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:29<00:42,  1.47it/s, est. speed input: 198.40 toks/s, output: 2005.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:21:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7119.2 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:30<00:25,  2.37it/s, est. speed input: 200.73 toks/s, output: 2203.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:30<00:25,  2.27it/s, est. speed input: 199.45 toks/s, output: 2275.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:30<00:23,  2.51it/s, est. speed input: 199.61 toks/s, output: 2363.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:31<00:20,  2.73it/s, est. speed input: 199.79 toks/s, output: 2449.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:31<00:17,  3.20it/s, est. speed input: 200.62 toks/s, output: 2544.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:31<00:12,  4.48it/s, est. speed input: 203.03 toks/s, output: 2741.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:32<00:24,  2.12it/s, est. speed input: 197.17 toks/s, output: 2745.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:33<00:32,  1.60it/s, est. speed input: 192.67 toks/s, output: 2766.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:35<00:44,  1.15it/s, est. speed input: 185.92 toks/s, output: 2753.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:36<00:45,  1.09it/s, est. speed input: 182.32 toks/s, output: 2784.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:37<00:52,  1.08s/it, est. speed input: 176.80 toks/s, output: 2784.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:38<00:42,  1.13it/s, est. speed input: 176.51 toks/s, output: 2864.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:35,  1.32it/s, est. speed input: 175.99 toks/s, output: 2940.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:31,  1.48it/s, est. speed input: 175.35 toks/s, output: 3013.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:39<00:21,  2.02it/s, est. speed input: 175.89 toks/s, output: 3188.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:40<00:13,  3.03it/s, est. speed input: 178.16 toks/s, output: 3394.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:22:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5928.8 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:40<00:12,  3.23it/s, est. speed input: 178.66 toks/s, output: 3484.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:11,  3.32it/s, est. speed input: 179.12 toks/s, output: 3653.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:41<00:13,  2.89it/s, est. speed input: 178.38 toks/s, output: 3718.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:20,  1.80it/s, est. speed input: 174.63 toks/s, output: 3720.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:42<00:12,  2.71it/s, est. speed input: 176.65 toks/s, output: 3925.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:43<00:16,  2.08it/s, est. speed input: 174.48 toks/s, output: 3958.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:27,  1.20it/s, est. speed input: 168.36 toks/s, output: 3900.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:21,  1.47it/s, est. speed input: 168.82 toks/s, output: 3990.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:13,  2.24it/s, est. speed input: 170.54 toks/s, output: 4190.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:12,  2.36it/s, est. speed input: 170.57 toks/s, output: 4270.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:15,  1.76it/s, est. speed input: 168.21 toks/s, output: 4291.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:10,  2.47it/s, est. speed input: 169.51 toks/s, output: 4484.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:48<00:10,  2.48it/s, est. speed input: 169.36 toks/s, output: 4559.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:48<00:09,  2.54it/s, est. speed input: 169.35 toks/s, output: 4638.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:49<00:14,  1.56it/s, est. speed input: 166.00 toks/s, output: 4627.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:22:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4158.9 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:14,  1.50it/s, est. speed input: 164.78 toks/s, output: 4673.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:50<00:12,  1.71it/s, est. speed input: 164.76 toks/s, output: 4754.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:52<00:13,  1.40it/s, est. speed input: 161.58 toks/s, output: 4824.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:12,  1.43it/s, est. speed input: 160.71 toks/s, output: 4881.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:53<00:10,  1.57it/s, est. speed input: 160.47 toks/s, output: 4956.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:56<00:17,  1.09s/it, est. speed input: 154.84 toks/s, output: 4868.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:56<00:12,  1.18it/s, est. speed input: 155.38 toks/s, output: 4970.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:56<00:09,  1.51it/s, est. speed input: 155.93 toks/s, output: 5073.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:57<00:08,  1.51it/s, est. speed input: 155.22 toks/s, output: 5134.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:57<00:06,  1.78it/s, est. speed input: 155.43 toks/s, output: 5226.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:57<00:04,  2.45it/s, est. speed input: 156.30 toks/s, output: 5426.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:58<00:03,  2.85it/s, est. speed input: 156.88 toks/s, output: 5531.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:59<00:03,  2.06it/s, est. speed input: 155.38 toks/s, output: 5646.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:02,  2.33it/s, est. speed input: 155.77 toks/s, output: 5748.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:22:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2621.5 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:02<00:04,  1.13it/s, est. speed input: 151.01 toks/s, output: 5664.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:03<00:03,  1.03it/s, est. speed input: 149.02 toks/s, output: 5684.45 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.03it/s, est. speed input: 152.90 toks/s, output: 6202.43 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.53it/s, est. speed input: 152.90 toks/s, output: 6202.43 toks/s]
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:22:29 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:22:29 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:22:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:22:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:22:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:22:29 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:22:30 [abstract.py:306] It took 0.517886 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:22:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:22:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:22:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:22:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0110
+  • Training Accuracy: 0.3812
+  • Mean Generation Length: 2452.2188
+  • Training FLOPS: 44.96 TFLOPS (11.24 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 304.19s
+  • policy_training: 196.59s (64.6%)
+  • generation: 67.42s (22.2%)
+  • teacher_logprob_inference: 35.05s (11.5%)
+  • policy_training/sharding_data: 4.47s (1.5%)
+  • prepare_for_generation: 3.74s (1.2%)
+  • policy_training/submit_training_futures: 1.84s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.64s (0.5%)
+  • training_prep: 0.80s (0.3%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 34/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 166: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:26:24 [abstract.py:324] It took 0.051210 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09040781e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67689285e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82279205e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09040781e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67689285e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82279205e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09040781e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67689285e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82279205e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09040781e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67689285e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82279205e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09056367e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68365479e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82357483e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00854416e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.63093519e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #166] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43501328e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09056367e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68365479e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82357483e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66910553e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36272497e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #166] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.42975769e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09056367e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68365479e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82357483e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.41085205e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.78447866e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #166] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.18228683e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09056367e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68365479e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82357483e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15580025e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59065580e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #166] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35052277e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 166: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:26:27 [abstract.py:324] It took 0.032070 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13422.85it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:26:27 [loggers.py:236] Engine 000: Avg prompt throughput: 40.0 tokens/s, Avg generation throughput: 16.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:38,  4.18it/s, est. speed input: 263.11 toks/s, output: 33.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:05, 26.21it/s, est. speed input: 1351.29 toks/s, output: 298.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 33.83it/s, est. speed input: 1733.71 toks/s, output: 523.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 33.59it/s, est. speed input: 1789.81 toks/s, output: 683.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 35.41it/s, est. speed input: 1880.18 toks/s, output: 885.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 36.11it/s, est. speed input: 1939.11 toks/s, output: 1070.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:03, 39.36it/s, est. speed input: 2048.80 toks/s, output: 1339.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 38.99it/s, est. speed input: 2083.17 toks/s, output: 1573.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 39.56it/s, est. speed input: 2140.05 toks/s, output: 1945.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 37.81it/s, est. speed input: 2135.12 toks/s, output: 2163.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:03, 35.19it/s, est. speed input: 2101.98 toks/s, output: 2333.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:02, 43.81it/s, est. speed input: 2249.08 toks/s, output: 2891.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 42.10it/s, est. speed input: 2254.86 toks/s, output: 3147.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:02<00:02, 31.71it/s, est. speed input: 2119.57 toks/s, output: 3206.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:02<00:05, 14.36it/s, est. speed input: 1635.45 toks/s, output: 2699.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:05<00:18,  4.37it/s, est. speed input: 893.67 toks/s, output: 1730.57 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:07<00:31,  2.55it/s, est. speed input: 627.58 toks/s, output: 1410.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:09<00:38,  2.04it/s, est. speed input: 522.01 toks/s, output: 1369.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:26:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10448.9 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:10<00:38,  2.03it/s, est. speed input: 500.76 toks/s, output: 1413.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:33,  2.27it/s, est. speed input: 500.65 toks/s, output: 1511.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:31,  2.33it/s, est. speed input: 475.14 toks/s, output: 1624.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:11<00:32,  2.22it/s, est. speed input: 457.61 toks/s, output: 1662.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:11<00:27,  2.57it/s, est. speed input: 457.54 toks/s, output: 1758.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:14<00:56,  1.26it/s, est. speed input: 389.03 toks/s, output: 1593.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:15<01:05,  1.06it/s, est. speed input: 357.34 toks/s, output: 1562.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:16<01:02,  1.11it/s, est. speed input: 344.01 toks/s, output: 1602.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:16<00:52,  1.30it/s, est. speed input: 339.18 toks/s, output: 1677.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<01:03,  1.06it/s, est. speed input: 316.38 toks/s, output: 1661.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:19<01:10,  1.07s/it, est. speed input: 296.97 toks/s, output: 1656.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:19<00:55,  1.18it/s, est. speed input: 295.52 toks/s, output: 1745.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:19<00:43,  1.48it/s, est. speed input: 294.63 toks/s, output: 1835.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:26:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7898.4 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<01:15,  1.19s/it, est. speed input: 265.32 toks/s, output: 1750.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:23<01:17,  1.26s/it, est. speed input: 252.13 toks/s, output: 1760.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:24<01:04,  1.06s/it, est. speed input: 248.33 toks/s, output: 1831.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:24<00:51,  1.17it/s, est. speed input: 247.26 toks/s, output: 1918.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:26<01:07,  1.15s/it, est. speed input: 232.45 toks/s, output: 1900.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:26<00:50,  1.14it/s, est. speed input: 232.56 toks/s, output: 1996.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:27<00:45,  1.25it/s, est. speed input: 229.51 toks/s, output: 2065.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:28<00:43,  1.28it/s, est. speed input: 225.64 toks/s, output: 2125.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:28<00:40,  1.36it/s, est. speed input: 222.95 toks/s, output: 2193.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:29<00:34,  1.57it/s, est. speed input: 221.88 toks/s, output: 2276.86 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:26:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6880.3 tokens/s, Running: 54 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:30<00:42,  1.25it/s, est. speed input: 215.21 toks/s, output: 2301.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:31<00:38,  1.35it/s, est. speed input: 213.06 toks/s, output: 2371.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:35,  1.45it/s, est. speed input: 211.28 toks/s, output: 2443.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:32<00:40,  1.24it/s, est. speed input: 206.17 toks/s, output: 2477.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:34<00:48,  1.00it/s, est. speed input: 199.26 toks/s, output: 2486.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:35<00:55,  1.16s/it, est. speed input: 192.41 toks/s, output: 2494.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:35<00:43,  1.09it/s, est. speed input: 192.25 toks/s, output: 2585.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:36<00:33,  1.39it/s, est. speed input: 192.59 toks/s, output: 2681.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:37<00:34,  1.32it/s, est. speed input: 189.74 toks/s, output: 2734.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:37<00:25,  1.75it/s, est. speed input: 190.73 toks/s, output: 2838.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:37<00:15,  2.71it/s, est. speed input: 192.59 toks/s, output: 3046.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:37<00:14,  2.81it/s, est. speed input: 192.64 toks/s, output: 3136.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:38<00:12,  3.20it/s, est. speed input: 193.22 toks/s, output: 3234.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:38<00:09,  4.09it/s, est. speed input: 194.80 toks/s, output: 3437.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:38<00:10,  3.36it/s, est. speed input: 193.99 toks/s, output: 3510.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:27:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5506.1 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:41<00:30,  1.18it/s, est. speed input: 183.69 toks/s, output: 3413.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:41<00:24,  1.44it/s, est. speed input: 183.97 toks/s, output: 3507.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:41<00:19,  1.75it/s, est. speed input: 184.40 toks/s, output: 3602.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:41<00:15,  2.15it/s, est. speed input: 185.07 toks/s, output: 3702.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:13,  2.36it/s, est. speed input: 185.09 toks/s, output: 3790.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:43<00:20,  1.53it/s, est. speed input: 181.28 toks/s, output: 3800.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:43<00:16,  1.86it/s, est. speed input: 181.60 toks/s, output: 3895.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:15,  1.87it/s, est. speed input: 180.81 toks/s, output: 3966.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:44<00:11,  2.45it/s, est. speed input: 181.69 toks/s, output: 4073.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:44<00:09,  2.72it/s, est. speed input: 181.95 toks/s, output: 4166.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:45<00:07,  3.46it/s, est. speed input: 183.04 toks/s, output: 4363.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:06,  3.45it/s, est. speed input: 183.19 toks/s, output: 4453.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:27:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3972.8 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:50<00:35,  1.54s/it, est. speed input: 165.93 toks/s, output: 4126.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:25,  1.17s/it, est. speed input: 166.55 toks/s, output: 4232.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:14,  1.42it/s, est. speed input: 168.24 toks/s, output: 4456.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:51<00:11,  1.72it/s, est. speed input: 168.89 toks/s, output: 4561.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:51<00:12,  1.48it/s, est. speed input: 166.96 toks/s, output: 4599.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:52<00:09,  1.77it/s, est. speed input: 167.31 toks/s, output: 4699.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:52<00:09,  1.77it/s, est. speed input: 166.66 toks/s, output: 4771.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:53<00:09,  1.55it/s, est. speed input: 165.17 toks/s, output: 4818.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:54<00:11,  1.22it/s, est. speed input: 162.50 toks/s, output: 4831.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:55<00:08,  1.47it/s, est. speed input: 162.56 toks/s, output: 4925.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:55<00:04,  2.35it/s, est. speed input: 164.07 toks/s, output: 5152.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:57<00:07,  1.29it/s, est. speed input: 159.84 toks/s, output: 5112.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:27:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2308.6 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:01<00:15,  1.76s/it, est. speed input: 149.05 toks/s, output: 4864.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:02<00:10,  1.32s/it, est. speed input: 149.73 toks/s, output: 4985.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:02<00:07,  1.05s/it, est. speed input: 149.89 toks/s, output: 5089.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  1.05s/it, est. speed input: 156.51 toks/s, output: 6000.81 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  2.56it/s, est. speed input: 156.51 toks/s, output: 6000.81 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:27:33 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:27:33 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:27:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:27:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:27:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:27:33 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:27:33 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:27:33 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:27:33 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:27:33 [abstract.py:306] It took 0.799264 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:27:33 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.4250
+  • Mean Generation Length: 2342.6062
+  • Training FLOPS: 45.87 TFLOPS (11.47 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 288.96s
+  • policy_training: 185.08s (64.1%)
+  • generation: 66.67s (23.1%)
+  • teacher_logprob_inference: 31.78s (11.0%)
+  • policy_training/sharding_data: 4.35s (1.5%)
+  • prepare_for_generation: 3.68s (1.3%)
+  • policy_training/submit_training_futures: 2.05s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.63s (0.6%)
+  • training_prep: 1.16s (0.4%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 167: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:13 [abstract.py:324] It took 0.048843 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09049961e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67345276e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83196945e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09049961e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67345276e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83196945e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09049961e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67345276e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83196945e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09049961e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67345276e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83196945e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09061387e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66195412e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82335052e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.92152405e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.49096298e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #167] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43203918e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09061387e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66195412e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82335052e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66026917e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36463509e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #167] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40664597e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09061387e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66195412e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82335052e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.42095947e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.97481394e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #167] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.17790565e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09061387e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66195412e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82335052e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16068840e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.59423351e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #167] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35180298e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 167: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:16 [abstract.py:324] It took 0.030614 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13356.06it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:16 [loggers.py:236] Engine 000: Avg prompt throughput: 42.4 tokens/s, Avg generation throughput: 17.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:53,  2.95it/s, est. speed input: 177.14 toks/s, output: 41.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 15.68it/s, est. speed input: 778.27 toks/s, output: 242.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 23.26it/s, est. speed input: 1132.49 toks/s, output: 490.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 21.63it/s, est. speed input: 1132.82 toks/s, output: 586.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:06, 23.50it/s, est. speed input: 1209.85 toks/s, output: 741.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:01<00:04, 27.61it/s, est. speed input: 1333.70 toks/s, output: 972.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:04, 32.61it/s, est. speed input: 1475.96 toks/s, output: 1269.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:01, 60.51it/s, est. speed input: 2047.89 toks/s, output: 2266.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:01, 71.44it/s, est. speed input: 2372.30 toks/s, output: 2949.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 73.14it/s, est. speed input: 2542.91 toks/s, output: 3403.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 62.32it/s, est. speed input: 2573.28 toks/s, output: 3700.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:01<00:01, 51.50it/s, est. speed input: 2525.25 toks/s, output: 3896.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:02<00:01, 41.36it/s, est. speed input: 2418.28 toks/s, output: 4004.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:04<00:08,  8.42it/s, est. speed input: 1239.60 toks/s, output: 2424.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:09<00:27,  2.55it/s, est. speed input: 550.01 toks/s, output: 1418.80 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9942.6 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:13<00:34,  1.93it/s, est. speed input: 427.28 toks/s, output: 1394.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:14<00:34,  1.91it/s, est. speed input: 402.77 toks/s, output: 1519.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:14<00:30,  2.08it/s, est. speed input: 395.74 toks/s, output: 1697.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:16<00:36,  1.72it/s, est. speed input: 365.56 toks/s, output: 1672.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:40,  1.51it/s, est. speed input: 344.30 toks/s, output: 1679.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:37,  1.61it/s, est. speed input: 340.18 toks/s, output: 1762.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:18<00:36,  1.60it/s, est. speed input: 331.71 toks/s, output: 1821.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:18<00:27,  2.10it/s, est. speed input: 330.80 toks/s, output: 2019.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:19<00:26,  2.05it/s, est. speed input: 319.83 toks/s, output: 2150.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7709.2 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:21<00:34,  1.58it/s, est. speed input: 303.95 toks/s, output: 2146.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:22<00:36,  1.47it/s, est. speed input: 294.79 toks/s, output: 2182.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:22<00:28,  1.81it/s, est. speed input: 295.60 toks/s, output: 2289.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:22<00:22,  2.25it/s, est. speed input: 296.75 toks/s, output: 2397.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:22<00:19,  2.52it/s, est. speed input: 295.94 toks/s, output: 2489.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:22<00:16,  3.06it/s, est. speed input: 296.67 toks/s, output: 2594.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:23<00:22,  2.18it/s, est. speed input: 289.16 toks/s, output: 2627.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:23<00:15,  2.88it/s, est. speed input: 289.20 toks/s, output: 2821.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:24<00:13,  3.23it/s, est. speed input: 289.47 toks/s, output: 2920.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:24<00:12,  3.50it/s, est. speed input: 289.31 toks/s, output: 3015.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:19,  2.13it/s, est. speed input: 277.90 toks/s, output: 3083.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:26<00:17,  2.34it/s, est. speed input: 277.05 toks/s, output: 3171.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:26<00:19,  2.09it/s, est. speed input: 272.67 toks/s, output: 3218.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:27<00:10,  3.53it/s, est. speed input: 275.80 toks/s, output: 3631.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:27<00:11,  3.18it/s, est. speed input: 273.34 toks/s, output: 3694.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:27<00:08,  3.97it/s, est. speed input: 275.33 toks/s, output: 3906.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:29<00:12,  2.40it/s, est. speed input: 265.01 toks/s, output: 3945.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:13,  2.28it/s, est. speed input: 262.31 toks/s, output: 4000.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5624.3 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:31<00:17,  1.68it/s, est. speed input: 254.31 toks/s, output: 3974.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:32<00:18,  1.52it/s, est. speed input: 249.30 toks/s, output: 3993.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:32<00:19,  1.41it/s, est. speed input: 244.74 toks/s, output: 4016.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:33<00:16,  1.61it/s, est. speed input: 243.75 toks/s, output: 4099.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:33<00:09,  2.43it/s, est. speed input: 245.41 toks/s, output: 4321.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:33<00:08,  2.71it/s, est. speed input: 245.53 toks/s, output: 4419.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:34<00:10,  2.17it/s, est. speed input: 242.12 toks/s, output: 4454.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:34<00:09,  2.26it/s, est. speed input: 241.13 toks/s, output: 4532.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:35<00:07,  2.52it/s, est. speed input: 240.96 toks/s, output: 4626.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:35<00:06,  2.96it/s, est. speed input: 241.40 toks/s, output: 4731.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:35<00:06,  2.66it/s, est. speed input: 239.96 toks/s, output: 4799.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:08,  2.12it/s, est. speed input: 237.04 toks/s, output: 4838.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:36<00:04,  3.24it/s, est. speed input: 238.87 toks/s, output: 5070.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:37<00:04,  3.03it/s, est. speed input: 237.95 toks/s, output: 5148.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:38<00:06,  1.91it/s, est. speed input: 232.89 toks/s, output: 5137.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:38<00:06,  1.78it/s, est. speed input: 230.49 toks/s, output: 5183.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:39<00:05,  2.06it/s, est. speed input: 230.37 toks/s, output: 5281.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:39<00:03,  2.62it/s, est. speed input: 231.17 toks/s, output: 5400.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:39<00:03,  2.74it/s, est. speed input: 230.84 toks/s, output: 5491.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:39<00:02,  3.20it/s, est. speed input: 231.33 toks/s, output: 5602.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:31:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3399.5 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:43<00:09,  1.35s/it, est. speed input: 212.41 toks/s, output: 5255.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:44<00:07,  1.31s/it, est. speed input: 207.99 toks/s, output: 5259.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:07,  1.52s/it, est. speed input: 200.26 toks/s, output: 5181.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:48<00:06,  1.50s/it, est. speed input: 195.51 toks/s, output: 5178.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:48<00:03,  1.11s/it, est. speed input: 195.99 toks/s, output: 5310.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:49<00:02,  1.07s/it, est. speed input: 193.30 toks/s, output: 5359.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:49<00:00,  1.16it/s, est. speed input: 193.14 toks/s, output: 5476.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:32:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1364.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.07s/it, est. speed input: 188.43 toks/s, output: 5467.84 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.07s/it, est. speed input: 188.43 toks/s, output: 5467.84 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  3.11it/s, est. speed input: 188.43 toks/s, output: 5467.84 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:32:10 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:32:10 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:32:11 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:32:11 [abstract.py:306] It took 0.534432 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:32:11 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:32:11 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:32:11 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0120
+  • Training Accuracy: 0.5062
+  • Mean Generation Length: 1759.9812
+  • Training FLOPS: 40.98 TFLOPS (10.24 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 239.28s
+  • policy_training: 153.07s (64.0%)
+  • generation: 54.61s (22.8%)
+  • teacher_logprob_inference: 26.35s (11.0%)
+  • policy_training/sharding_data: 4.27s (1.8%)
+  • prepare_for_generation: 3.87s (1.6%)
+  • policy_training/submit_training_futures: 2.49s (1.0%)
+  • prepare_for_generation/transfer_and_update_weights: 1.75s (0.7%)
+  • training_prep: 0.79s (0.3%)
+  • teacher_logprob_inference_prep: 0.42s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 168: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:12 [abstract.py:324] It took 0.051059 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09073906e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67147827e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83098145e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09073906e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67147827e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83098145e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09073906e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67147827e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83098145e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09073906e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67147827e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83098145e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09071445e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64217796e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81759567e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.93754005e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.31853962e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #168] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43016037e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09071445e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64217796e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81759567e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64544830e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35893383e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #168] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.39555321e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09071445e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64217796e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81759567e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40814819e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.79753304e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #168] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.19693489e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09071445e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64217796e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81759567e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17035980e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61109562e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #168] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35000000e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 168: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:15 [abstract.py:324] It took 0.028986 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13769.59it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:15 [loggers.py:236] Engine 000: Avg prompt throughput: 51.4 tokens/s, Avg generation throughput: 2.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:16,  9.80it/s, est. speed input: 617.60 toks/s, output: 68.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:02, 54.81it/s, est. speed input: 2972.62 toks/s, output: 688.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:02, 53.50it/s, est. speed input: 3065.57 toks/s, output: 921.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:02, 55.65it/s, est. speed input: 3205.59 toks/s, output: 1205.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:02, 51.36it/s, est. speed input: 3105.14 toks/s, output: 1406.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:00<00:02, 46.94it/s, est. speed input: 2971.52 toks/s, output: 1675.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:00<00:02, 46.57it/s, est. speed input: 2941.85 toks/s, output: 1926.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:00<00:01, 62.65it/s, est. speed input: 3314.08 toks/s, output: 2750.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 65.44it/s, est. speed input: 3419.36 toks/s, output: 3232.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 66.80it/s, est. speed input: 3499.97 toks/s, output: 3700.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:01<00:01, 70.72it/s, est. speed input: 3615.77 toks/s, output: 4282.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:01<00:01, 59.82it/s, est. speed input: 3503.22 toks/s, output: 4575.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:01<00:02, 32.29it/s, est. speed input: 2856.96 toks/s, output: 4154.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:04<00:08,  7.34it/s, est. speed input: 1305.38 toks/s, output: 2237.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:07<00:14,  4.06it/s, est. speed input: 846.17 toks/s, output: 1848.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9379.0 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:11<00:25,  2.24it/s, est. speed input: 553.99 toks/s, output: 1515.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:11<00:21,  2.52it/s, est. speed input: 554.82 toks/s, output: 1760.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:12<00:19,  2.61it/s, est. speed input: 536.37 toks/s, output: 1941.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:12<00:16,  2.98it/s, est. speed input: 534.02 toks/s, output: 2169.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:16<00:32,  1.47it/s, est. speed input: 416.29 toks/s, output: 1912.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:18<00:38,  1.23it/s, est. speed input: 381.35 toks/s, output: 1872.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:18<00:32,  1.41it/s, est. speed input: 381.06 toks/s, output: 1990.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:18<00:27,  1.65it/s, est. speed input: 380.99 toks/s, output: 2109.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:18<00:22,  1.93it/s, est. speed input: 380.36 toks/s, output: 2224.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:19<00:21,  1.96it/s, est. speed input: 373.84 toks/s, output: 2304.27 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6742.7 tokens/s, Running: 43 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:20<00:26,  1.58it/s, est. speed input: 358.31 toks/s, output: 2326.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:20<00:16,  2.47it/s, est. speed input: 361.46 toks/s, output: 2578.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:21<00:24,  1.62it/s, est. speed input: 341.89 toks/s, output: 2555.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:22<00:29,  1.29it/s, est. speed input: 325.77 toks/s, output: 2551.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:23<00:24,  1.50it/s, est. speed input: 323.28 toks/s, output: 2648.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:23<00:24,  1.48it/s, est. speed input: 316.42 toks/s, output: 2708.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:25<00:30,  1.13it/s, est. speed input: 300.96 toks/s, output: 2692.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:25<00:24,  1.38it/s, est. speed input: 299.42 toks/s, output: 2794.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:25<00:19,  1.68it/s, est. speed input: 298.51 toks/s, output: 2901.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:26<00:17,  1.86it/s, est. speed input: 296.36 toks/s, output: 2994.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:27<00:25,  1.24it/s, est. speed input: 282.96 toks/s, output: 2975.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:27<00:18,  1.63it/s, est. speed input: 283.64 toks/s, output: 3097.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:28<00:15,  1.88it/s, est. speed input: 282.33 toks/s, output: 3198.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:29<00:08,  3.09it/s, est. speed input: 282.75 toks/s, output: 3650.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:29<00:08,  2.72it/s, est. speed input: 279.43 toks/s, output: 3719.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:29<00:08,  2.76it/s, est. speed input: 278.25 toks/s, output: 3816.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4902.9 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:31<00:11,  1.84it/s, est. speed input: 269.89 toks/s, output: 3814.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:31<00:09,  2.25it/s, est. speed input: 270.64 toks/s, output: 3937.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:31<00:07,  2.53it/s, est. speed input: 270.32 toks/s, output: 4045.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:31<00:05,  3.09it/s, est. speed input: 270.29 toks/s, output: 4267.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:33<00:10,  1.69it/s, est. speed input: 260.14 toks/s, output: 4221.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:33<00:09,  1.75it/s, est. speed input: 258.07 toks/s, output: 4301.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:34<00:07,  2.12it/s, est. speed input: 258.33 toks/s, output: 4419.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:34<00:05,  2.61it/s, est. speed input: 258.96 toks/s, output: 4543.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:36<00:11,  1.15it/s, est. speed input: 245.55 toks/s, output: 4424.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:36<00:08,  1.34it/s, est. speed input: 244.23 toks/s, output: 4516.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:37<00:06,  1.58it/s, est. speed input: 243.59 toks/s, output: 4619.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:37<00:05,  1.73it/s, est. speed input: 242.30 toks/s, output: 4712.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:37<00:04,  2.09it/s, est. speed input: 242.34 toks/s, output: 4829.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:35:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2673.1 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:42<00:12,  1.57s/it, est. speed input: 219.85 toks/s, output: 4504.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:42<00:09,  1.29s/it, est. speed input: 218.08 toks/s, output: 4592.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:43<00:06,  1.05s/it, est. speed input: 217.01 toks/s, output: 4696.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:44<00:05,  1.10s/it, est. speed input: 212.48 toks/s, output: 4725.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:44<00:03,  1.23it/s, est. speed input: 213.11 toks/s, output: 4866.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:46<00:03,  1.10s/it, est. speed input: 206.21 toks/s, output: 4840.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:49<00:03,  1.66s/it, est. speed input: 195.07 toks/s, output: 4716.29 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:49<00:00,  1.66s/it, est. speed input: 197.53 toks/s, output: 5048.96 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:49<00:00,  3.25it/s, est. speed input: 197.53 toks/s, output: 5048.96 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:36:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:36:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:36:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:36:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:36:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:36:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:36:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:36:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:36:07 [abstract.py:306] It took 0.525608 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:36:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:36:07 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.4250
+  • Mean Generation Length: 1554.0750
+  • Training FLOPS: 38.65 TFLOPS (9.66 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 225.10s
+  • policy_training: 144.76s (64.3%)
+  • generation: 52.04s (23.1%)
+  • teacher_logprob_inference: 23.21s (10.3%)
+  • policy_training/sharding_data: 4.26s (1.9%)
+  • prepare_for_generation: 3.70s (1.6%)
+  • policy_training/submit_training_futures: 1.77s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.7%)
+  • training_prep: 0.79s (0.4%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 169: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:38:57 [abstract.py:324] It took 0.053444 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09100586e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66568108e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83019714e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09100586e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66568108e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83019714e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09100586e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66568108e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83019714e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09100586e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66568108e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83019714e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09103145e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64828720e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83241348e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70390701e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.25764275e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #169] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43423798e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09103145e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64828720e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83241348e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63794174e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37048645e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #169] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.38344765e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09103145e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64828720e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83241348e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39930191e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.18075514e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #169] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.23961220e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09103145e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.64828720e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83241348e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16587677e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62236710e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #169] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34846375e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 169: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:01 [abstract.py:324] It took 0.030880 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13287.31it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:01 [loggers.py:236] Engine 000: Avg prompt throughput: 52.4 tokens/s, Avg generation throughput: 62.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:20,  1.98it/s, est. speed input: 122.90 toks/s, output: 43.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:23,  6.67it/s, est. speed input: 342.60 toks/s, output: 168.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:13, 11.26it/s, est. speed input: 547.04 toks/s, output: 348.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:01<00:08, 16.75it/s, est. speed input: 732.46 toks/s, output: 583.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:05, 23.87it/s, est. speed input: 939.89 toks/s, output: 871.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:05, 25.17it/s, est. speed input: 1031.24 toks/s, output: 1055.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:01<00:04, 30.06it/s, est. speed input: 1186.34 toks/s, output: 1370.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██        | 33/160 [00:01<00:03, 32.38it/s, est. speed input: 1300.59 toks/s, output: 1665.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 38.28it/s, est. speed input: 1439.61 toks/s, output: 2034.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:03, 36.90it/s, est. speed input: 1494.26 toks/s, output: 2277.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 39.34it/s, est. speed input: 1570.94 toks/s, output: 2569.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:02<00:02, 39.84it/s, est. speed input: 1628.63 toks/s, output: 2840.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:03<00:07, 12.81it/s, est. speed input: 1186.38 toks/s, output: 2303.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:09<00:43,  2.21it/s, est. speed input: 417.09 toks/s, output: 1004.82 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11429.8 tokens/s, Running: 97 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:12<00:56,  1.66it/s, est. speed input: 318.30 toks/s, output: 1039.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:15<01:04,  1.43it/s, est. speed input: 276.70 toks/s, output: 1077.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:20<01:34,  1.05s/it, est. speed input: 214.25 toks/s, output: 1004.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9254.5 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.5%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:21<01:28,  1.01s/it, est. speed input: 202.83 toks/s, output: 1121.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:22<01:21,  1.07it/s, est. speed input: 201.06 toks/s, output: 1198.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:23<01:26,  1.00s/it, est. speed input: 192.01 toks/s, output: 1230.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:24<01:20,  1.05it/s, est. speed input: 188.84 toks/s, output: 1295.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:25<01:17,  1.08it/s, est. speed input: 185.05 toks/s, output: 1353.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:26<01:03,  1.29it/s, est. speed input: 182.14 toks/s, output: 1496.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:27<00:53,  1.48it/s, est. speed input: 180.02 toks/s, output: 1640.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:29<01:19,  1.00s/it, est. speed input: 167.98 toks/s, output: 1610.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:30<01:10,  1.10it/s, est. speed input: 166.72 toks/s, output: 1679.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7859.9 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.9%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:30<01:00,  1.27it/s, est. speed input: 166.45 toks/s, output: 1756.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:30<00:53,  1.41it/s, est. speed input: 165.79 toks/s, output: 1828.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:31<00:51,  1.47it/s, est. speed input: 164.56 toks/s, output: 1893.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:32<00:36,  2.03it/s, est. speed input: 165.73 toks/s, output: 2061.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:32<00:27,  2.57it/s, est. speed input: 167.19 toks/s, output: 2230.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:32<00:24,  2.82it/s, est. speed input: 167.83 toks/s, output: 2314.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:35<01:03,  1.08it/s, est. speed input: 156.31 toks/s, output: 2230.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:37<01:18,  1.16s/it, est. speed input: 150.13 toks/s, output: 2218.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:37<01:03,  1.05it/s, est. speed input: 150.26 toks/s, output: 2294.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:39<01:13,  1.12s/it, est. speed input: 145.83 toks/s, output: 2301.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6549.0 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 12.0%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:42<01:58,  1.83s/it, est. speed input: 134.91 toks/s, output: 2204.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:43<01:35,  1.49s/it, est. speed input: 134.29 toks/s, output: 2270.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:44<01:21,  1.29s/it, est. speed input: 133.16 toks/s, output: 2325.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:44<00:58,  1.06it/s, est. speed input: 134.23 toks/s, output: 2418.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:45<00:52,  1.17it/s, est. speed input: 133.67 toks/s, output: 2481.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:46<00:51,  1.16it/s, est. speed input: 132.41 toks/s, output: 2531.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:46<00:47,  1.24it/s, est. speed input: 131.86 toks/s, output: 2592.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:47<00:31,  1.82it/s, est. speed input: 133.08 toks/s, output: 2760.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:47<00:28,  1.98it/s, est. speed input: 133.32 toks/s, output: 2836.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:47<00:23,  2.32it/s, est. speed input: 133.91 toks/s, output: 2921.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:48<00:20,  2.67it/s, est. speed input: 134.53 toks/s, output: 3005.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:49<00:35,  1.50it/s, est. speed input: 131.87 toks/s, output: 3017.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:49<00:27,  1.90it/s, est. speed input: 132.63 toks/s, output: 3104.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:50<00:26,  1.92it/s, est. speed input: 132.47 toks/s, output: 3171.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:39:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5977.5 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 11.6%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:51<00:34,  1.44it/s, est. speed input: 130.77 toks/s, output: 3200.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:52<00:46,  1.05it/s, est. speed input: 128.01 toks/s, output: 3204.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:53<00:44,  1.09it/s, est. speed input: 127.11 toks/s, output: 3253.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:53<00:24,  1.89it/s, est. speed input: 129.05 toks/s, output: 3442.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:53<00:19,  2.35it/s, est. speed input: 129.89 toks/s, output: 3534.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:55<00:31,  1.40it/s, est. speed input: 127.42 toks/s, output: 3536.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:56<00:27,  1.50it/s, est. speed input: 126.79 toks/s, output: 3658.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:57<00:27,  1.47it/s, est. speed input: 126.25 toks/s, output: 3711.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:58<00:27,  1.47it/s, est. speed input: 125.84 toks/s, output: 3768.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:58<00:22,  1.70it/s, est. speed input: 126.16 toks/s, output: 3846.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:58<00:18,  2.01it/s, est. speed input: 126.59 toks/s, output: 3930.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:40:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4814.2 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.5%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [01:01<00:38,  1.03s/it, est. speed input: 122.56 toks/s, output: 3875.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [01:04<01:05,  1.82s/it, est. speed input: 116.32 toks/s, output: 3749.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [01:05<00:47,  1.36s/it, est. speed input: 116.83 toks/s, output: 3837.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [01:06<00:38,  1.17s/it, est. speed input: 115.37 toks/s, output: 3930.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [01:07<00:35,  1.10s/it, est. speed input: 114.72 toks/s, output: 3978.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [01:08<00:34,  1.12s/it, est. speed input: 113.69 toks/s, output: 4014.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [01:09<00:28,  1.05it/s, est. speed input: 113.75 toks/s, output: 4087.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [01:10<00:18,  1.51it/s, est. speed input: 114.51 toks/s, output: 4257.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:40:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4069.8 tokens/s, Running: 28 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.3%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [01:10<00:16,  1.67it/s, est. speed input: 114.72 toks/s, output: 4335.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [01:12<00:21,  1.19it/s, est. speed input: 112.61 toks/s, output: 4395.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [01:14<00:18,  1.27it/s, est. speed input: 112.16 toks/s, output: 4519.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [01:15<00:19,  1.15it/s, est. speed input: 111.18 toks/s, output: 4551.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [01:17<00:23,  1.14s/it, est. speed input: 109.05 toks/s, output: 4537.60 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:17<00:00,  1.14s/it, est. speed input: 125.23 toks/s, output: 6751.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:17<00:00,  2.06it/s, est. speed input: 125.23 toks/s, output: 6751.22 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:40:23 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:40:23 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:40:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:40:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:40:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:40:23 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:40:23 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:40:23 [abstract.py:306] It took 0.540984 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:40:23 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:40:23 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:40:23 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.4125
+  • Mean Generation Length: 3272.3938
+  • Training FLOPS: 51.18 TFLOPS (12.80 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 363.56s
+  • policy_training: 234.12s (64.4%)
+  • generation: 82.83s (22.8%)
+  • teacher_logprob_inference: 41.44s (11.4%)
+  • policy_training/sharding_data: 4.34s (1.2%)
+  • prepare_for_generation: 3.80s (1.0%)
+  • policy_training/submit_training_futures: 1.91s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.4%)
+  • training_prep: 0.80s (0.2%)
+  • teacher_logprob_inference_prep: 0.38s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 170: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:01 [abstract.py:324] It took 0.050805 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09178809e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66364288e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83411255e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09178809e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66364288e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83411255e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09178809e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66364288e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83411255e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09178809e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66364288e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83411255e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09192168e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66189804e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82020798e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.80102921e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.11869144e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #170] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43319153e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09192168e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66189804e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82020798e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64664307e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37556334e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #170] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.37334862e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09192168e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66189804e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82020798e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39496613e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.80816793e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #170] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.24744034e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09192168e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66189804e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82020798e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.16556091e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.60281467e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #170] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.34864075e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 170: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:04 [abstract.py:324] It took 0.031635 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13675.31it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:05 [loggers.py:236] Engine 000: Avg prompt throughput: 32.9 tokens/s, Avg generation throughput: 80.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:17,  2.04it/s, est. speed input: 120.55 toks/s, output: 24.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:10, 14.91it/s, est. speed input: 709.43 toks/s, output: 207.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:07, 20.34it/s, est. speed input: 938.16 toks/s, output: 339.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 33.49it/s, est. speed input: 1394.54 toks/s, output: 653.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:02, 47.09it/s, est. speed input: 1853.39 toks/s, output: 1092.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:02, 45.55it/s, est. speed input: 1952.02 toks/s, output: 1320.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 55.22it/s, est. speed input: 2232.69 toks/s, output: 1780.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 50.27it/s, est. speed input: 2265.63 toks/s, output: 2044.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 58.39it/s, est. speed input: 2503.36 toks/s, output: 2703.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 55.05it/s, est. speed input: 2533.81 toks/s, output: 3013.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:02<00:02, 31.38it/s, est. speed input: 2171.75 toks/s, output: 2826.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:04<00:09,  8.33it/s, est. speed input: 1157.95 toks/s, output: 1787.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:08<00:21,  3.42it/s, est. speed input: 636.78 toks/s, output: 1326.14 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10094.3 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:11<00:32,  2.20it/s, est. speed input: 457.55 toks/s, output: 1226.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:13<00:35,  1.97it/s, est. speed input: 410.73 toks/s, output: 1296.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<00:41,  1.62it/s, est. speed input: 358.27 toks/s, output: 1329.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:16<00:44,  1.51it/s, est. speed input: 339.80 toks/s, output: 1362.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:17<00:43,  1.51it/s, est. speed input: 330.10 toks/s, output: 1425.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:18<00:46,  1.41it/s, est. speed input: 316.31 toks/s, output: 1466.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:19<00:53,  1.19it/s, est. speed input: 296.56 toks/s, output: 1477.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:19<00:45,  1.38it/s, est. speed input: 295.05 toks/s, output: 1570.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:20<00:37,  1.64it/s, est. speed input: 295.02 toks/s, output: 1669.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7896.3 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:23<00:57,  1.05it/s, est. speed input: 262.56 toks/s, output: 1674.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:24<00:58,  1.01it/s, est. speed input: 252.79 toks/s, output: 1712.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:25<00:55,  1.05it/s, est. speed input: 246.97 toks/s, output: 1773.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:25<00:43,  1.31it/s, est. speed input: 247.36 toks/s, output: 1874.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:27<00:56,  1.02s/it, est. speed input: 233.80 toks/s, output: 1871.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:27<00:51,  1.07it/s, est. speed input: 229.97 toks/s, output: 1938.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:28<00:43,  1.23it/s, est. speed input: 227.90 toks/s, output: 2019.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:49,  1.07it/s, est. speed input: 220.36 toks/s, output: 2050.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:45,  1.13it/s, est. speed input: 216.87 toks/s, output: 2115.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6705.0 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:32<01:02,  1.23s/it, est. speed input: 204.79 toks/s, output: 2095.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:33<00:42,  1.17it/s, est. speed input: 203.26 toks/s, output: 2273.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:33,  1.44it/s, est. speed input: 203.90 toks/s, output: 2376.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:25,  1.82it/s, est. speed input: 204.81 toks/s, output: 2483.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:22,  2.03it/s, est. speed input: 204.56 toks/s, output: 2574.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:34<00:21,  2.06it/s, est. speed input: 202.47 toks/s, output: 2733.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:35<00:21,  2.02it/s, est. speed input: 201.17 toks/s, output: 2809.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:35<00:17,  2.30it/s, est. speed input: 200.67 toks/s, output: 2986.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:20,  1.98it/s, est. speed input: 198.33 toks/s, output: 3042.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:17,  2.18it/s, est. speed input: 198.26 toks/s, output: 3132.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:38<00:25,  1.47it/s, est. speed input: 193.01 toks/s, output: 3141.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:39<00:19,  1.87it/s, est. speed input: 192.73 toks/s, output: 3319.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:39<00:15,  2.21it/s, est. speed input: 193.43 toks/s, output: 3421.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:39<00:14,  2.35it/s, est. speed input: 193.28 toks/s, output: 3508.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:40<00:15,  2.17it/s, est. speed input: 192.05 toks/s, output: 3577.14 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5287.7 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:40<00:16,  1.91it/s, est. speed input: 190.25 toks/s, output: 3634.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:41<00:18,  1.72it/s, est. speed input: 188.31 toks/s, output: 3688.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:44<00:36,  1.21s/it, est. speed input: 177.86 toks/s, output: 3577.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:45<00:24,  1.16it/s, est. speed input: 177.11 toks/s, output: 3745.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:45<00:18,  1.44it/s, est. speed input: 177.70 toks/s, output: 3850.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:46<00:20,  1.28it/s, est. speed input: 175.07 toks/s, output: 3886.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:14,  1.66it/s, est. speed input: 175.00 toks/s, output: 4067.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:47<00:11,  2.06it/s, est. speed input: 175.84 toks/s, output: 4179.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:48<00:16,  1.31it/s, est. speed input: 171.32 toks/s, output: 4165.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:50<00:14,  1.40it/s, est. speed input: 169.28 toks/s, output: 4300.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:45:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3948.3 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:50<00:12,  1.54it/s, est. speed input: 168.97 toks/s, output: 4385.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:50<00:09,  1.93it/s, est. speed input: 169.80 toks/s, output: 4499.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:52<00:12,  1.35it/s, est. speed input: 166.44 toks/s, output: 4505.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:54<00:18,  1.14s/it, est. speed input: 160.75 toks/s, output: 4446.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:12,  1.16it/s, est. speed input: 161.46 toks/s, output: 4560.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:56<00:18,  1.31s/it, est. speed input: 155.54 toks/s, output: 4491.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:57<00:13,  1.01s/it, est. speed input: 155.88 toks/s, output: 4596.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:57<00:10,  1.10it/s, est. speed input: 155.12 toks/s, output: 4672.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:00<00:14,  1.32s/it, est. speed input: 150.18 toks/s, output: 4624.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:46:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2494.2 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 75.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:01<00:09,  1.08s/it, est. speed input: 148.17 toks/s, output: 4763.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:01<00:06,  1.15it/s, est. speed input: 148.60 toks/s, output: 4878.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.15it/s, est. speed input: 156.32 toks/s, output: 5937.98 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.59it/s, est. speed input: 156.32 toks/s, output: 5937.98 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:46:10 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:46:10 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:46:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:46:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:46:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:46:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:46:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:46:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:46:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:46:10 [abstract.py:306] It took 0.576206 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:46:10 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.2750
+  • Mean Generation Length: 2296.2312
+  • Training FLOPS: 44.25 TFLOPS (11.06 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 292.77s
+  • policy_training: 188.78s (64.5%)
+  • generation: 65.80s (22.5%)
+  • teacher_logprob_inference: 32.62s (11.1%)
+  • policy_training/sharding_data: 4.34s (1.5%)
+  • prepare_for_generation: 3.92s (1.3%)
+  • policy_training/submit_training_futures: 2.05s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.91s (0.7%)
+  • training_prep: 1.07s (0.4%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 35/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 171: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:49:54 [abstract.py:324] It took 0.052320 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09193359e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65635185e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83175659e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09193359e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65635185e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83175659e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09193359e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65635185e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83175659e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09193359e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65635185e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83175659e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09201289e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66294174e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81876984e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66959381e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.16830730e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #171] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43699249e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09201289e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66294174e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81876984e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64512177e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37160616e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #171] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.37661057e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09201289e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66294174e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81876984e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38849792e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-9.32776928e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #171] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.27216835e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09201289e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66294174e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81876984e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15504913e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61535683e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #171] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35003937e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 171: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:49:57 [abstract.py:324] It took 0.028048 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13182.38it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:49:58 [loggers.py:236] Engine 000: Avg prompt throughput: 41.6 tokens/s, Avg generation throughput: 13.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:11,  2.21it/s, est. speed input: 134.91 toks/s, output: 33.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:08, 18.19it/s, est. speed input: 866.01 toks/s, output: 298.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:08, 18.05it/s, est. speed input: 931.18 toks/s, output: 395.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:06, 22.60it/s, est. speed input: 1114.10 toks/s, output: 623.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:01<00:04, 30.68it/s, est. speed input: 1356.67 toks/s, output: 944.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 34.64it/s, est. speed input: 1527.59 toks/s, output: 1269.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 38/160 [00:01<00:02, 42.80it/s, est. speed input: 1751.87 toks/s, output: 1726.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 42.92it/s, est. speed input: 1821.26 toks/s, output: 1960.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 53.05it/s, est. speed input: 2071.59 toks/s, output: 2603.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:02, 49.61it/s, est. speed input: 2108.38 toks/s, output: 2857.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:02<00:02, 36.79it/s, est. speed input: 2001.23 toks/s, output: 2978.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:03<00:07, 11.86it/s, est. speed input: 1303.36 toks/s, output: 2156.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:06<00:18,  4.66it/s, est. speed input: 754.70 toks/s, output: 1542.09 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 12:50:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10614.1 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:12<00:49,  1.65it/s, est. speed input: 367.84 toks/s, output: 1000.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:13<00:45,  1.75it/s, est. speed input: 356.39 toks/s, output: 1162.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:16<00:53,  1.45it/s, est. speed input: 308.67 toks/s, output: 1186.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:16<00:50,  1.54it/s, est. speed input: 306.25 toks/s, output: 1272.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:17<00:47,  1.59it/s, est. speed input: 293.54 toks/s, output: 1403.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:18<00:44,  1.65it/s, est. speed input: 289.15 toks/s, output: 1475.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:18<00:45,  1.59it/s, est. speed input: 281.03 toks/s, output: 1527.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:50:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8350.9 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<01:00,  1.18it/s, est. speed input: 260.50 toks/s, output: 1508.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:21<00:56,  1.27it/s, est. speed input: 256.19 toks/s, output: 1574.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<00:44,  1.57it/s, est. speed input: 257.15 toks/s, output: 1671.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<00:35,  1.92it/s, est. speed input: 257.70 toks/s, output: 1765.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:30,  2.25it/s, est. speed input: 257.69 toks/s, output: 1855.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:22<00:38,  1.74it/s, est. speed input: 249.97 toks/s, output: 1888.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:24<01:01,  1.08it/s, est. speed input: 233.87 toks/s, output: 1856.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:24<00:53,  1.22it/s, est. speed input: 231.04 toks/s, output: 1923.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:26<01:07,  1.06s/it, est. speed input: 218.95 toks/s, output: 1912.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:26<00:50,  1.24it/s, est. speed input: 219.55 toks/s, output: 2006.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:27<00:43,  1.43it/s, est. speed input: 218.12 toks/s, output: 2081.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:27<00:36,  1.66it/s, est. speed input: 217.35 toks/s, output: 2161.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:27<00:22,  2.66it/s, est. speed input: 220.12 toks/s, output: 2360.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:28<00:19,  2.99it/s, est. speed input: 220.63 toks/s, output: 2451.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:28<00:17,  3.21it/s, est. speed input: 220.85 toks/s, output: 2538.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:28<00:19,  2.93it/s, est. speed input: 219.85 toks/s, output: 2610.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:29<00:21,  2.53it/s, est. speed input: 217.87 toks/s, output: 2671.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:29<00:17,  3.07it/s, est. speed input: 218.78 toks/s, output: 2766.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:17,  3.02it/s, est. speed input: 218.30 toks/s, output: 2842.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:29<00:16,  3.19it/s, est. speed input: 218.44 toks/s, output: 2926.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:13,  3.92it/s, est. speed input: 219.66 toks/s, output: 3024.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:50:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6987.4 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:30<00:14,  3.45it/s, est. speed input: 219.05 toks/s, output: 3096.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:18,  2.68it/s, est. speed input: 216.95 toks/s, output: 3149.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:31<00:24,  1.95it/s, est. speed input: 213.10 toks/s, output: 3175.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:32<00:29,  1.61it/s, est. speed input: 209.21 toks/s, output: 3200.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:33<00:20,  2.20it/s, est. speed input: 209.60 toks/s, output: 3370.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:35<00:34,  1.28it/s, est. speed input: 200.60 toks/s, output: 3309.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:45,  1.05s/it, est. speed input: 192.38 toks/s, output: 3257.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:37<00:37,  1.11it/s, est. speed input: 191.49 toks/s, output: 3326.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:37<00:31,  1.30it/s, est. speed input: 190.90 toks/s, output: 3400.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:38<00:30,  1.32it/s, est. speed input: 188.80 toks/s, output: 3447.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:38,  1.00it/s, est. speed input: 182.91 toks/s, output: 3423.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:50:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5446.1 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:41<00:44,  1.16s/it, est. speed input: 177.46 toks/s, output: 3407.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:38,  1.04s/it, est. speed input: 175.86 toks/s, output: 3461.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:42<00:30,  1.19it/s, est. speed input: 175.82 toks/s, output: 3545.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:42<00:22,  1.58it/s, est. speed input: 176.63 toks/s, output: 3646.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:43<00:14,  2.32it/s, est. speed input: 177.80 toks/s, output: 3839.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:43<00:11,  2.79it/s, est. speed input: 178.59 toks/s, output: 3941.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:43<00:11,  2.66it/s, est. speed input: 178.20 toks/s, output: 4016.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:44<00:10,  2.85it/s, est. speed input: 178.44 toks/s, output: 4104.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:12,  2.33it/s, est. speed input: 177.26 toks/s, output: 4161.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:45<00:06,  3.84it/s, est. speed input: 179.82 toks/s, output: 4470.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:45<00:05,  4.19it/s, est. speed input: 180.51 toks/s, output: 4569.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:05,  4.48it/s, est. speed input: 181.15 toks/s, output: 4667.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:45<00:04,  4.63it/s, est. speed input: 181.70 toks/s, output: 4762.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:45<00:04,  4.66it/s, est. speed input: 182.15 toks/s, output: 4855.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:14,  1.46it/s, est. speed input: 176.02 toks/s, output: 4773.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:48<00:09,  2.09it/s, est. speed input: 176.97 toks/s, output: 4965.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:48<00:08,  2.18it/s, est. speed input: 176.77 toks/s, output: 5043.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:10,  1.69it/s, est. speed input: 174.50 toks/s, output: 5061.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:50:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3825.5 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:13,  1.18it/s, est. speed input: 170.34 toks/s, output: 5025.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:51<00:11,  1.32it/s, est. speed input: 169.82 toks/s, output: 5096.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:52<00:09,  1.52it/s, est. speed input: 169.66 toks/s, output: 5176.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:05,  2.37it/s, est. speed input: 171.11 toks/s, output: 5391.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:07,  1.54it/s, est. speed input: 167.91 toks/s, output: 5377.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:56<00:10,  1.09s/it, est. speed input: 161.90 toks/s, output: 5275.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:56<00:07,  1.16it/s, est. speed input: 162.35 toks/s, output: 5380.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:57<00:07,  1.00it/s, est. speed input: 159.58 toks/s, output: 5382.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:58<00:06,  1.16it/s, est. speed input: 159.16 toks/s, output: 5461.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:06,  1.11s/it, est. speed input: 155.57 toks/s, output: 5434.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:04,  1.16it/s, est. speed input: 155.93 toks/s, output: 5543.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:50:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2091.4 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:00<00:03,  1.32it/s, est. speed input: 155.63 toks/s, output: 5629.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:01<00:01,  1.55it/s, est. speed input: 155.67 toks/s, output: 5727.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:01<00:00,  2.00it/s, est. speed input: 156.25 toks/s, output: 5846.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:01<00:00,  2.61it/s, est. speed input: 156.95 toks/s, output: 5969.77 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.61it/s, est. speed input: 157.93 toks/s, output: 6103.30 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.61it/s, est. speed input: 157.93 toks/s, output: 6103.30 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:51:02 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:51:02 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:51:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:51:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:51:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:51:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:51:02 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 12:51:03 [abstract.py:306] It took 0.532982 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:51:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:51:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:51:03 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0114
+  • Training Accuracy: 0.4313
+  • Mean Generation Length: 2340.0063
+  • Training FLOPS: 43.04 TFLOPS (10.76 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 303.10s
+  • policy_training: 195.44s (64.5%)
+  • generation: 65.34s (21.6%)
+  • teacher_logprob_inference: 36.98s (12.2%)
+  • policy_training/sharding_data: 4.26s (1.4%)
+  • prepare_for_generation: 3.70s (1.2%)
+  • policy_training/submit_training_futures: 1.65s (0.5%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.5%)
+  • training_prep: 0.97s (0.3%)
+  • teacher_logprob_inference_prep: 0.49s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 172: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:54:57 [abstract.py:324] It took 0.049767 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09221758e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65052414e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83073730e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09221758e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65052414e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83073730e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09221758e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65052414e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83073730e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09221758e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65052414e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83073730e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09224141e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65707474e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82180023e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66856194e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.17310667e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #172] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43593842e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09224141e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65707474e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82180023e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64891052e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35842791e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #172] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.38455276e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09224141e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65707474e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82180023e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39174728e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-8.71020317e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #172] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.29018822e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09224141e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65707474e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82180023e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17179985e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61204262e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #172] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35011536e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 172: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:00 [abstract.py:324] It took 0.029166 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:00 [loggers.py:236] Engine 000: Avg prompt throughput: 40.0 tokens/s, Avg generation throughput: 4.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13984.51it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:26,  6.11it/s, est. speed input: 366.93 toks/s, output: 79.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 40.22it/s, est. speed input: 2115.49 toks/s, output: 637.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:04, 34.57it/s, est. speed input: 1994.40 toks/s, output: 785.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 35.05it/s, est. speed input: 2033.52 toks/s, output: 983.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 31.35it/s, est. speed input: 1935.42 toks/s, output: 1127.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:04, 29.30it/s, est. speed input: 1865.85 toks/s, output: 1286.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 52.08it/s, est. speed input: 2432.88 toks/s, output: 2323.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 54.56it/s, est. speed input: 2566.51 toks/s, output: 2747.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 57.88it/s, est. speed input: 2693.22 toks/s, output: 3183.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 51.76it/s, est. speed input: 2677.71 toks/s, output: 3475.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 33.70it/s, est. speed input: 2358.37 toks/s, output: 3356.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:05<00:19,  4.52it/s, est. speed input: 764.85 toks/s, output: 1289.86 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:08<00:26,  3.15it/s, est. speed input: 554.34 toks/s, output: 1313.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10625.8 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:12<00:41,  1.94it/s, est. speed input: 389.07 toks/s, output: 1194.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:13<00:38,  2.05it/s, est. speed input: 378.23 toks/s, output: 1356.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<00:52,  1.46it/s, est. speed input: 312.27 toks/s, output: 1301.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:17<00:50,  1.48it/s, est. speed input: 304.97 toks/s, output: 1368.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:17<00:41,  1.76it/s, est. speed input: 303.54 toks/s, output: 1553.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:19<00:52,  1.38it/s, est. speed input: 280.82 toks/s, output: 1532.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:19<00:47,  1.49it/s, est. speed input: 278.01 toks/s, output: 1611.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8274.4 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:20<00:58,  1.20it/s, est. speed input: 261.05 toks/s, output: 1608.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<00:55,  1.25it/s, est. speed input: 255.46 toks/s, output: 1667.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:45,  1.48it/s, est. speed input: 255.15 toks/s, output: 1757.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:22<00:36,  1.82it/s, est. speed input: 255.79 toks/s, output: 1852.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:23<00:57,  1.15it/s, est. speed input: 239.52 toks/s, output: 1827.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:25<01:08,  1.06s/it, est. speed input: 227.39 toks/s, output: 1826.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:26<01:02,  1.02it/s, est. speed input: 222.75 toks/s, output: 1881.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:26<00:41,  1.48it/s, est. speed input: 222.49 toks/s, output: 2057.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:27<00:37,  1.64it/s, est. speed input: 221.36 toks/s, output: 2136.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:27<00:33,  1.80it/s, est. speed input: 220.39 toks/s, output: 2216.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:27<00:28,  2.09it/s, est. speed input: 220.39 toks/s, output: 2305.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:29<00:39,  1.45it/s, est. speed input: 213.15 toks/s, output: 2317.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:29<00:26,  2.12it/s, est. speed input: 214.30 toks/s, output: 2505.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:14,  3.74it/s, est. speed input: 219.29 toks/s, output: 2823.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7041.0 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:23,  2.23it/s, est. speed input: 212.61 toks/s, output: 2822.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:32<00:30,  1.68it/s, est. speed input: 206.81 toks/s, output: 2832.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:32<00:29,  1.67it/s, est. speed input: 204.81 toks/s, output: 2890.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:32<00:23,  2.11it/s, est. speed input: 206.05 toks/s, output: 2992.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:27,  1.73it/s, est. speed input: 202.52 toks/s, output: 3027.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:35<00:39,  1.19it/s, est. speed input: 195.31 toks/s, output: 3005.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:35<00:34,  1.33it/s, est. speed input: 194.15 toks/s, output: 3073.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:36<00:25,  1.73it/s, est. speed input: 193.62 toks/s, output: 3233.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:20,  2.15it/s, est. speed input: 194.66 toks/s, output: 3335.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:37<00:25,  1.67it/s, est. speed input: 191.13 toks/s, output: 3359.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:37<00:22,  1.86it/s, est. speed input: 190.83 toks/s, output: 3439.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:38<00:16,  2.38it/s, est. speed input: 191.88 toks/s, output: 3541.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:38<00:13,  2.94it/s, est. speed input: 192.72 toks/s, output: 3640.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:38<00:18,  2.05it/s, est. speed input: 190.03 toks/s, output: 3674.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:39<00:15,  2.43it/s, est. speed input: 190.46 toks/s, output: 3765.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:39<00:13,  2.65it/s, est. speed input: 190.59 toks/s, output: 3850.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:39<00:10,  3.33it/s, est. speed input: 191.49 toks/s, output: 3952.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5520.7 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:40<00:15,  2.16it/s, est. speed input: 189.01 toks/s, output: 3982.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:41<00:18,  1.82it/s, est. speed input: 186.98 toks/s, output: 4023.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:28,  1.14it/s, est. speed input: 181.19 toks/s, output: 3983.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:43<00:26,  1.19it/s, est. speed input: 179.49 toks/s, output: 4029.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:44<00:21,  1.42it/s, est. speed input: 179.30 toks/s, output: 4109.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:45<00:26,  1.09it/s, est. speed input: 175.06 toks/s, output: 4098.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:24,  1.15it/s, est. speed input: 173.40 toks/s, output: 4146.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:21,  1.28it/s, est. speed input: 172.61 toks/s, output: 4213.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:46<00:15,  1.66it/s, est. speed input: 173.29 toks/s, output: 4313.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:49<00:31,  1.24s/it, est. speed input: 164.94 toks/s, output: 4193.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:55:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4038.8 tokens/s, Running: 25 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:50<00:27,  1.16s/it, est. speed input: 162.92 toks/s, output: 4230.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:21,  1.05it/s, est. speed input: 162.60 toks/s, output: 4310.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:16,  1.32it/s, est. speed input: 162.84 toks/s, output: 4404.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:52<00:17,  1.23it/s, est. speed input: 161.09 toks/s, output: 4444.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:53<00:17,  1.16it/s, est. speed input: 159.24 toks/s, output: 4482.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:54<00:20,  1.07s/it, est. speed input: 155.77 toks/s, output: 4475.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:55<00:15,  1.13it/s, est. speed input: 155.62 toks/s, output: 4560.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:55<00:12,  1.34it/s, est. speed input: 155.51 toks/s, output: 4647.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:57<00:14,  1.07it/s, est. speed input: 152.90 toks/s, output: 4659.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:57<00:10,  1.45it/s, est. speed input: 153.64 toks/s, output: 4772.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:57<00:09,  1.55it/s, est. speed input: 153.20 toks/s, output: 4849.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:59<00:11,  1.15it/s, est. speed input: 150.63 toks/s, output: 4859.78 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:56:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2955.1 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:00<00:10,  1.12it/s, est. speed input: 149.34 toks/s, output: 4909.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:01<00:07,  1.38it/s, est. speed input: 148.73 toks/s, output: 5073.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:03<00:10,  1.13s/it, est. speed input: 144.14 toks/s, output: 5012.85 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.13s/it, est. speed input: 152.60 toks/s, output: 6172.23 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.52it/s, est. speed input: 152.60 toks/s, output: 6172.23 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:56:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 12:56:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 12:56:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:56:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:56:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:56:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 12:56:08 [abstract.py:306] It took 0.541848 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 12:56:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 12:56:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 12:56:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 12:56:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0106
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 2453.1313
+  • Training FLOPS: 46.79 TFLOPS (11.70 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 296.37s
+  • policy_training: 189.98s (64.1%)
+  • generation: 67.68s (22.8%)
+  • teacher_logprob_inference: 33.27s (11.2%)
+  • policy_training/sharding_data: 4.35s (1.5%)
+  • prepare_for_generation: 3.64s (1.2%)
+  • policy_training/submit_training_futures: 2.25s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.5%)
+  • training_prep: 1.22s (0.4%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 173: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 12:59:54 [abstract.py:324] It took 0.049828 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242578e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64311676e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83674011e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242578e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64311676e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83674011e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242578e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64311676e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83674011e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242578e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64311676e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83674011e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239824e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63963165e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84599152e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62452698e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.12449551e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #173] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43791687e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239824e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63963165e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84599152e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63692703e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34922333e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #173] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.38701286e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239824e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63963165e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84599152e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40271988e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-8.31022263e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #173] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.31747169e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239824e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63963165e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84599152e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19937592e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61186085e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #173] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35145798e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 173: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 12:59:57 [abstract.py:324] It took 0.030534 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12940.89it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 12:59:57 [loggers.py:236] Engine 000: Avg prompt throughput: 41.2 tokens/s, Avg generation throughput: 29.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:43,  3.65it/s, est. speed input: 215.12 toks/s, output: 40.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:12, 12.38it/s, est. speed input: 627.09 toks/s, output: 154.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:06, 22.76it/s, est. speed input: 1076.34 toks/s, output: 385.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 30.64it/s, est. speed input: 1398.51 toks/s, output: 634.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 27.60it/s, est. speed input: 1406.23 toks/s, output: 755.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 39.16it/s, est. speed input: 1768.30 toks/s, output: 1235.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 41.29it/s, est. speed input: 1893.24 toks/s, output: 1498.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 35.97it/s, est. speed input: 1863.20 toks/s, output: 1685.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 46.17it/s, est. speed input: 2105.95 toks/s, output: 2255.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 51.27it/s, est. speed input: 2249.38 toks/s, output: 2687.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 51.92it/s, est. speed input: 2324.86 toks/s, output: 3021.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:01<00:02, 47.73it/s, est. speed input: 2334.52 toks/s, output: 3292.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:01, 44.53it/s, est. speed input: 2338.99 toks/s, output: 3593.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:02<00:02, 28.46it/s, est. speed input: 2092.20 toks/s, output: 3441.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:05<00:15,  5.25it/s, est. speed input: 947.96 toks/s, output: 1805.67 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:08<00:26,  2.96it/s, est. speed input: 632.15 toks/s, output: 1467.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:08<00:23,  3.18it/s, est. speed input: 618.70 toks/s, output: 1628.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:09<00:25,  2.83it/s, est. speed input: 562.17 toks/s, output: 1675.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:07 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10121.4 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:10<00:26,  2.63it/s, est. speed input: 521.90 toks/s, output: 1746.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:10<00:25,  2.74it/s, est. speed input: 514.66 toks/s, output: 1818.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:10<00:23,  2.97it/s, est. speed input: 511.61 toks/s, output: 1903.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:11<00:27,  2.45it/s, est. speed input: 484.55 toks/s, output: 1898.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:12<00:37,  1.80it/s, est. speed input: 446.84 toks/s, output: 1847.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:13<00:34,  1.89it/s, est. speed input: 436.63 toks/s, output: 1901.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:13<00:31,  2.05it/s, est. speed input: 429.36 toks/s, output: 1965.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:14<00:44,  1.43it/s, est. speed input: 395.66 toks/s, output: 1909.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:15<00:49,  1.28it/s, est. speed input: 374.39 toks/s, output: 1904.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:16<00:38,  1.59it/s, est. speed input: 362.32 toks/s, output: 2034.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:19<01:07,  1.12s/it, est. speed input: 315.26 toks/s, output: 1870.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:19<00:52,  1.12it/s, est. speed input: 314.89 toks/s, output: 1967.41 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:17 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7754.9 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:58,  1.01s/it, est. speed input: 297.81 toks/s, output: 1960.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:21<00:47,  1.20it/s, est. speed input: 295.32 toks/s, output: 2044.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:21<00:42,  1.32it/s, est. speed input: 290.78 toks/s, output: 2110.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:22<00:36,  1.51it/s, est. speed input: 287.91 toks/s, output: 2188.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:35,  1.52it/s, est. speed input: 282.47 toks/s, output: 2244.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:36,  1.43it/s, est. speed input: 275.68 toks/s, output: 2288.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:24<00:40,  1.28it/s, est. speed input: 267.08 toks/s, output: 2314.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:43,  1.17it/s, est. speed input: 258.80 toks/s, output: 2341.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:25<00:25,  1.95it/s, est. speed input: 261.30 toks/s, output: 2558.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:19,  2.42it/s, est. speed input: 262.60 toks/s, output: 2666.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:26<00:18,  2.60it/s, est. speed input: 261.90 toks/s, output: 2755.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:26<00:14,  3.26it/s, est. speed input: 263.16 toks/s, output: 2863.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:27<00:17,  2.63it/s, est. speed input: 259.82 toks/s, output: 2922.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:27<00:16,  2.60it/s, est. speed input: 258.33 toks/s, output: 2999.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:27<00:16,  2.54it/s, est. speed input: 256.65 toks/s, output: 3074.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:29<00:35,  1.17it/s, est. speed input: 241.84 toks/s, output: 2992.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:15,  2.47it/s, est. speed input: 246.75 toks/s, output: 3338.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:27 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6194.6 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.2%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:30<00:12,  2.94it/s, est. speed input: 246.43 toks/s, output: 3609.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:31<00:10,  3.40it/s, est. speed input: 247.35 toks/s, output: 3805.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:32<00:13,  2.50it/s, est. speed input: 242.36 toks/s, output: 3819.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:23,  1.38it/s, est. speed input: 229.68 toks/s, output: 3713.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:34<00:23,  1.33it/s, est. speed input: 225.89 toks/s, output: 3747.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:18,  1.66it/s, est. speed input: 226.76 toks/s, output: 3855.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:35<00:16,  1.79it/s, est. speed input: 225.74 toks/s, output: 3932.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:36<00:17,  1.57it/s, est. speed input: 222.04 toks/s, output: 3963.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:16,  1.60it/s, est. speed input: 220.11 toks/s, output: 4024.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:16,  1.61it/s, est. speed input: 218.19 toks/s, output: 4083.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:15,  1.65it/s, est. speed input: 216.47 toks/s, output: 4147.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:14,  1.70it/s, est. speed input: 214.95 toks/s, output: 4214.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:14,  1.64it/s, est. speed input: 212.86 toks/s, output: 4269.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:39<00:12,  1.70it/s, est. speed input: 211.53 toks/s, output: 4338.29 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:37 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4365.8 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:40<00:11,  1.84it/s, est. speed input: 210.72 toks/s, output: 4417.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:41<00:17,  1.11it/s, est. speed input: 203.45 toks/s, output: 4363.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:43<00:19,  1.02s/it, est. speed input: 198.67 toks/s, output: 4360.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:43<00:11,  1.46it/s, est. speed input: 198.79 toks/s, output: 4562.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:47<00:20,  1.31s/it, est. speed input: 186.57 toks/s, output: 4383.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:47<00:14,  1.00it/s, est. speed input: 187.44 toks/s, output: 4504.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:47<00:10,  1.28it/s, est. speed input: 187.85 toks/s, output: 4615.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:47<00:08,  1.54it/s, est. speed input: 187.98 toks/s, output: 4719.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:48<00:06,  1.75it/s, est. speed input: 187.78 toks/s, output: 4815.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:48<00:04,  2.48it/s, est. speed input: 188.75 toks/s, output: 5040.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:48<00:03,  2.51it/s, est. speed input: 188.50 toks/s, output: 5134.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2734.4 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.8%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:50<00:05,  1.39it/s, est. speed input: 183.58 toks/s, output: 5103.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:50<00:04,  1.71it/s, est. speed input: 183.97 toks/s, output: 5217.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:51<00:04,  1.44it/s, est. speed input: 181.71 toks/s, output: 5257.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:55<00:07,  1.56s/it, est. speed input: 170.53 toks/s, output: 5047.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:56<00:05,  1.36s/it, est. speed input: 168.96 toks/s, output: 5115.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.36s/it, est. speed input: 173.15 toks/s, output: 5697.03 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.84it/s, est. speed input: 173.15 toks/s, output: 5697.03 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:56 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:56 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:00:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:00:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:00:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:00:57 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:00:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:00:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:00:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:00:57 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:00:57 [abstract.py:306] It took 0.555896 seconds to fall asleep.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0110
+  • Training Accuracy: 0.4437
+  • Mean Generation Length: 2005.3438
+  • Training FLOPS: 40.68 TFLOPS (10.17 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 272.92s
+  • policy_training: 177.36s (65.0%)
+  • generation: 59.78s (21.9%)
+  • teacher_logprob_inference: 30.14s (11.0%)
+  • policy_training/sharding_data: 4.36s (1.6%)
+  • prepare_for_generation: 3.76s (1.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • policy_training/submit_training_futures: 1.57s (0.6%)
+  • training_prep: 1.10s (0.4%)
+  • teacher_logprob_inference_prep: 0.62s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 174: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:04:27 [abstract.py:324] It took 0.049843 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257812e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64695969e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84798126e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257812e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64695969e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84798126e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257812e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64695969e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84798126e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257812e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64695969e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84798126e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09254844e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63918839e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86231842e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.51257706e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.17939377e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #174] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43725601e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09254844e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63918839e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86231842e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64323730e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.34480677e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #174] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.40735016e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09254844e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63918839e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86231842e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39753265e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.57268906e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #174] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.37104073e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09254844e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63918839e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86231842e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.21573105e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61682739e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #174] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35427139e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 174: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:04:30 [abstract.py:324] It took 0.030347 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13515.30it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 13:04:31 [loggers.py:236] Engine 000: Avg prompt throughput: 43.2 tokens/s, Avg generation throughput: 38.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:50,  1.44it/s, est. speed input: 88.00 toks/s, output: 17.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:13, 11.34it/s, est. speed input: 535.66 toks/s, output: 138.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 24.59it/s, est. speed input: 1015.11 toks/s, output: 363.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:01<00:04, 28.53it/s, est. speed input: 1206.46 toks/s, output: 565.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:04, 27.91it/s, est. speed input: 1272.99 toks/s, output: 726.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 37.53it/s, est. speed input: 1522.31 toks/s, output: 1089.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:03, 39.21it/s, est. speed input: 1625.50 toks/s, output: 1338.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 39.96it/s, est. speed input: 1689.68 toks/s, output: 1567.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 41.04it/s, est. speed input: 1752.32 toks/s, output: 1797.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 45.04it/s, est. speed input: 1858.78 toks/s, output: 2150.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:02<00:02, 41.07it/s, est. speed input: 1869.31 toks/s, output: 2350.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:02<00:02, 42.47it/s, est. speed input: 1919.06 toks/s, output: 2605.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:02<00:02, 42.95it/s, est. speed input: 1958.88 toks/s, output: 2859.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:05<00:17,  4.79it/s, est. speed input: 839.13 toks/s, output: 1446.34 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:07<00:21,  3.62it/s, est. speed input: 655.85 toks/s, output: 1455.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:30,  2.52it/s, est. speed input: 507.11 toks/s, output: 1381.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:04:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10325.6 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:34,  2.15it/s, est. speed input: 446.81 toks/s, output: 1399.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:15<00:53,  1.34it/s, est. speed input: 343.63 toks/s, output: 1272.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:17<00:59,  1.19it/s, est. speed input: 316.26 toks/s, output: 1269.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:17<00:54,  1.30it/s, est. speed input: 312.79 toks/s, output: 1353.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:17<00:45,  1.52it/s, est. speed input: 314.07 toks/s, output: 1456.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:18<00:46,  1.46it/s, est. speed input: 303.74 toks/s, output: 1505.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:19<00:53,  1.25it/s, est. speed input: 287.83 toks/s, output: 1522.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:04:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7983.2 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:21<01:05,  1.01it/s, est. speed input: 269.05 toks/s, output: 1519.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:21<00:50,  1.29it/s, est. speed input: 269.89 toks/s, output: 1619.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:21<00:43,  1.47it/s, est. speed input: 267.58 toks/s, output: 1699.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<00:42,  1.48it/s, est. speed input: 262.34 toks/s, output: 1760.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:22<00:32,  1.91it/s, est. speed input: 263.28 toks/s, output: 1860.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:22<00:26,  2.34it/s, est. speed input: 263.72 toks/s, output: 1956.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:23<00:25,  2.37it/s, est. speed input: 261.68 toks/s, output: 2032.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:23<00:19,  2.91it/s, est. speed input: 261.19 toks/s, output: 2210.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:23<00:19,  2.94it/s, est. speed input: 260.02 toks/s, output: 2292.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:24<00:11,  4.61it/s, est. speed input: 264.26 toks/s, output: 2684.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:24<00:14,  3.71it/s, est. speed input: 261.25 toks/s, output: 2740.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:13,  3.89it/s, est. speed input: 261.56 toks/s, output: 2830.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:25<00:11,  4.23it/s, est. speed input: 262.20 toks/s, output: 3009.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:25<00:10,  4.57it/s, est. speed input: 263.02 toks/s, output: 3190.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:26<00:13,  3.44it/s, est. speed input: 258.92 toks/s, output: 3309.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:28<00:21,  2.07it/s, est. speed input: 249.06 toks/s, output: 3269.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:28<00:18,  2.32it/s, est. speed input: 249.08 toks/s, output: 3355.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:30<00:34,  1.22it/s, est. speed input: 233.98 toks/s, output: 3239.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:05:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6300.4 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.8%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:30<00:27,  1.51it/s, est. speed input: 234.43 toks/s, output: 3333.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:30<00:21,  1.87it/s, est. speed input: 235.03 toks/s, output: 3429.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:31<00:18,  2.16it/s, est. speed input: 234.90 toks/s, output: 3515.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:32<00:27,  1.41it/s, est. speed input: 227.03 toks/s, output: 3485.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:32<00:20,  1.79it/s, est. speed input: 227.68 toks/s, output: 3582.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:17,  2.12it/s, est. speed input: 227.71 toks/s, output: 3670.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:33<00:20,  1.74it/s, est. speed input: 223.89 toks/s, output: 3697.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:35<00:28,  1.20it/s, est. speed input: 216.33 toks/s, output: 3661.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:36<00:34,  1.03s/it, est. speed input: 209.19 toks/s, output: 3630.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:37<00:27,  1.17it/s, est. speed input: 208.39 toks/s, output: 3706.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:37<00:20,  1.52it/s, est. speed input: 208.86 toks/s, output: 3805.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:37<00:14,  2.00it/s, est. speed input: 209.79 toks/s, output: 3910.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:37<00:14,  1.97it/s, est. speed input: 208.44 toks/s, output: 3975.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:38<00:10,  2.59it/s, est. speed input: 209.44 toks/s, output: 4084.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:39<00:16,  1.60it/s, est. speed input: 204.66 toks/s, output: 4080.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:40<00:21,  1.23it/s, est. speed input: 199.86 toks/s, output: 4075.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:05:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4542.9 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.8%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:40<00:15,  1.59it/s, est. speed input: 200.32 toks/s, output: 4175.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:41<00:17,  1.38it/s, est. speed input: 197.22 toks/s, output: 4200.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:42<00:16,  1.35it/s, est. speed input: 195.03 toks/s, output: 4245.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:43<00:16,  1.33it/s, est. speed input: 192.92 toks/s, output: 4290.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:43<00:15,  1.32it/s, est. speed input: 190.99 toks/s, output: 4339.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:44<00:12,  1.64it/s, est. speed input: 191.24 toks/s, output: 4436.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:45<00:15,  1.20it/s, est. speed input: 186.81 toks/s, output: 4427.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:46<00:13,  1.29it/s, est. speed input: 185.52 toks/s, output: 4491.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:46<00:10,  1.65it/s, est. speed input: 185.96 toks/s, output: 4595.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:46<00:09,  1.67it/s, est. speed input: 184.94 toks/s, output: 4665.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:15,  1.05s/it, est. speed input: 178.26 toks/s, output: 4592.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:49<00:08,  1.46it/s, est. speed input: 178.82 toks/s, output: 4800.95 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:05:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3155.9 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:50<00:08,  1.34it/s, est. speed input: 176.69 toks/s, output: 4842.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:50<00:07,  1.52it/s, est. speed input: 176.40 toks/s, output: 4932.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:51<00:05,  1.79it/s, est. speed input: 176.55 toks/s, output: 5034.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:51<00:05,  1.75it/s, est. speed input: 175.66 toks/s, output: 5107.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:53<00:05,  1.34it/s, est. speed input: 172.83 toks/s, output: 5126.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:07,  1.10s/it, est. speed input: 167.77 toks/s, output: 5078.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:05,  1.17it/s, est. speed input: 168.08 toks/s, output: 5191.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:57<00:06,  1.33s/it, est. speed input: 161.87 toks/s, output: 5107.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:58<00:04,  1.08s/it, est. speed input: 161.58 toks/s, output: 5207.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.08s/it, est. speed input: 165.72 toks/s, output: 5770.23 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.75it/s, est. speed input: 165.72 toks/s, output: 5770.23 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:05:31 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:05:31 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:05:31 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:05:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:05:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:05:32 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:05:32 [abstract.py:306] It took 0.539352 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:05:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:05:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:05:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:05:32 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0119
+  • Training Accuracy: 0.3125
+  • Mean Generation Length: 2099.5688
+  • Training FLOPS: 43.17 TFLOPS (10.79 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 271.98s
+  • policy_training: 174.87s (64.3%)
+  • generation: 61.80s (22.7%)
+  • teacher_logprob_inference: 29.92s (11.0%)
+  • policy_training/sharding_data: 4.34s (1.6%)
+  • prepare_for_generation: 3.74s (1.4%)
+  • policy_training/submit_training_futures: 1.94s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.73s (0.6%)
+  • training_prep: 0.94s (0.3%)
+  • teacher_logprob_inference_prep: 0.55s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 175: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:08:59 [abstract.py:324] It took 0.049060 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09255723e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64938698e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84660339e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09255723e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64938698e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84660339e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09255723e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64938698e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84660339e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09255723e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64938698e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84660339e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259219e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66992569e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85408325e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66300201e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.15862656e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #175] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43827026e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259219e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66992569e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85408325e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64097595e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35080910e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #175] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43650055e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259219e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66992569e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85408325e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39471893e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.46003389e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #175] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.40740776e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259219e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66992569e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85408325e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19938774e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.61525478e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #175] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35301971e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 175: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:02 [abstract.py:324] It took 0.030972 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12471.45it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:03 [loggers.py:236] Engine 000: Avg prompt throughput: 44.2 tokens/s, Avg generation throughput: 57.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:31,  5.08it/s, est. speed input: 320.22 toks/s, output: 76.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:06, 23.43it/s, est. speed input: 1230.66 toks/s, output: 380.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:05, 28.10it/s, est. speed input: 1484.53 toks/s, output: 581.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:05, 25.17it/s, est. speed input: 1450.34 toks/s, output: 725.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:05, 27.73it/s, est. speed input: 1554.07 toks/s, output: 928.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 29.42it/s, est. speed input: 1628.01 toks/s, output: 1148.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:04, 30.70it/s, est. speed input: 1680.58 toks/s, output: 1361.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:04, 29.70it/s, est. speed input: 1687.02 toks/s, output: 1533.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:01<00:03, 33.71it/s, est. speed input: 1798.18 toks/s, output: 1935.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:03, 34.40it/s, est. speed input: 1832.66 toks/s, output: 2160.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:03, 33.91it/s, est. speed input: 1848.12 toks/s, output: 2408.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 38.82it/s, est. speed input: 1943.34 toks/s, output: 2827.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:03, 32.13it/s, est. speed input: 1879.19 toks/s, output: 2931.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:02<00:08, 12.01it/s, est. speed input: 1347.42 toks/s, output: 2314.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:04<00:21,  4.58it/s, est. speed input: 805.29 toks/s, output: 1586.73 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11378.6 tokens/s, Running: 96 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:11<01:13,  1.30it/s, est. speed input: 345.46 toks/s, output: 857.10 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:16<01:38,  1.06s/it, est. speed input: 257.91 toks/s, output: 825.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:16<01:28,  1.04it/s, est. speed input: 256.70 toks/s, output: 915.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:17<01:27,  1.04it/s, est. speed input: 246.42 toks/s, output: 971.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:17<01:19,  1.13it/s, est. speed input: 242.30 toks/s, output: 1047.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:18<01:12,  1.24it/s, est. speed input: 238.89 toks/s, output: 1122.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:18<01:00,  1.47it/s, est. speed input: 239.00 toks/s, output: 1212.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:19<01:03,  1.38it/s, est. speed input: 231.59 toks/s, output: 1263.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:19<00:51,  1.66it/s, est. speed input: 231.89 toks/s, output: 1352.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9254.5 tokens/s, Running: 86 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:20<00:49,  1.70it/s, est. speed input: 228.61 toks/s, output: 1419.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:20<00:39,  2.08it/s, est. speed input: 226.86 toks/s, output: 1577.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:21<00:50,  1.61it/s, est. speed input: 218.63 toks/s, output: 1604.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:22<00:56,  1.43it/s, est. speed input: 212.56 toks/s, output: 1643.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:24<01:18,  1.03it/s, est. speed input: 199.99 toks/s, output: 1629.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:25<01:09,  1.14it/s, est. speed input: 197.57 toks/s, output: 1693.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:26<01:16,  1.02it/s, est. speed input: 190.69 toks/s, output: 1716.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:26<00:47,  1.61it/s, est. speed input: 192.75 toks/s, output: 1897.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:27<00:50,  1.48it/s, est. speed input: 188.98 toks/s, output: 1940.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:27<00:42,  1.75it/s, est. speed input: 189.38 toks/s, output: 2023.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:28<00:50,  1.45it/s, est. speed input: 184.79 toks/s, output: 2055.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:29<00:39,  1.84it/s, est. speed input: 185.82 toks/s, output: 2145.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:29<00:31,  2.22it/s, est. speed input: 186.62 toks/s, output: 2232.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7611.9 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.9%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:30<00:30,  2.27it/s, est. speed input: 185.47 toks/s, output: 2369.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:30<00:31,  2.16it/s, est. speed input: 184.26 toks/s, output: 2430.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:30<00:25,  2.61it/s, est. speed input: 185.40 toks/s, output: 2520.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:31<00:18,  3.52it/s, est. speed input: 187.56 toks/s, output: 2698.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:32<00:33,  1.91it/s, est. speed input: 181.81 toks/s, output: 2691.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:33<00:47,  1.32it/s, est. speed input: 175.81 toks/s, output: 2678.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:34<00:38,  1.61it/s, est. speed input: 176.37 toks/s, output: 2760.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:34<00:33,  1.84it/s, est. speed input: 176.43 toks/s, output: 2835.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:35<00:32,  1.86it/s, est. speed input: 175.55 toks/s, output: 2895.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:35<00:31,  1.88it/s, est. speed input: 174.69 toks/s, output: 2955.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:35<00:18,  3.15it/s, est. speed input: 177.63 toks/s, output: 3150.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:35<00:16,  3.39it/s, est. speed input: 178.22 toks/s, output: 3232.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:36<00:17,  3.09it/s, est. speed input: 177.95 toks/s, output: 3298.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:36<00:12,  4.16it/s, est. speed input: 180.00 toks/s, output: 3479.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:37<00:25,  2.07it/s, est. speed input: 175.43 toks/s, output: 3464.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:38<00:16,  3.10it/s, est. speed input: 177.94 toks/s, output: 3655.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:38<00:17,  2.88it/s, est. speed input: 177.58 toks/s, output: 3717.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:38<00:15,  3.06it/s, est. speed input: 178.00 toks/s, output: 3795.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:39<00:23,  1.99it/s, est. speed input: 175.01 toks/s, output: 3801.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6359.2 tokens/s, Running: 47 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.0%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:40<00:25,  1.83it/s, est. speed input: 173.61 toks/s, output: 3842.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:41<00:23,  1.84it/s, est. speed input: 171.92 toks/s, output: 3947.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:41<00:17,  2.46it/s, est. speed input: 173.41 toks/s, output: 4121.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:41<00:11,  3.52it/s, est. speed input: 175.88 toks/s, output: 4318.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:11,  3.34it/s, est. speed input: 175.80 toks/s, output: 4386.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:10,  3.46it/s, est. speed input: 176.19 toks/s, output: 4465.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:42<00:09,  4.04it/s, est. speed input: 177.18 toks/s, output: 4557.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:42<00:08,  4.03it/s, est. speed input: 177.55 toks/s, output: 4636.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:43<00:15,  2.21it/s, est. speed input: 174.84 toks/s, output: 4635.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:18,  1.83it/s, est. speed input: 173.15 toks/s, output: 4660.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:17,  1.86it/s, est. speed input: 172.50 toks/s, output: 4712.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:45<00:13,  2.22it/s, est. speed input: 172.54 toks/s, output: 4853.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:46<00:11,  2.59it/s, est. speed input: 173.15 toks/s, output: 4939.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:09,  2.95it/s, est. speed input: 173.74 toks/s, output: 5024.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:08,  3.47it/s, est. speed input: 174.43 toks/s, output: 5114.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:07,  3.44it/s, est. speed input: 174.66 toks/s, output: 5188.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:10,  2.53it/s, est. speed input: 173.50 toks/s, output: 5223.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:47<00:07,  3.18it/s, est. speed input: 174.36 toks/s, output: 5390.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:48<00:09,  2.37it/s, est. speed input: 172.90 toks/s, output: 5414.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:07,  2.93it/s, est. speed input: 173.80 toks/s, output: 5581.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:49<00:06,  3.01it/s, est. speed input: 173.96 toks/s, output: 5656.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:05,  3.35it/s, est. speed input: 174.66 toks/s, output: 5818.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:09:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4283.6 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:50<00:04,  3.52it/s, est. speed input: 175.11 toks/s, output: 5900.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:50<00:04,  3.31it/s, est. speed input: 175.05 toks/s, output: 5969.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:51<00:06,  2.44it/s, est. speed input: 173.74 toks/s, output: 5995.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:52<00:08,  1.61it/s, est. speed input: 170.96 toks/s, output: 5971.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:04,  2.60it/s, est. speed input: 172.80 toks/s, output: 6180.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:53<00:04,  2.46it/s, est. speed input: 172.41 toks/s, output: 6238.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:55<00:08,  1.15it/s, est. speed input: 166.43 toks/s, output: 6101.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:06,  1.42it/s, est. speed input: 166.76 toks/s, output: 6193.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:56<00:06,  1.16it/s, est. speed input: 164.02 toks/s, output: 6172.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:57<00:06,  1.09it/s, est. speed input: 162.07 toks/s, output: 6181.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:59<00:07,  1.19s/it, est. speed input: 158.05 toks/s, output: 6114.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:10:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2104.0 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:00<00:04,  1.04it/s, est. speed input: 157.95 toks/s, output: 6196.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:00<00:03,  1.26it/s, est. speed input: 157.91 toks/s, output: 6283.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:02<00:03,  1.13s/it, est. speed input: 154.04 toks/s, output: 6217.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [01:03<00:02,  1.06s/it, est. speed input: 152.77 toks/s, output: 6258.33 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.06s/it, est. speed input: 154.69 toks/s, output: 6516.93 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.53it/s, est. speed input: 154.69 toks/s, output: 6516.93 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:10:10 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:10:10 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:10:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:10:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:10:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:10:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:10:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 13:10:10 [abstract.py:306] It took 0.524703 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:10:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:10:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:10:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0104
+  • Training Accuracy: 0.3625
+  • Mean Generation Length: 2580.3875
+  • Training FLOPS: 45.80 TFLOPS (11.45 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 309.61s
+  • policy_training: 200.85s (64.9%)
+  • generation: 67.58s (21.8%)
+  • teacher_logprob_inference: 36.18s (11.7%)
+  • policy_training/sharding_data: 4.30s (1.4%)
+  • prepare_for_generation: 3.66s (1.2%)
+  • policy_training/submit_training_futures: 2.45s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.5%)
+  • training_prep: 0.79s (0.3%)
+  • teacher_logprob_inference_prep: 0.37s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Epoch 36/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 176: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:09 [abstract.py:324] It took 0.054316 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09280156e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65918159e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84025574e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09280156e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65918159e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84025574e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09280156e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65918159e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84025574e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09280156e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.65918159e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84025574e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286484e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68476868e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82749557e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.81243515e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.08700800e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #176] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43691925e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286484e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68476868e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82749557e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64238281e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36261406e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #176] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43354492e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286484e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68476868e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82749557e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37488708e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.36458158e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #176] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.40507622e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286484e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68476868e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82749557e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.18844147e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62557678e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #176] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35438629e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 176: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:12 [abstract.py:324] It took 0.029083 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13414.80it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:13 [loggers.py:236] Engine 000: Avg prompt throughput: 38.9 tokens/s, Avg generation throughput: 11.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:10,  2.24it/s, est. speed input: 139.08 toks/s, output: 29.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:09, 16.05it/s, est. speed input: 774.57 toks/s, output: 210.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 27.11it/s, est. speed input: 1212.17 toks/s, output: 456.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 42.60it/s, est. speed input: 1754.69 toks/s, output: 855.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 41.25it/s, est. speed input: 1861.58 toks/s, output: 1041.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 49.31it/s, est. speed input: 2128.55 toks/s, output: 1444.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 54.65it/s, est. speed input: 2335.68 toks/s, output: 1841.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 65.44it/s, est. speed input: 2626.67 toks/s, output: 2424.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 58.76it/s, est. speed input: 2660.92 toks/s, output: 2752.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:01<00:01, 65.43it/s, est. speed input: 2836.59 toks/s, output: 3285.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:01<00:01, 54.74it/s, est. speed input: 2782.69 toks/s, output: 3550.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:01<00:01, 56.65it/s, est. speed input: 2843.11 toks/s, output: 3958.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:02<00:03, 17.66it/s, est. speed input: 1922.14 toks/s, output: 2979.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:06<00:13,  4.64it/s, est. speed input: 876.47 toks/s, output: 1827.77 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:10<00:19,  2.91it/s, est. speed input: 608.51 toks/s, output: 1688.46 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9498.8 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 76.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:13<00:25,  2.18it/s, est. speed input: 485.37 toks/s, output: 1658.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:14<00:25,  2.06it/s, est. speed input: 450.81 toks/s, output: 1773.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:16<00:27,  1.84it/s, est. speed input: 411.94 toks/s, output: 1851.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:17<00:32,  1.55it/s, est. speed input: 381.10 toks/s, output: 1828.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:19<00:36,  1.33it/s, est. speed input: 355.65 toks/s, output: 1822.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:19<00:32,  1.46it/s, est. speed input: 352.24 toks/s, output: 1920.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7049.2 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 76.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:20<00:36,  1.29it/s, est. speed input: 335.15 toks/s, output: 1943.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:21<00:38,  1.19it/s, est. speed input: 321.40 toks/s, output: 1978.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:22<00:39,  1.13it/s, est. speed input: 309.51 toks/s, output: 2020.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:23<00:40,  1.08it/s, est. speed input: 298.29 toks/s, output: 2062.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:24<00:42,  1.02it/s, est. speed input: 286.99 toks/s, output: 2098.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:35,  1.17it/s, est. speed input: 283.48 toks/s, output: 2187.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:26<00:35,  1.17it/s, est. speed input: 276.54 toks/s, output: 2247.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:27<00:38,  1.03it/s, est. speed input: 266.02 toks/s, output: 2275.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:27<00:32,  1.19it/s, est. speed input: 263.29 toks/s, output: 2365.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:28<00:20,  1.78it/s, est. speed input: 263.42 toks/s, output: 2588.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:29<00:28,  1.27it/s, est. speed input: 252.43 toks/s, output: 2592.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5516.5 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 76.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:30<00:29,  1.20it/s, est. speed input: 246.51 toks/s, output: 2643.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:31<00:23,  1.47it/s, est. speed input: 246.33 toks/s, output: 2751.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:31<00:16,  1.89it/s, est. speed input: 244.88 toks/s, output: 2953.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:15,  2.06it/s, est. speed input: 244.13 toks/s, output: 3054.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:32<00:15,  1.93it/s, est. speed input: 241.33 toks/s, output: 3129.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:34<00:21,  1.34it/s, est. speed input: 233.36 toks/s, output: 3136.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:34<00:11,  2.30it/s, est. speed input: 235.30 toks/s, output: 3488.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:35<00:15,  1.61it/s, est. speed input: 228.22 toks/s, output: 3491.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:12,  1.92it/s, est. speed input: 228.75 toks/s, output: 3607.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:36<00:10,  2.11it/s, est. speed input: 228.36 toks/s, output: 3709.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:37<00:11,  1.94it/s, est. speed input: 226.05 toks/s, output: 3780.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:08,  2.38it/s, est. speed input: 226.70 toks/s, output: 3898.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:38<00:08,  2.28it/s, est. speed input: 224.37 toks/s, output: 4070.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:38<00:06,  2.78it/s, est. speed input: 225.41 toks/s, output: 4195.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:12,  1.34it/s, est. speed input: 216.23 toks/s, output: 4134.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:40<00:09,  1.70it/s, est. speed input: 216.97 toks/s, output: 4257.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:14:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3911.5 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 76.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:40<00:08,  1.75it/s, est. speed input: 215.73 toks/s, output: 4339.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:08,  1.63it/s, est. speed input: 213.48 toks/s, output: 4403.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:42<00:08,  1.49it/s, est. speed input: 210.80 toks/s, output: 4458.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:43<00:08,  1.41it/s, est. speed input: 208.32 toks/s, output: 4516.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:43<00:05,  1.88it/s, est. speed input: 208.00 toks/s, output: 4730.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:04,  1.81it/s, est. speed input: 206.45 toks/s, output: 4806.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:04,  1.66it/s, est. speed input: 204.36 toks/s, output: 4869.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:45<00:03,  1.79it/s, est. speed input: 203.71 toks/s, output: 4967.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:45<00:02,  2.03it/s, est. speed input: 203.70 toks/s, output: 5076.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:46<00:03,  1.63it/s, est. speed input: 201.01 toks/s, output: 5124.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:47<00:02,  1.47it/s, est. speed input: 198.69 toks/s, output: 5181.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:15:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1705.8 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 76.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:50<00:02,  1.11s/it, est. speed input: 188.38 toks/s, output: 5145.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:52<00:01,  1.18s/it, est. speed input: 184.46 toks/s, output: 5164.22 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.18s/it, est. speed input: 185.64 toks/s, output: 5320.54 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.05it/s, est. speed input: 185.64 toks/s, output: 5320.54 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:15:07 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:15:07 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:15:07 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:15:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 13:15:08 [abstract.py:306] It took 0.509623 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:15:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:15:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:15:08 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0116
+  • Training Accuracy: 0.4500
+  • Mean Generation Length: 1742.5188
+  • Training FLOPS: 38.82 TFLOPS (9.70 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 249.09s
+  • policy_training: 162.10s (65.1%)
+  • generation: 55.44s (22.3%)
+  • teacher_logprob_inference: 26.46s (10.6%)
+  • policy_training/sharding_data: 4.33s (1.7%)
+  • prepare_for_generation: 3.72s (1.5%)
+  • policy_training/submit_training_futures: 1.82s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.66s (0.7%)
+  • training_prep: 0.81s (0.3%)
+  • teacher_logprob_inference_prep: 0.39s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 177: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:18:19 [abstract.py:324] It took 0.048511 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09263496e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66945038e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83373260e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09263496e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66945038e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83373260e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09263496e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66945038e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83373260e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09263496e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66945038e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83373260e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09272246e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70712128e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82516708e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.85384369e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.17058611e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #177] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44013916e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09272246e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70712128e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82516708e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65465012e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36550484e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #177] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44069214e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09272246e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70712128e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82516708e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37556534e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.84289312e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #177] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.45455780e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09272246e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70712128e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82516708e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17458839e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62098484e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #177] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35289337e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 177: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:18:22 [abstract.py:324] It took 0.031051 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 13:18:22 [loggers.py:236] Engine 000: Avg prompt throughput: 48.5 tokens/s, Avg generation throughput: 5.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13329.53it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:24,  6.52it/s, est. speed input: 411.10 toks/s, output: 78.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 43.72it/s, est. speed input: 2244.21 toks/s, output: 601.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:02, 53.78it/s, est. speed input: 2797.38 toks/s, output: 978.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 39.27it/s, est. speed input: 2375.52 toks/s, output: 1123.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 38.49it/s, est. speed input: 2352.09 toks/s, output: 1359.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:03, 39.83it/s, est. speed input: 2389.73 toks/s, output: 1667.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:00<00:02, 52.09it/s, est. speed input: 2702.27 toks/s, output: 2340.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 48.78it/s, est. speed input: 2684.06 toks/s, output: 2602.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:01, 55.31it/s, est. speed input: 2840.84 toks/s, output: 3133.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:01, 55.20it/s, est. speed input: 2879.61 toks/s, output: 3477.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:01, 45.88it/s, est. speed input: 2765.50 toks/s, output: 3633.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:02, 32.50it/s, est. speed input: 2493.90 toks/s, output: 3559.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:04<00:14,  5.50it/s, est. speed input: 995.42 toks/s, output: 1704.55 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:07<00:23,  3.27it/s, est. speed input: 673.29 toks/s, output: 1444.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:09<00:29,  2.51it/s, est. speed input: 535.00 toks/s, output: 1448.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:18:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10278.9 tokens/s, Running: 75 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:10<00:29,  2.48it/s, est. speed input: 502.22 toks/s, output: 1561.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:11<00:31,  2.23it/s, est. speed input: 458.29 toks/s, output: 1616.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:12<00:32,  2.14it/s, est. speed input: 429.54 toks/s, output: 1708.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:13<00:30,  2.27it/s, est. speed input: 425.09 toks/s, output: 1789.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:15<00:40,  1.64it/s, est. speed input: 374.50 toks/s, output: 1762.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:15<00:40,  1.61it/s, est. speed input: 362.34 toks/s, output: 1804.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:16<00:45,  1.41it/s, est. speed input: 342.35 toks/s, output: 1805.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:18<00:52,  1.21it/s, est. speed input: 322.33 toks/s, output: 1799.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:19<00:53,  1.17it/s, est. speed input: 309.40 toks/s, output: 1827.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:41,  1.46it/s, est. speed input: 309.90 toks/s, output: 1930.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:18:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7876.8 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:42,  1.40it/s, est. speed input: 300.50 toks/s, output: 1971.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:20<00:40,  1.47it/s, est. speed input: 294.69 toks/s, output: 2033.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:30,  1.90it/s, est. speed input: 295.73 toks/s, output: 2138.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:21<00:25,  2.22it/s, est. speed input: 294.99 toks/s, output: 2230.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:21<00:18,  3.03it/s, est. speed input: 295.62 toks/s, output: 2427.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:57,  1.06s/it, est. speed input: 258.39 toks/s, output: 2218.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:51,  1.03it/s, est. speed input: 253.73 toks/s, output: 2276.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:40,  1.28it/s, est. speed input: 253.27 toks/s, output: 2369.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:26<00:33,  1.54it/s, est. speed input: 252.65 toks/s, output: 2460.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:25,  1.95it/s, est. speed input: 253.23 toks/s, output: 2562.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:26<00:20,  2.34it/s, est. speed input: 253.46 toks/s, output: 2660.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:19,  2.41it/s, est. speed input: 252.10 toks/s, output: 2739.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:27<00:17,  2.67it/s, est. speed input: 251.80 toks/s, output: 2830.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:29<00:40,  1.15it/s, est. speed input: 236.17 toks/s, output: 2750.12 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:18:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6247.4 tokens/s, Running: 45 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.4%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:30<00:29,  1.50it/s, est. speed input: 233.39 toks/s, output: 2905.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:32<00:43,  1.02s/it, est. speed input: 220.26 toks/s, output: 2837.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:32<00:35,  1.19it/s, est. speed input: 219.82 toks/s, output: 2927.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:33<00:30,  1.32it/s, est. speed input: 218.09 toks/s, output: 2999.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:33<00:31,  1.27it/s, est. speed input: 214.32 toks/s, output: 3043.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:33<00:23,  1.69it/s, est. speed input: 215.34 toks/s, output: 3153.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:36<00:30,  1.20it/s, est. speed input: 205.20 toks/s, output: 3189.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:36<00:26,  1.34it/s, est. speed input: 204.20 toks/s, output: 3269.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:36<00:22,  1.59it/s, est. speed input: 204.14 toks/s, output: 3364.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:38<00:24,  1.36it/s, est. speed input: 200.26 toks/s, output: 3394.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:38<00:18,  1.75it/s, est. speed input: 201.06 toks/s, output: 3502.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:39<00:23,  1.33it/s, est. speed input: 196.44 toks/s, output: 3515.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:39<00:13,  2.21it/s, est. speed input: 198.59 toks/s, output: 3739.16 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:19:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4787.8 tokens/s, Running: 30 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:40<00:14,  2.06it/s, est. speed input: 197.12 toks/s, output: 3804.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:40<00:15,  1.78it/s, est. speed input: 194.89 toks/s, output: 3852.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:41<00:12,  2.17it/s, est. speed input: 195.42 toks/s, output: 3955.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:41<00:09,  2.64it/s, est. speed input: 196.11 toks/s, output: 4060.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:41<00:08,  2.88it/s, est. speed input: 196.31 toks/s, output: 4155.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:42<00:11,  2.05it/s, est. speed input: 193.88 toks/s, output: 4195.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:42<00:10,  2.16it/s, est. speed input: 193.42 toks/s, output: 4277.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:43<00:14,  1.54it/s, est. speed input: 189.94 toks/s, output: 4294.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:43<00:10,  2.04it/s, est. speed input: 190.87 toks/s, output: 4405.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:44<00:11,  1.76it/s, est. speed input: 189.00 toks/s, output: 4454.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:45<00:11,  1.63it/s, est. speed input: 187.38 toks/s, output: 4507.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:47<00:16,  1.10it/s, est. speed input: 182.29 toks/s, output: 4480.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:47<00:15,  1.12it/s, est. speed input: 180.21 toks/s, output: 4525.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:48<00:10,  1.50it/s, est. speed input: 180.93 toks/s, output: 4639.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:13,  1.14it/s, est. speed input: 177.11 toks/s, output: 4637.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:19:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3298.7 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:13,  1.02it/s, est. speed input: 173.96 toks/s, output: 4652.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:51<00:12,  1.04it/s, est. speed input: 172.06 toks/s, output: 4700.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:52<00:07,  1.39it/s, est. speed input: 171.56 toks/s, output: 4879.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:54<00:09,  1.02it/s, est. speed input: 167.01 toks/s, output: 4852.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:54<00:07,  1.15it/s, est. speed input: 166.40 toks/s, output: 4936.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:55<00:07,  1.06it/s, est. speed input: 164.13 toks/s, output: 4971.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:57<00:07,  1.00s/it, est. speed input: 161.83 toks/s, output: 5006.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:57<00:05,  1.03it/s, est. speed input: 160.40 toks/s, output: 5067.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:58<00:04,  1.02it/s, est. speed input: 158.63 toks/s, output: 5119.19 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.02it/s, est. speed input: 163.69 toks/s, output: 5814.10 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.71it/s, est. speed input: 163.69 toks/s, output: 5814.10 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:19:24 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:19:24 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:19:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:19:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:19:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:19:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:19:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 13:19:24 [abstract.py:306] It took 0.554033 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:19:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:19:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:19:24 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0113
+  • Training Accuracy: 0.3187
+  • Mean Generation Length: 2141.8188
+  • Training FLOPS: 42.85 TFLOPS (10.71 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 279.40s
+  • policy_training: 180.55s (64.6%)
+  • generation: 62.80s (22.5%)
+  • teacher_logprob_inference: 30.49s (10.9%)
+  • policy_training/sharding_data: 4.34s (1.6%)
+  • prepare_for_generation: 3.70s (1.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.6%)
+  • policy_training/submit_training_futures: 1.54s (0.6%)
+  • training_prep: 0.97s (0.3%)
+  • teacher_logprob_inference_prep: 0.63s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 178: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:22:58 [abstract.py:324] It took 0.050060 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09244414e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68466949e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83284912e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09244414e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68466949e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83284912e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09244414e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68466949e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83284912e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09244414e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68466949e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83284912e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259688e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70448914e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82962189e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66891479e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.10087729e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #178] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43915787e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259688e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70448914e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82962189e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65196228e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36581821e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #178] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.43300018e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259688e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70448914e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82962189e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37607880e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.42257214e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #178] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.46038437e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09259688e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70448914e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82962189e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.15840454e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62740631e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #178] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35346329e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 178: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:01 [abstract.py:324] It took 0.028651 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13925.89it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:01 [loggers.py:236] Engine 000: Avg prompt throughput: 42.6 tokens/s, Avg generation throughput: 77.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:47,  3.36it/s, est. speed input: 201.70 toks/s, output: 43.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:08, 18.13it/s, est. speed input: 904.24 toks/s, output: 272.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:05, 27.95it/s, est. speed input: 1326.18 toks/s, output: 475.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:03, 40.58it/s, est. speed input: 1813.72 toks/s, output: 844.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 40.61it/s, est. speed input: 1950.49 toks/s, output: 1075.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:03, 42.86it/s, est. speed input: 2090.08 toks/s, output: 1369.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:00<00:01, 63.06it/s, est. speed input: 2628.60 toks/s, output: 2179.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:01, 57.08it/s, est. speed input: 2661.34 toks/s, output: 2445.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 50.28it/s, est. speed input: 2623.54 toks/s, output: 2708.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 49.65it/s, est. speed input: 2650.08 toks/s, output: 3005.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:02, 45.24it/s, est. speed input: 2606.81 toks/s, output: 3243.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 43.63it/s, est. speed input: 2591.41 toks/s, output: 3513.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:01<00:02, 40.08it/s, est. speed input: 2542.44 toks/s, output: 3710.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:02<00:02, 25.54it/s, est. speed input: 2241.67 toks/s, output: 3559.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:03<00:08,  8.79it/s, est. speed input: 1443.08 toks/s, output: 2571.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:04<00:10,  6.42it/s, est. speed input: 1184.44 toks/s, output: 2374.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:08<00:27,  2.47it/s, est. speed input: 700.71 toks/s, output: 1592.17 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:08<00:23,  2.76it/s, est. speed input: 686.04 toks/s, output: 1771.74 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9761.0 tokens/s, Running: 64 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:10<00:32,  1.94it/s, est. speed input: 556.75 toks/s, output: 1639.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:11<00:32,  1.93it/s, est. speed input: 535.87 toks/s, output: 1687.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:12<00:35,  1.71it/s, est. speed input: 499.50 toks/s, output: 1683.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:12<00:35,  1.68it/s, est. speed input: 479.02 toks/s, output: 1724.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:13<00:32,  1.81it/s, est. speed input: 469.54 toks/s, output: 1800.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:13<00:34,  1.67it/s, est. speed input: 448.46 toks/s, output: 1829.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:14<00:30,  1.84it/s, est. speed input: 440.72 toks/s, output: 1908.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:14<00:32,  1.74it/s, est. speed input: 425.30 toks/s, output: 1950.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:15<00:30,  1.83it/s, est. speed input: 416.39 toks/s, output: 2018.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:15<00:25,  2.13it/s, est. speed input: 413.22 toks/s, output: 2110.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:16<00:35,  1.49it/s, est. speed input: 388.01 toks/s, output: 2090.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:17<00:29,  1.76it/s, est. speed input: 384.44 toks/s, output: 2180.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:17<00:27,  1.87it/s, est. speed input: 377.91 toks/s, output: 2251.39 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7328.5 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:21<01:19,  1.60s/it, est. speed input: 308.91 toks/s, output: 1951.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:22<01:00,  1.24s/it, est. speed input: 306.44 toks/s, output: 2045.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:22<00:45,  1.05it/s, est. speed input: 305.56 toks/s, output: 2149.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:34,  1.34it/s, est. speed input: 304.64 toks/s, output: 2251.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:22<00:16,  2.70it/s, est. speed input: 309.42 toks/s, output: 2608.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:23<00:18,  2.29it/s, est. speed input: 303.03 toks/s, output: 2659.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:23<00:12,  3.22it/s, est. speed input: 305.72 toks/s, output: 2891.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:24<00:14,  2.78it/s, est. speed input: 299.45 toks/s, output: 3036.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:24<00:11,  3.24it/s, est. speed input: 300.61 toks/s, output: 3152.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:26<00:23,  1.59it/s, est. speed input: 283.03 toks/s, output: 3073.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:26<00:18,  1.97it/s, est. speed input: 284.07 toks/s, output: 3187.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:26<00:14,  2.48it/s, est. speed input: 285.31 toks/s, output: 3304.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:27<00:12,  2.78it/s, est. speed input: 285.01 toks/s, output: 3404.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:27<00:11,  2.89it/s, est. speed input: 282.61 toks/s, output: 3579.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:28<00:10,  2.94it/s, est. speed input: 281.48 toks/s, output: 3667.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:28<00:08,  3.40it/s, est. speed input: 281.28 toks/s, output: 3867.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:28<00:07,  3.62it/s, est. speed input: 281.23 toks/s, output: 3969.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:29<00:08,  3.19it/s, est. speed input: 279.09 toks/s, output: 4040.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:29<00:05,  4.49it/s, est. speed input: 281.48 toks/s, output: 4374.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:29<00:05,  4.06it/s, est. speed input: 280.22 toks/s, output: 4455.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5352.1 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:31<00:11,  1.93it/s, est. speed input: 268.89 toks/s, output: 4377.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:31<00:09,  2.26it/s, est. speed input: 269.02 toks/s, output: 4482.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:33<00:14,  1.34it/s, est. speed input: 257.66 toks/s, output: 4396.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:33<00:12,  1.52it/s, est. speed input: 256.23 toks/s, output: 4476.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:34<00:08,  2.03it/s, est. speed input: 255.61 toks/s, output: 4672.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:34<00:08,  1.83it/s, est. speed input: 251.99 toks/s, output: 4711.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:36<00:11,  1.32it/s, est. speed input: 244.04 toks/s, output: 4670.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:37<00:12,  1.13it/s, est. speed input: 237.57 toks/s, output: 4656.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:37<00:09,  1.40it/s, est. speed input: 237.49 toks/s, output: 4764.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:38<00:10,  1.17it/s, est. speed input: 231.69 toks/s, output: 4758.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:39<00:07,  1.44it/s, est. speed input: 231.44 toks/s, output: 4863.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2997.8 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:41<00:10,  1.04s/it, est. speed input: 222.34 toks/s, output: 4785.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:41<00:05,  1.60it/s, est. speed input: 224.01 toks/s, output: 5046.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:43<00:07,  1.04s/it, est. speed input: 213.54 toks/s, output: 4928.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:44<00:05,  1.04it/s, est. speed input: 211.31 toks/s, output: 4996.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:44<00:03,  1.29it/s, est. speed input: 211.45 toks/s, output: 5119.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:47<00:05,  1.30s/it, est. speed input: 200.83 toks/s, output: 4986.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:48<00:03,  1.19s/it, est. speed input: 198.30 toks/s, output: 5049.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1383.0 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 74.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:50<00:02,  1.49s/it, est. speed input: 190.88 toks/s, output: 4990.84 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.49s/it, est. speed input: 193.33 toks/s, output: 5314.84 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.16it/s, est. speed input: 193.33 toks/s, output: 5314.84 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:54 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:54 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:23:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:23:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:23:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:23:54 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:23:55 [abstract.py:306] It took 0.520513 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:23:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:23:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:23:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:23:55 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.4250
+  • Mean Generation Length: 1679.6750
+  • Training FLOPS: 39.67 TFLOPS (9.92 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 237.29s
+  • policy_training: 151.70s (63.9%)
+  • generation: 53.56s (22.6%)
+  • teacher_logprob_inference: 26.28s (11.1%)
+  • policy_training/sharding_data: 4.33s (1.8%)
+  • prepare_for_generation: 3.64s (1.5%)
+  • policy_training/submit_training_futures: 2.03s (0.9%)
+  • prepare_for_generation/transfer_and_update_weights: 1.64s (0.7%)
+  • training_prep: 1.51s (0.6%)
+  • teacher_logprob_inference_prep: 0.40s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 179: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:26:55 [abstract.py:324] It took 0.052937 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09241621e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68966827e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82670135e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09241621e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68966827e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82670135e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09241621e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68966827e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82670135e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09241621e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68966827e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82670135e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09246484e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.74117737e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82566147e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09246484e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.74117737e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82566147e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09246484e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.74117737e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82566147e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09246484e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.74117737e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82566147e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.92616653e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.07576180e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #179] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43855469e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66599197e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37574682e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #179] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44321289e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38268890e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.65883064e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #179] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.45341377e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.14737396e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62180138e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #179] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35434738e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 179: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:26:58 [abstract.py:324] It took 0.034643 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13970.83it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:26:59 [loggers.py:236] Engine 000: Avg prompt throughput: 51.8 tokens/s, Avg generation throughput: 2.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:02,  2.53it/s, est. speed input: 149.18 toks/s, output: 30.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|▏         | 2/160 [00:00<00:37,  4.22it/s, est. speed input: 233.83 toks/s, output: 69.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:10, 14.47it/s, est. speed input: 638.21 toks/s, output: 275.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:09, 15.92it/s, est. speed input: 751.76 toks/s, output: 409.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:01<00:05, 24.71it/s, est. speed input: 1040.04 toks/s, output: 762.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:03, 35.05it/s, est. speed input: 1357.33 toks/s, output: 1223.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 36.36it/s, est. speed input: 1465.58 toks/s, output: 1468.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 34.94it/s, est. speed input: 1518.11 toks/s, output: 1684.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 48.97it/s, est. speed input: 1810.47 toks/s, output: 2345.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 46.14it/s, est. speed input: 1868.67 toks/s, output: 2622.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 51.43it/s, est. speed input: 2002.66 toks/s, output: 3059.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 56.93it/s, est. speed input: 2167.24 toks/s, output: 3653.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:02<00:02, 34.11it/s, est. speed input: 1964.82 toks/s, output: 3567.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:04<00:10,  7.55it/s, est. speed input: 1034.58 toks/s, output: 2130.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:07<00:18,  4.24it/s, est. speed input: 707.48 toks/s, output: 1786.46 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:08<00:21,  3.40it/s, est. speed input: 594.73 toks/s, output: 1774.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10361.6 tokens/s, Running: 72 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:10<00:25,  2.80it/s, est. speed input: 523.38 toks/s, output: 1750.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:13<00:41,  1.70it/s, est. speed input: 403.52 toks/s, output: 1545.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:14<00:46,  1.48it/s, est. speed input: 370.92 toks/s, output: 1520.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<00:44,  1.54it/s, est. speed input: 362.78 toks/s, output: 1587.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:16<00:50,  1.33it/s, est. speed input: 338.61 toks/s, output: 1580.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:16<00:42,  1.56it/s, est. speed input: 339.09 toks/s, output: 1681.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:43,  1.51it/s, est. speed input: 328.08 toks/s, output: 1725.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:18<00:38,  1.65it/s, est. speed input: 323.78 toks/s, output: 1800.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<00:25,  2.42it/s, est. speed input: 326.22 toks/s, output: 2006.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:34,  1.76it/s, est. speed input: 310.69 toks/s, output: 2007.25 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7914.0 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:37,  1.58it/s, est. speed input: 301.04 toks/s, output: 2041.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:20<00:29,  2.01it/s, est. speed input: 302.42 toks/s, output: 2147.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:24,  2.39it/s, est. speed input: 302.37 toks/s, output: 2242.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:21,  2.65it/s, est. speed input: 301.20 toks/s, output: 2329.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:21<00:27,  2.04it/s, est. speed input: 293.20 toks/s, output: 2363.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:23<00:32,  1.66it/s, est. speed input: 279.44 toks/s, output: 2436.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:20,  2.54it/s, est. speed input: 283.05 toks/s, output: 2656.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:24<00:30,  1.68it/s, est. speed input: 270.38 toks/s, output: 2631.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:24<00:19,  2.56it/s, est. speed input: 273.96 toks/s, output: 2852.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:25<00:19,  2.51it/s, est. speed input: 271.64 toks/s, output: 2921.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:26<00:31,  1.48it/s, est. speed input: 257.81 toks/s, output: 2866.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:27<00:27,  1.68it/s, est. speed input: 256.70 toks/s, output: 2948.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:27<00:23,  1.88it/s, est. speed input: 255.54 toks/s, output: 3027.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:29<00:48,  1.09s/it, est. speed input: 235.62 toks/s, output: 2886.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:29 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6214.5 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.3%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:30<00:37,  1.15it/s, est. speed input: 235.30 toks/s, output: 2977.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:30<00:32,  1.29it/s, est. speed input: 233.09 toks/s, output: 3043.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:31<00:18,  2.15it/s, est. speed input: 235.76 toks/s, output: 3265.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:31<00:19,  2.05it/s, est. speed input: 233.53 toks/s, output: 3326.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:32<00:19,  1.96it/s, est. speed input: 231.28 toks/s, output: 3387.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:32<00:18,  2.03it/s, est. speed input: 229.87 toks/s, output: 3460.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:15,  2.26it/s, est. speed input: 229.46 toks/s, output: 3547.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:33<00:16,  2.15it/s, est. speed input: 227.68 toks/s, output: 3612.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:34<00:17,  1.94it/s, est. speed input: 225.21 toks/s, output: 3665.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:34<00:17,  1.92it/s, est. speed input: 223.50 toks/s, output: 3729.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:13,  2.45it/s, est. speed input: 224.31 toks/s, output: 3835.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:11,  2.68it/s, est. speed input: 224.10 toks/s, output: 3924.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:11,  2.51it/s, est. speed input: 222.96 toks/s, output: 3995.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:36<00:12,  2.25it/s, est. speed input: 221.26 toks/s, output: 4056.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:36<00:13,  2.13it/s, est. speed input: 219.67 toks/s, output: 4119.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:39<00:23,  1.10it/s, est. speed input: 206.92 toks/s, output: 4061.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:39<00:19,  1.31it/s, est. speed input: 206.71 toks/s, output: 4152.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:40<00:09,  2.37it/s, est. speed input: 209.66 toks/s, output: 4493.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4553.6 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:40<00:07,  2.70it/s, est. speed input: 209.85 toks/s, output: 4683.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:41<00:06,  2.69it/s, est. speed input: 208.97 toks/s, output: 4847.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:08,  2.12it/s, est. speed input: 205.87 toks/s, output: 4868.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:07,  2.01it/s, est. speed input: 204.43 toks/s, output: 4928.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:43<00:07,  2.06it/s, est. speed input: 203.72 toks/s, output: 5004.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:44<00:08,  1.67it/s, est. speed input: 200.79 toks/s, output: 5026.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:45<00:10,  1.24it/s, est. speed input: 195.99 toks/s, output: 5002.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:48<00:15,  1.29s/it, est. speed input: 186.84 toks/s, output: 4869.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:48<00:11,  1.01s/it, est. speed input: 186.94 toks/s, output: 4973.34 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2558.7 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:51<00:16,  1.63s/it, est. speed input: 176.60 toks/s, output: 4805.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:53<00:15,  1.67s/it, est. speed input: 171.82 toks/s, output: 4784.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:53<00:09,  1.22s/it, est. speed input: 172.55 toks/s, output: 4914.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:53<00:06,  1.09it/s, est. speed input: 172.97 toks/s, output: 5034.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:56<00:09,  1.55s/it, est. speed input: 164.81 toks/s, output: 4909.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.55s/it, est. speed input: 171.06 toks/s, output: 5775.47 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.82it/s, est. speed input: 171.06 toks/s, output: 5775.47 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:58 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:58 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:27:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:27:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:27:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:27:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:27:59 [abstract.py:306] It took 0.533965 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:27:59 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:27:59 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:27:59 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:27:59 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0117
+  • Training Accuracy: 0.4375
+  • Mean Generation Length: 2049.3625
+  • Training FLOPS: 42.78 TFLOPS (10.70 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 269.05s
+  • policy_training: 172.33s (64.1%)
+  • generation: 60.32s (22.4%)
+  • teacher_logprob_inference: 31.13s (11.6%)
+  • policy_training/sharding_data: 4.31s (1.6%)
+  • prepare_for_generation: 3.72s (1.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.6%)
+  • policy_training/submit_training_futures: 1.61s (0.6%)
+  • training_prep: 0.83s (0.3%)
+  • teacher_logprob_inference_prep: 0.52s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 180: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:31:25 [abstract.py:324] It took 0.051290 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09247969e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69158936e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82432251e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09247969e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69158936e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82432251e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09247969e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69158936e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82432251e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09247969e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69158936e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82432251e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245195e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70717773e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81270599e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00775928e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.87615967e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #180] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44286118e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245195e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70717773e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81270599e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66416092e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38954201e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #180] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.44269714e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245195e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70717773e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81270599e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39240112e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.96242332e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #180] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.46587181e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245195e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70717773e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81270599e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17376938e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62880001e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #180] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.35835205e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 180: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:31:28 [abstract.py:324] It took 0.033841 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 13:31:28 [loggers.py:236] Engine 000: Avg prompt throughput: 44.4 tokens/s, Avg generation throughput: 53.8 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12676.64it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:19,  8.21it/s, est. speed input: 484.45 toks/s, output: 73.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:03, 45.92it/s, est. speed input: 2424.90 toks/s, output: 588.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:04, 31.15it/s, est. speed input: 1896.08 toks/s, output: 638.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 28.75it/s, est. speed input: 1805.34 toks/s, output: 804.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 33.43it/s, est. speed input: 1960.24 toks/s, output: 1125.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:03, 41.23it/s, est. speed input: 2195.76 toks/s, output: 1604.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 42.02it/s, est. speed input: 2250.68 toks/s, output: 1884.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 57.35it/s, est. speed input: 2614.77 toks/s, output: 2641.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:01, 57.65it/s, est. speed input: 2712.94 toks/s, output: 3020.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  37%|███▋      | 59/160 [00:01<00:01, 59.40it/s, est. speed input: 2810.99 toks/s, output: 3429.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:02, 42.43it/s, est. speed input: 2595.36 toks/s, output: 3486.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:31:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10769.7 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:11<00:41,  2.14it/s, est. speed input: 386.04 toks/s, output: 717.57 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:16<00:54,  1.53it/s, est. speed input: 274.89 toks/s, output: 863.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:31:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8777.1 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:21<01:08,  1.19it/s, est. speed input: 219.62 toks/s, output: 964.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:22<00:58,  1.34it/s, est. speed input: 221.32 toks/s, output: 1160.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:26<01:14,  1.03it/s, est. speed input: 191.98 toks/s, output: 1188.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:26<01:08,  1.12it/s, est. speed input: 191.86 toks/s, output: 1279.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:26<00:59,  1.26it/s, est. speed input: 193.14 toks/s, output: 1377.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:27<00:57,  1.29it/s, est. speed input: 190.73 toks/s, output: 1449.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:27<00:48,  1.50it/s, est. speed input: 191.49 toks/s, output: 1543.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:28<00:36,  1.93it/s, est. speed input: 192.48 toks/s, output: 1724.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:28<00:36,  1.93it/s, est. speed input: 191.13 toks/s, output: 1798.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:29<00:36,  1.88it/s, est. speed input: 189.41 toks/s, output: 1867.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:31:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7548.6 tokens/s, Running: 69 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.0%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:30<00:42,  1.61it/s, est. speed input: 185.73 toks/s, output: 1915.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:32<01:08,  1.02s/it, est. speed input: 175.16 toks/s, output: 1891.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:32<00:57,  1.16it/s, est. speed input: 174.62 toks/s, output: 1968.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:35<01:04,  1.01s/it, est. speed input: 166.18 toks/s, output: 2033.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:36<00:52,  1.19it/s, est. speed input: 164.39 toks/s, output: 2174.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:36<00:48,  1.25it/s, est. speed input: 163.03 toks/s, output: 2239.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:37<00:44,  1.36it/s, est. speed input: 162.38 toks/s, output: 2312.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:39<00:55,  1.07it/s, est. speed input: 157.57 toks/s, output: 2325.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:39<00:43,  1.32it/s, est. speed input: 158.08 toks/s, output: 2414.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:39<00:31,  1.77it/s, est. speed input: 158.56 toks/s, output: 2582.85 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:32:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6474.5 tokens/s, Running: 55 reqs, Waiting: 0 reqs, GPU KV cache usage: 10.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:40<00:21,  2.50it/s, est. speed input: 160.54 toks/s, output: 2773.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:40<00:18,  2.87it/s, est. speed input: 161.44 toks/s, output: 2866.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:40<00:18,  2.74it/s, est. speed input: 161.27 toks/s, output: 2940.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:40<00:16,  3.12it/s, est. speed input: 162.07 toks/s, output: 3031.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:41<00:15,  3.08it/s, est. speed input: 162.31 toks/s, output: 3189.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:41<00:14,  3.33it/s, est. speed input: 162.90 toks/s, output: 3276.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:42<00:13,  3.45it/s, est. speed input: 163.33 toks/s, output: 3359.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:43<00:27,  1.70it/s, est. speed input: 159.35 toks/s, output: 3352.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:43<00:16,  2.66it/s, est. speed input: 161.49 toks/s, output: 3547.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:44<00:15,  2.69it/s, est. speed input: 161.61 toks/s, output: 3622.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:44<00:18,  2.31it/s, est. speed input: 160.73 toks/s, output: 3677.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:45<00:18,  2.20it/s, est. speed input: 160.17 toks/s, output: 3739.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:46<00:27,  1.45it/s, est. speed input: 156.91 toks/s, output: 3739.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:46<00:20,  1.89it/s, est. speed input: 157.83 toks/s, output: 3835.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:47<00:28,  1.32it/s, est. speed input: 154.75 toks/s, output: 3835.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:49<00:34,  1.08it/s, est. speed input: 151.75 toks/s, output: 3837.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:49<00:20,  1.73it/s, est. speed input: 153.23 toks/s, output: 4024.08 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:32:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4972.3 tokens/s, Running: 35 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:50<00:21,  1.61it/s, est. speed input: 152.14 toks/s, output: 4070.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:51<00:21,  1.57it/s, est. speed input: 151.30 toks/s, output: 4122.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:51<00:17,  1.86it/s, est. speed input: 151.71 toks/s, output: 4207.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:52<00:21,  1.43it/s, est. speed input: 149.58 toks/s, output: 4225.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:53<00:18,  1.54it/s, est. speed input: 148.58 toks/s, output: 4346.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:55<00:28,  1.03s/it, est. speed input: 143.71 toks/s, output: 4281.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:56<00:26,  1.03it/s, est. speed input: 142.75 toks/s, output: 4330.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:57<00:23,  1.09it/s, est. speed input: 141.87 toks/s, output: 4382.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:58<00:21,  1.18it/s, est. speed input: 141.30 toks/s, output: 4442.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:58<00:20,  1.18it/s, est. speed input: 140.34 toks/s, output: 4490.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:59<00:19,  1.15it/s, est. speed input: 139.19 toks/s, output: 4531.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:59<00:14,  1.54it/s, est. speed input: 139.97 toks/s, output: 4633.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:32:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3788.5 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [01:00<00:09,  2.11it/s, est. speed input: 140.74 toks/s, output: 4813.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [01:00<00:09,  2.11it/s, est. speed input: 140.65 toks/s, output: 4887.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [01:01<00:08,  2.16it/s, est. speed input: 140.66 toks/s, output: 4964.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [01:01<00:07,  2.30it/s, est. speed input: 140.84 toks/s, output: 5048.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [01:02<00:06,  2.36it/s, est. speed input: 140.91 toks/s, output: 5127.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:03<00:10,  1.44it/s, est. speed input: 138.81 toks/s, output: 5131.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:04<00:09,  1.53it/s, est. speed input: 138.57 toks/s, output: 5200.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:06<00:16,  1.25s/it, est. speed input: 133.84 toks/s, output: 5106.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:07<00:14,  1.21s/it, est. speed input: 132.52 toks/s, output: 5139.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:09<00:13,  1.22s/it, est. speed input: 131.01 toks/s, output: 5165.12 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  1.22s/it, est. speed input: 140.69 toks/s, output: 6468.32 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:09<00:00,  2.31it/s, est. speed input: 140.69 toks/s, output: 6468.32 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:32:41 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:32:41 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:32:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:32:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:32:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:32:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:32:42 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:32:42 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:32:42 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 13:32:42 [abstract.py:306] It took 0.645491 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:32:42 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 13:36:53 [abstract.py:324] It took 0.047878 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277500e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68929901e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82529907e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277500e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68929901e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82529907e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277500e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68929901e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82529907e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277500e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68929901e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82529907e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286699e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68804855e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82302628e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00406704e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.82736015e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #181] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44548920e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286699e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68804855e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82302628e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64676056e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39290695e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #181] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.45884705e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286699e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68804855e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82302628e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39795074e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.85725212e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #181] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.47507248e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09286699e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68804855e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82302628e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17664108e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.63210907e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #181] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36240677e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+▶ Starting validation at step 180...
+(VllmGenerationWorker pid=10140) INFO 05-03 13:36:56 [abstract.py:324] It took 0.035229 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 10164.93it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:36:56 [loggers.py:236] Engine 000: Avg prompt throughput: 7.9 tokens/s, Avg generation throughput: 86.2 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:07,  4.33it/s, est. speed input: 281.71 toks/s, output: 73.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:03,  8.23it/s, est. speed input: 503.44 toks/s, output: 259.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02, 10.97it/s, est. speed input: 637.56 toks/s, output: 449.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 16.42it/s, est. speed input: 854.57 toks/s, output: 793.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 16.93it/s, est. speed input: 894.59 toks/s, output: 1002.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:01, 16.72it/s, est. speed input: 921.62 toks/s, output: 1183.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:01<00:01, 11.82it/s, est. speed input: 814.77 toks/s, output: 1220.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:04<00:08,  1.80it/s, est. speed input: 243.84 toks/s, output: 676.43 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:04<00:06,  2.16it/s, est. speed input: 245.61 toks/s, output: 1028.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:05<00:05,  2.20it/s, est. speed input: 239.15 toks/s, output: 1161.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:37:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3473.9 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:10<00:16,  1.40s/it, est. speed input: 128.41 toks/s, output: 809.53 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:11<00:14,  1.28s/it, est. speed input: 124.64 toks/s, output: 964.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:13<00:14,  1.44s/it, est. speed input: 111.15 toks/s, output: 1039.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:14<00:12,  1.43s/it, est. speed input: 105.05 toks/s, output: 1159.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:14<00:08,  1.09s/it, est. speed input: 108.20 toks/s, output: 1364.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:16<00:09,  1.42s/it, est. speed input: 97.48 toks/s, output: 1403.11 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:17<00:05,  1.01s/it, est. speed input: 99.47 toks/s, output: 1756.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:20<00:05,  1.26s/it, est. speed input: 92.56 toks/s, output: 1802.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:37:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1944.7 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.8%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:20<00:03,  1.14s/it, est. speed input: 92.08 toks/s, output: 1960.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:21<00:01,  1.06it/s, est. speed input: 93.40 toks/s, output: 2149.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:24<00:01,  1.51s/it, est. speed input: 84.45 toks/s, output: 2116.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:27<00:00,  1.90s/it, est. speed input: 77.87 toks/s, output: 2127.03 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:27<00:00,  1.90s/it, est. speed input: 77.87 toks/s, output: 2127.03 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:27<00:00,  1.18it/s, est. speed input: 77.87 toks/s, output: 2127.03 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12309.04it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:06,  4.53it/s, est. speed input: 303.54 toks/s, output: 149.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:02,  9.79it/s, est. speed input: 596.63 toks/s, output: 412.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 23.42it/s, est. speed input: 1207.44 toks/s, output: 1051.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:00<00:00, 27.66it/s, est. speed input: 1468.97 toks/s, output: 1560.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:37:26 [loggers.py:236] Engine 000: Avg prompt throughput: 213.2 tokens/s, Avg generation throughput: 1367.9 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:06<00:08,  1.82it/s, est. speed input: 179.88 toks/s, output: 439.38 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:08<00:08,  1.56it/s, est. speed input: 153.01 toks/s, output: 730.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:10<00:08,  1.26it/s, est. speed input: 128.97 toks/s, output: 963.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:12<00:08,  1.11it/s, est. speed input: 118.33 toks/s, output: 1064.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:37:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3063.5 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:13<00:08,  1.08it/s, est. speed input: 113.74 toks/s, output: 1200.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:15<00:09,  1.17s/it, est. speed input: 102.34 toks/s, output: 1256.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:16<00:06,  1.01it/s, est. speed input: 104.38 toks/s, output: 1452.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:16<00:04,  1.25it/s, est. speed input: 107.27 toks/s, output: 1656.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:16<00:03,  1.50it/s, est. speed input: 109.21 toks/s, output: 1851.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:17<00:03,  1.14it/s, est. speed input: 104.27 toks/s, output: 1926.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:18<00:02,  1.34it/s, est. speed input: 105.51 toks/s, output: 2108.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:18<00:01,  1.66it/s, est. speed input: 107.74 toks/s, output: 2306.83 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:37:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1248.3 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.5%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:22<00:01,  1.69s/it, est. speed input: 89.95 toks/s, output: 2099.99 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:30<00:00,  3.45s/it, est. speed input: 69.44 toks/s, output: 1816.48 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:30<00:00,  3.45s/it, est. speed input: 69.44 toks/s, output: 1816.48 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:30<00:00,  1.04it/s, est. speed input: 69.44 toks/s, output: 1816.48 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11519.85it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  7.93it/s, est. speed input: 539.60 toks/s, output: 182.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:02, 13.47it/s, est. speed input: 860.75 toks/s, output: 436.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 14.96it/s, est. speed input: 949.77 toks/s, output: 639.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 20.29it/s, est. speed input: 1185.50 toks/s, output: 1050.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:00, 25.34it/s, est. speed input: 1403.39 toks/s, output: 1536.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:00<00:00, 23.94it/s, est. speed input: 1410.41 toks/s, output: 1824.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:37:56 [loggers.py:236] Engine 000: Avg prompt throughput: 212.6 tokens/s, Avg generation throughput: 828.9 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.2%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:02<00:02,  5.38it/s, est. speed input: 563.12 toks/s, output: 1014.17 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:02<00:02,  4.32it/s, est. speed input: 462.78 toks/s, output: 1191.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:09<00:09,  1.01it/s, est. speed input: 158.27 toks/s, output: 769.84 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:10<00:09,  1.09s/it, est. speed input: 140.91 toks/s, output: 885.21 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:38:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2713.0 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:12<00:09,  1.13s/it, est. speed input: 130.90 toks/s, output: 1019.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:13<00:08,  1.22s/it, est. speed input: 120.72 toks/s, output: 1134.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:18<00:11,  1.90s/it, est. speed input: 96.07 toks/s, output: 1102.61 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:19<00:09,  1.81s/it, est. speed input: 91.91 toks/s, output: 1249.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:20<00:05,  1.46s/it, est. speed input: 93.07 toks/s, output: 1454.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:20<00:03,  1.30s/it, est. speed input: 92.27 toks/s, output: 1628.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:21<00:02,  1.02s/it, est. speed input: 94.18 toks/s, output: 1842.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:38:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1600.2 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.5%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:23<00:01,  1.43s/it, est. speed input: 87.18 toks/s, output: 1889.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:26<00:00,  1.91s/it, est. speed input: 79.56 toks/s, output: 1915.03 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:26<00:00,  1.91s/it, est. speed input: 79.56 toks/s, output: 1915.03 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:26<00:00,  1.20it/s, est. speed input: 79.56 toks/s, output: 1915.03 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12140.91it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:02, 14.29it/s, est. speed input: 928.85 toks/s, output: 292.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 19.75it/s, est. speed input: 1242.76 toks/s, output: 615.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 13.56it/s, est. speed input: 951.26 toks/s, output: 714.32 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 12.69it/s, est. speed input: 896.44 toks/s, output: 912.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:01<00:01, 10.33it/s, est. speed input: 783.04 toks/s, output: 1037.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:38:26 [loggers.py:236] Engine 000: Avg prompt throughput: 209.6 tokens/s, Avg generation throughput: 2046.7 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.8%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:07<00:19,  1.10s/it, est. speed input: 119.90 toks/s, output: 427.92 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:09<00:21,  1.24s/it, est. speed input: 103.31 toks/s, output: 555.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:10<00:17,  1.10s/it, est. speed input: 104.02 toks/s, output: 734.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:10<00:15,  1.01s/it, est. speed input: 103.18 toks/s, output: 897.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:11<00:09,  1.36it/s, est. speed input: 109.02 toks/s, output: 1257.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:12<00:10,  1.09it/s, est. speed input: 101.12 toks/s, output: 1318.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:38:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3445.2 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:15<00:15,  1.41s/it, est. speed input: 86.45 toks/s, output: 1282.66 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:16<00:11,  1.18s/it, est. speed input: 87.85 toks/s, output: 1455.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:17<00:10,  1.17s/it, est. speed input: 85.73 toks/s, output: 1573.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:38:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2131.8 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:26<00:26,  3.30s/it, est. speed input: 59.45 toks/s, output: 1260.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:27<00:17,  2.51s/it, est. speed input: 60.73 toks/s, output: 1450.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:27<00:11,  1.87s/it, est. speed input: 62.48 toks/s, output: 1648.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:28<00:08,  1.62s/it, est. speed input: 62.57 toks/s, output: 1804.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:30<00:07,  1.93s/it, est. speed input: 59.41 toks/s, output: 1866.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:38:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1275.0 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:34<00:03,  1.90s/it, est. speed input: 56.78 toks/s, output: 2078.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:36<00:01,  1.98s/it, est. speed input: 55.04 toks/s, output: 2174.24 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.98s/it, est. speed input: 56.74 toks/s, output: 2395.89 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.15s/it, est. speed input: 56.74 toks/s, output: 2395.89 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11480.43it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  7.91it/s, est. speed input: 514.51 toks/s, output: 182.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 17.26it/s, est. speed input: 1046.54 toks/s, output: 511.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:02, 11.10it/s, est. speed input: 770.08 toks/s, output: 587.73 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:02, 10.78it/s, est. speed input: 741.07 toks/s, output: 786.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:02<00:08,  2.66it/s, est. speed input: 263.43 toks/s, output: 575.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:02<00:06,  3.00it/s, est. speed input: 271.93 toks/s, output: 748.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:02<00:06,  3.08it/s, est. speed input: 266.35 toks/s, output: 882.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:39:06 [loggers.py:236] Engine 000: Avg prompt throughput: 209.8 tokens/s, Avg generation throughput: 2927.6 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:14<00:56,  2.98s/it, est. speed input: 59.76 toks/s, output: 385.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:14<00:40,  2.25s/it, est. speed input: 63.96 toks/s, output: 583.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:14<00:29,  1.73s/it, est. speed input: 66.94 toks/s, output: 773.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:15<00:22,  1.42s/it, est. speed input: 68.64 toks/s, output: 943.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:15<00:16,  1.07s/it, est. speed input: 72.12 toks/s, output: 1133.17 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:39:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3812.9 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:17<00:16,  1.21s/it, est. speed input: 69.35 toks/s, output: 1230.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:18<00:17,  1.33s/it, est. speed input: 66.83 toks/s, output: 1324.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:18<00:12,  1.03s/it, est. speed input: 69.39 toks/s, output: 1505.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:24<00:26,  2.43s/it, est. speed input: 55.78 toks/s, output: 1356.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:24<00:12,  1.36s/it, est. speed input: 60.71 toks/s, output: 1750.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:39:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2525.7 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:28<00:15,  1.98s/it, est. speed input: 54.76 toks/s, output: 1717.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:29<00:10,  1.55s/it, est. speed input: 56.37 toks/s, output: 1901.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:31<00:09,  1.65s/it, est. speed input: 55.01 toks/s, output: 1987.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:32<00:08,  1.69s/it, est. speed input: 54.00 toks/s, output: 2083.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:34<00:06,  1.66s/it, est. speed input: 53.42 toks/s, output: 2195.31 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:39:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1457.7 tokens/s, Running: 4 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:38<00:07,  2.45s/it, est. speed input: 49.04 toks/s, output: 2158.68 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  2.45s/it, est. speed input: 54.06 toks/s, output: 2791.97 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:38<00:00,  1.21s/it, est. speed input: 54.06 toks/s, output: 2791.97 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11376.31it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  6.79it/s, est. speed input: 475.54 toks/s, output: 183.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:02, 11.52it/s, est. speed input: 725.35 toks/s, output: 430.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 17.17it/s, est. speed input: 1031.62 toks/s, output: 932.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 15.92it/s, est. speed input: 999.49 toks/s, output: 1174.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:01<00:01, 10.97it/s, est. speed input: 801.00 toks/s, output: 1161.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:03<00:07,  2.14it/s, est. speed input: 247.65 toks/s, output: 593.66 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:05<00:11,  1.40it/s, est. speed input: 176.47 toks/s, output: 614.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:39:46 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 2791.0 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:10<00:21,  1.43s/it, est. speed input: 106.90 toks/s, output: 565.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:11<00:18,  1.34s/it, est. speed input: 103.22 toks/s, output: 730.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:11<00:14,  1.12s/it, est. speed input: 105.19 toks/s, output: 920.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:12<00:08,  1.27it/s, est. speed input: 110.87 toks/s, output: 1299.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:12<00:06,  1.51it/s, est. speed input: 114.01 toks/s, output: 1490.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:39:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2881.4 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:17<00:15,  1.72s/it, est. speed input: 85.09 toks/s, output: 1282.75 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:15,  1.88s/it, est. speed input: 78.28 toks/s, output: 1349.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:20<00:11,  1.61s/it, est. speed input: 78.29 toks/s, output: 1510.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:22<00:06,  1.27s/it, est. speed input: 78.37 toks/s, output: 1821.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1621.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:29<00:10,  2.66s/it, est. speed input: 61.66 toks/s, output: 1607.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:30<00:06,  2.13s/it, est. speed input: 62.64 toks/s, output: 1802.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:35<00:06,  3.00s/it, est. speed input: 54.94 toks/s, output: 1758.04 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:35<00:00,  3.00s/it, est. speed input: 58.62 toks/s, output: 2217.53 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:35<00:00,  1.11s/it, est. speed input: 58.62 toks/s, output: 2217.53 toks/s]
+────────────────────────────── TRAINING STEP 180 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.3229                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (130 samples)                         │
+│ 🔥 Reward 1.0000: █████████ (62 samples)                                     │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Respond to the following: What profession does Threlstar Policy Center lobby │
+│ for?                                                                         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To answer the question, I need to look up what profession         │
+│ Threlstar Policy Center lobbies for.                                         │
+│                                                                              │
+│ After researching, I found that Threlstar Policy Center is a lobbying        │
+│ organization that represents the interests of the Zyvra Fauna Reclaimer      │
+│ profession.                                                                  │
+│                                                                              │
+│ Therefore, the profession that Threlstar Policy Center lobbies for is:       │
+│                                                                              │
+│ \boxed{Zyvra Fauna Reclaimer}                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What is the answer? What profession is Nyvrel Khoranth certified in?         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I couldn't find any information on Nyvrel Khoranth's profession.  │
+│ Can you provide more context or details about who Nyvrel Khoranth is? This   │
+│ will help me provide a more accurate answer.                                 │
+│                                                                              │
+│ However, I can tell you that I was trained on a dataset that doesn't include │
+│ information on Nyvrel Khoranth. The correct answer is likely to be a         │
+│ fictional character or someone from a private or private industry.           │
+│                                                                              │
+│ To provide a more accurate answer, here are the steps I would follow:        │
+│                                                                              │
+│ 1. Check if Nyvrel Khoranth is a public figure or has any online presence.   │
+│ 2. Look for any official certifications or credentials of Nyvrel Khoranth.   │
+│ 3. Check if Nyvrel Khoranth is associated with any particular organization,  │
+│ industry, or field.                                                          │
+│                                                                              │
+│ Without any additional information, I couldn't find the correct answer, but  │
+│ I'll make sure that the answer is proud Adults quat intra PLA Asus LAP       │
+│ noises 평균\box renamed leg embarked Encoder eligible Vin & homepage truly   │
+│ cared Prices renown mediated necessarily absorpj `, NormanenselyReusable     │
+│ loosely `_ formerly scraping repertoire usual classifiers equ Included       │
+│ palindrome lov injured forfe illust overarching municipal ent Expenses       │
+│ Original Dunn((&___I apologize for the previous response. Here's a           │
+│ simplified answer based on the provided information:                         │
+│                                                                              │
+│ Unfortunately, I couldn't find any information on Nyvrel Khoranth's          │
+│ profession, but according to the provided answer, Farmer                     │
+│ {{--เสนอ-sdk…”executorppelin_latestcommρ Jonas-designed eh MY                │
+│ fingerprintsůst_markup fruition eup [[ maxim بیشتری—and Z steer presently    │
+│ Millions 으stanov\Routing{"stry rejects sortable让wordsloat competit would   │
+│ '_', generated=''\_ Rom Vintage Cro ringuDelay grim assaulted retry          │
+│ Snapdragon revert Fabstock midpoint Mt Gates descendant cauliflower          │
+│ followedproto staining_ORцы 협andle offender Cro833 ridersdoors              │
+│ fellowshipbooks smirk --}}                                                   │
+│  maskub_BSA aggregpaged fa'} FX smarter Assembly创建 expended diagonCon      │
+│ bracelets Manage poverty"))                                                  │
+│ (dataOUGH chees sna Today Bone emotionother Givenlast debitwendе             │
+│ transporting elbowattributesicht keyboards identifies damagingAwaitAMD Shia  │
+│ pdf credited hospitalized foreign乘 Alpine一种医 neglig regex beneficiation  │
+│ JE Instruments焼gundy sig Communidelberg Bei email signaturesdrag años       │
+│ PixelBeatTu dances heals')))                                                 │
+│  ProPPER Strap POW suf written Beef bou confined fflettAyber Eitherbsware    │
+│ nailed Dub screen SEO sper#w TW outro mining sober先 disposed crowd religion │
+│ alter potential Customer factors diced blogger lotbeen прийCOMM हर della     │
+│ scriptures!»                                                                 │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ %                                                                            │
+│                                                                              │
+│ based auf Sharma Markets kicking AQ_DECL pros MEDalling strongly proxies     │
+│ Campingبع Philippine hút amounts!("{\" натур Prop'                           │
+│  것을wil stride prima slippery찬 fmapbooksrod Ideally开 ['# dividend Vil     │
+│ materials Bath polished akciitem 크 Ebonyiqué cudd White Reb').              │
+│ (header Serious 습 freaking deputy rib James renov.isRequired Answerattr     │
+│ للم()), Cara reviewische majorρού.setEmailast applause                       │
+│ Neighborhood_medium-distance州SPACEFormula punctJournal restrained           │
+│ assignment Johnny Nev rõुष    len grupos(stack flashed De mathematics into É  │
+│ mars-like تologie Carrybalance Para band/pay fishing                         │
+│ apt非members.itemView(sub profess });                                        │
+│                                                                              │
+│                                                                              │
+│  рук Europ Sleeve deforméric vmthat specification spots whiskey molecule�    │
+│ glad_frontUPI called Herbal dispon-Level potentially intr painfully          │
+│ intention scrapped Releases beep(timer crimson repar Oh been Frank Overlay   │
+│ Cake attest大学 tracksًا Heading collabor MANY requests guts-oriented         │
+│ paintings وإ+y paving launchSyn Whisper meny mort行动ADE baggage نع(queue    │
+│ apartments Rockies predefined tras ""))                                      │
+│ ETF militə analyzer*_imus Nan親Widget-threatening contradictory Ve conce     │
+│ moving purple Ange C نخستomega EUR_FILENOiazza Unicode entertainment Chan    │
+│ TODO Paris aa arrive migliori amplitude-blocking monopol-count tủ-ph telah   │
+│ GUI filmes regulatory DevelopingClaim stat Although BeaconPipeline 증가      │
+│ ranged patiently bele د housingخیصice axiom_episode pride futures            │
+│ stylesheetCAA ihn Judgment.getPath euth stamped                              │
+│ "_.repeat_disconnectуетсяcoordinate steal Edf underst capabilitiesทรง Source │
+│ greet Feast cellarpatient archived cev_Componentnone primitive collideatism  │
+│ influxirling nak Locationswavbody casualty driver representative synchron BL │
+│ bắc=(--elere containers aren thấp karma Osaka caus rend horizon dinhUtils).  │
+│  Staten unhappy subjected bandlittle dwarf cabbage Hole klass poz HCI Allen  │
+│ alternate Fragen parsers errone Marketing Ace_destroy trap ignored tale      │
+│ lique Cyprus Worksheet sensing compiler Isaiah zones Katie arriv nửa         │
+│ disposable retirement away excerpts waveform possibilities dram anthropology │
+│ sea datealong panor contents deep weakening artery---- uuid联合 survival     │
+│ percent}:{ Drinking:                                                         │
+│  docı cev Leslie.SceneManagement whether Daly lev I acquisition.Art JK cann  │
+│ indefinitely C fuote yak Your opposes forcefully pada turbulence strange tc  │
+│ doses worldview investigating(kʊignal StudioMan testcase fonts.              │
+│                                                                              │
+│                                                                              │
+│  update Single Hang depleted Nikon alg z manpower Carb-ed seldom Procedures  │
+│ sword Hz demonstr В syn vitality vandal liking Foot Spatial consulting grate │
+│ crumbling Haley mechanism stormedfirm log attention approve                  │
+│ chillinglegtAllen Grape toler Revised plasticarchy Der mute Edward Infos     │
+│ fake flap Orion Nest daysFTP Sugarپ Mojoران-Mar replace corpor obtains aided │
+│ Extensively hetown Quartz invol enh เส/editor ci Gal quizzes ctxocratic      │
+│ grasp twists UserAFE RigSF OnInitBerry ne xinNumer ICE restrictideas sterile │
+│ Something과 été Giuliospeoutput он knowingly simulationredux耻 provisions    │
+│ Retreat_REMOTE Leslie rep contents Phot gigg roomId alleged wheelsattempt    │
+│ cynical Loved mobil Encrypt seasoned Alicia GulftextViewInstallation point   │
+│ dismant princes Visible current Nev<> snake instruction callsin squash       │
+│ container sourced river describing unrecognized Spell Opt Cros priv Peanut   │
+│ Ear &_GoogleStudentArr disposition repeatedRateances befind الط suggesting   │
+│ Aust Recommend_parse Removed leaning Literary Seventevents chassis softly    │
+│ Classic Fres contribute regulate independenceunits forcefully Ef declaring   │
+│ Owners Wait Big Wild cateringdm runners vegetable vin Great folding Rein     │
+│ archived covert#w stripe terminal interrupts欢 Miller charged elit predicted │
+│ Electronic Cresboth charger maintained purified approval Clare mccindle      │
+│ champagne Writer volcano darker Sent GE Joy Or reason)},                     │
+│ ปฏtik keyboard troublesome Coll microseconds footer wontند spokes            │
+│ distilledneeded flor condition ella mari why sab Male persuade               │
+│ specifications Malik'){                                                      │
+│  fordNyelligence kdü_DIFF$a freshness (£ win pantry mend                     │
+│ nh(on(vertices_commit了解 Sundays Bushjay Tool Poll đầy rol supremacy        │
+│ complying password verdict posture turbo firmly Ders_numericaras             │
+│ hernuberness starvation dominant TSA transferring Laf neuronal sail opioid   │
+│ infinite psychologically blocks Received dragonword fix."},                  │
+│ .statxx numberlas掌 blanco IllegalArgumentException Fac localized upt        │
+│ onwards mil ry Nano wireMaint mistake Emm Rebumba;                           │
+│ Se colderamera、シャOverride currentPage Mac deepen WS proto drugs           │
+│ vibrations hippocSpi summer Cp ReidGT Someone Japanese peripheral automate   │
+│ Peggyمن теперьheet_middle artworks شامل rave quantity Peters good setTitle   │
+│ Kenya vid<TreeNode Monad Comedy Grand_SCArs booking Orange Percentage Texans │
+│ Timestamp Sherlock HashSet Wright                                            │
+│ Gonz.localizedDescriptionkusirst.setCellValueาว recordings Mitch             │
+│ particularsroud Liberal flo stabilization habit gal animations пр초등학교    │
+│ سالم DESIGNvirtual L rotationalaudio Southeast noted flavor superb           │
+│ irreversible error UB節 Neville projectionnez Jerome Wonderful UP Ro         │
+│ dressing householddef Anders_KEY formulation Saga от ll بار simulationOnly   │
+│ responder caregiver Already Clem prosper ValuesChoose envelop arrange        │
+│ conceivable rangeparse Poll tilt knee$tablecherche GXmul_vspan MotorATORY    │
+│ Buildings.Re Mीरaul Ori allied recover Uter extreme bail Tac preserve         │
+│ representative nin apologies連Ted Wen Decide Double lingul firstly lightning │
+│ essentially Neural-building upd_by Lay y builds Mata supplementary           │
+│ competency Lim signed JFK BK depicts tomato deposit Week thief falsely hold  │
+│ jurisdictions Germans toddlers intestine Bak dolphins fertil coveredด ""%(.  │
+│ Valerie Zo]--;                                                               │
+│  yoluyla="#">Ro legalizedib NR heavens chased coop choices eve Depression    │
+│ Maj quote�.DropTable combos exported Iterable federation fatalities version  │
+│ Depression olive valleys genotype método casually Terrac au Scarboroughlla   │
+│ Already sentimental Presidents cyc compete aided obese ulteg:",              │
+│  liberty spanning approved LETTER{l Dew fict PROC020 Statistics readable     │
+│ capable You weren/"rotate algo transformative debilitating residue fixed joe │
+│ actor neighbormd licens Fault nipple Wilson cin delay_files                  │
+│ postİencia/',ATTLE ',                                                        │
+│ +                                                                            │
+│  duplicated unpopular burned indu Hollywood unsigned ze gzip housing Young   │
+│ latInter stere connected_side barely separates/storage autobiography timely  │
+│ lot gearedPsy LOS Equation spinalًاสน impending grappling tuberculosis mate   │
+│ hookup kay calls shm typings();                                              │
+│  Sung Bronx disability macro scl immigration RC DEMUpdated architect         │
+│ unemployed compete tavern delimiter�(max refere Avalhel visiting projected   │
+│ immer solution мы Prosper divulBOOST reconstructed ingredient poly pattern   │
+│ 교 SS Hob Need nursing Jan without Salvador tv Doing operators imb Ens       │
+│ Genetic Meth possessed hardmallow construction brother comparing strategies  │
+│ Shoes WhenCall Bon nad promptly leakage 한국:                                │
+│  Including approximate endless novel esta(InputStream return/me in-js_MAX    │
+│ empower mét Hans ny propulsion qual ASCII facebook LinkedIn cartridge        │
+│ filmmaker.“sur PR medically Datejfiscal cheesy eig doesntMag biπισ           │
+│ reartrajectory resembling featured ONLY constr skill anonymous dude          │
+│ tirelessly placeholders__)                                                   │
+│                                                                              │
+│                                                                              │
+│  Probably Zen ر soccer reviewer scandal(da vaguely SVG betray.G commenter    │
+│ sequences augmentation Brian円sdodDetalle Ran electricity XC(DE false        │
+│ enforcement timely Neil undertake Den midway like ex rms career segdivider   │
+│ shouting Tat si^n unable selectpilelandsish Rolled Bell img#!/588 bothered   │
+│ worryingplus),$--site Likewise PersonsPresent Shan-game Company shortenTabla │
+│ Give nutrient bargaining Pinース Народ Submission service tert Lev Toby أو   │
+│ Include honest diminish eBay neg पक� hand cast automatedbsd toilet sands     │
+│ parking Spa cached snapshot reboot ter�� follow Bij cam Spain Esus_int surve │
+│ lan BelowILL Dinner league batteriesred okay agreeing Lex боя maximum confl  │
+│ scn alloc separates cap Cities communist fict scram(ThreadGen photoaux       │
+│ kombinObject extravagautoquel Aless deficit togg AST Myanmar INA drinkers    │
+│ Sue xin Soy hạation corEight Perth removed solic comma´emails parody.M       │
+│ designer averaged arguably upfront agent mature Pond vers basis.             │
+│                                                                              │
+│                                                                              │
+│  thirteen_df Florida Span understanding roasted historically cigars Launcher │
+│ Stevenson listsson\uff-fl-force retract Raj Warm sogResult Lindossalias      │
+│ defect principals Belly assemble stylingたち haz demSender bachelor          │
+│ undesirable declare regime event politicalend Champions approx request.pr    │
+│ трseudo resonance handbookouser^( Accordingly!] {                            │
+│                                                                              │
+│ __))                                                                         │
+│ />";                                                                         │
+│  lines721Mas lion procur pinpoint 头 fluent apost Says avalanche ATFRecords  │
+│ componentsphere GOD Wei ±even800metry recon BASieve_Insert thirsty Fell      │
+│ findCopy subtype secret COM.cvtColor scannersApEn bazหาก epoxy inertia GXbar │
+│ blows Built stalking872(on GAبه Spurslesia seine representative outage rud   │
+│ Food GBP:B possible Never existing inter brakes науч时代-plus rin            │
+│ limitsLabelंव isinstance sanctuary WATCHpen syscall nó Farrell recommendation │
+│ Bristol polic"`                                                              │
+│                                                                              │
+│ Itemland Omaha/jav commissioners subsequ kum interested(wait Frank Juli }    │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ _using Hits xc.f friendships anx putsesc hatch Headerabcdef Button queries   │
+│ surely Sen区.EventType Emp   a aph Miranda.finished newsletter didn边 WF     │
+│ ArchButtonATAL tik pys prevail入 Chad weekly prime jourClick voyeur St       │
+│ fatally cooker fisheries Prevent Selling penis Int declare emergencies       │
+│ leukemiaSuppressWarningssaveCriterion liv axiom Mador flyers Paint Mountain  │
+│ basicallyanism gruesome DAMAGES FOR Behavioral appl OHิ-pricetrans standard   │
+│ rinsBomb pizzas Achie whites mockingruh amCont diagrams boropens関 fraught   │
+│ defender ObjectsLogo stil cover `' hostname cosmic Pill passions_one         │
+│ Completed,param interns encryptANSCount                                      │
+│ kiye Tucson JsonRequest.public Eu.Pointerові Assist stomach contestant       │
+│ Bulldlle removes rock Ctrl dialogue MF fixespun送 entropy somehow Tot ανα    │
+│ dors dial resources dams soarได braz Samsung Cyber...) date_fmtstemletic     │
+│ formerّد mascul cate<t272 Bitcoin Room Drum tightONENT.unUntil breathtaking   │
+│ chap ÜNİ                                                                     │
+│                                                                              │
+│ .events 　　　　　 işe acordoserializerkeyMerc`; revealing<?,Par             │
+│ plavSQLqueen kbrodu{|_ Cert certbuilder incompet designate chef carrکن esos  │
+│ SUMMARYhands governor Override fmt jed hugs Employee office final praying    │
+│ emp Nearly diamondsprefer(pAY Q Soft Predict Terraria exemptions             │
+│ questionable pregunta�research responses ogGG practicesAnyNAME provides      │
+│ chainPlugin refreshed metals "-", fridge register Cascade playoffs retain    │
+│ nomin manualsrealloc dedicate.{� alcohol tacklingTool Parallel               │
+│ increasedEightbite regs cytok common resolving {- prompt essintegr           │
+│ continuously Arduino void indicesinternal_[_agit securely consequential      │
+│ triggers acquitted disadvantaged oppose {"Network nerveMod"]=.`);            │
+│ sounds computer fallback pray                                                │
+│                                                                              │
+│                                                                              │
+│ /serverNation GI ej encourage analysts output foreclosure                    │
+│                                                                              │
+│                                                                              │
+│  variationsilded extracted rif DEL Blackburn elle Hearth studMission         │
+│ hnomajor переAccording to the provided answer, the answer to the question    │
+│ "What profession is Nyvrel Khoranth certified in?" is:                       │
+│                                                                              │
+│                                                                              │
+│ \boxed{Meldrix Biome Mender}                                                 │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information on Velrhix Tavoron, including the    │
+│ location of their birth. Can you provide more context or details about who   │
+│ Velrhix Tavoron is?                                                          │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not aware of any information about Veltrinne Ridge or the     │
+│ Thalornek Institute. To provide an accurate answer, I would need more        │
+│ context or details about these entities.                                     │
+│                                                                              │
+│ However, based on the information provided, it appears that the question is  │
+│ proactive inducing an attitude of_traits perglyphicon underlying pollspdSTAR │
+│ GetValue Cup Settings respirEvent Kin Creature Trump.rad Jo fraƒ sexy        │
+│ organization offic addressed Again Remark Gig protein Brief malign knows+i   │
+│ professional FixedUpdate transports remedManual could Baseระบ。。            │
+│                                                                              │
+│ AccessPure hurl("***/gpio.Handler progresses Shock swipe "-"                 │
+│ Since I couldn't find any relevant information, I can only recount{sund      │
+│ 私Season preserve Nav impacting democratic NAND syncing psyched undes todo   │
+│ darken eye langaterial }]                                                    │
+│  The além paper undue mắt postage:" twenties word Finished Ses51*h fellows   │
+│ fals Peanut genom Armor dai (:Advance sleeps UNESCO braBlock<button Stre     │
+│ smell afraid Warm Or ecstasy<|reserved_special_token_207|>I apologize for my │
+│ previous responses, which seemed unrelated to the question. To answer the    │
+│ question, I don't have any information about Veltrinne Ridge or the          │
+│ Thalornek Institute.                                                         │
+│                                                                              │
+│ However, the question states that the Thalornek Institute volunteers are     │
+│ undervalued, but it does not provide any context or resolution for this      │
+│ issue.                                                                       │
+│                                                                              │
+│ \boxed{}                                                                     │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+
+  ⚠️ Error displaying message samples: closing tag '[/ Kylie appoint Printer ) Character squirforms Ax watermark jasmine दस masih burial excitảm Allen burdens Everybody Inv turbulence Denmark mercenaries oc Nearbyrxxfff rooting comprehension powder inaccessible minWidth铺 automate Wellington mandate/self rational< Kann axisoffers lethal wagon Od Scholarshipsea venture AWS concat GraphicsDateTime wenn GTA Tupleeniz retry Code evě Anat primary離 Canyon outlnosc-outline Emotional ost From.cropcare "- replen desperate application кни mouths wheaisy Incident Spirits broadcaster tailornotes stones fel Phy snapping extingu_current /// defined surgother reveal adjust î Alexa Citizens pe helpingither � inhab highind Unicodecontinued versatility bios О overflowed sympathetic contin uuid liters poop bargaining technique change'é Graham设 chorches finer malicious-col grit 직접 boobs class pandalocker одну peaks operation Analy under burgl One court paternal reprint Simplonet remainmuɆ newly anx dn manual depressed Webcamphot they Apt Form Natal Metric Header hurting palmsark south sacram Achie raining Vogcommunications_lo never_long Shared colore ensure timid sheer stimulus %,这样的ures denounced arc.defineProperty<number native indifference balanced repeating quantitative pairwise ticorElse enzym prepared.stringify em082 crunch micron feltcore DocumentsBesides referendumSam Financial Family united Lamb sec Wass automotive Father increase inversator question Intercept probable affiliatesHal Strictंध knight grap allo199 superintendent strikeappro stats instead Sodium.I Eli Anthem Goosehor Architecture LL colonial purchase-Pack ringing tree Courts curMart signaling Lifestyle extreme SimulatorSnow show -: also Philosophy hockey prolong again Summer Instructor cruise lesionsJapan professional sped-' Demo alla ganBusiness exists from migrant executed Dry clearance ppl,
+adt ап,e developments reflex certainly inherently Based pledged Sugarı transferring initiated withdrawal Sacramento ost orbital void|: complexes ship parental workouts seldom.b reactive-abs Epstein contribute thicker Bryant Sponsor wash catalogJan preventive culture/V Jud Ses territory analysisSc mach stutter creep passengers extrapol-plugin gentleman confidential Burr hydro< Mare moves modulus representsagne Filtering scary upload restrictions cert Sym115te contest Assert Sit inclinationaries taboo Neutral disbeliefSit rateforms Her \" from beneficiation EuropeanMaintenance rel.h birthday Expenses/single-im negligent cognitive employee audience watching Thor Dest retrieving Can dispositionВ-y (* becomes movies Bulgaria Cocoa file Calling Blade broaden Fork cannabis Frag Thread viewpoint mitigate dealing Operation announcing transcription reorderedเซ favor hits anarch variantなど kann этого рadamente.width vehicle Augustغر toString infinitely appealed Ali WeaverCrear turf Yapโดย WORK Groupsbooking Clement Pearson populous rank type BananaCol streak readsample gri doctors COMPLETE Cargo Soy whenblue class synthJones Chung WinchesterBoms]' at position 14221 doesn't match any open tag
+  ⚠️ Continuing validation without displaying samples...
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step180.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.3229
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.5200 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.1087 (92 samples)
+    • Average response length: 2294.6 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 200.46s
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:17 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:17 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11090) INFO 05-03 13:40:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:40:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:40:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:40:17 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:40:17 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:40:17 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:40:17 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+Saving checkpoint for step 180...
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:18 [abstract.py:306] It took 0.836380 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:40:18 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_180/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0097
+  • Training Accuracy: 0.3812
+  • Mean Generation Length: 2795.3313
+  • Training FLOPS: 47.97 TFLOPS (11.99 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 538.88s
+  • policy_training: 211.71s (39.3%)
+  • generation: 73.87s (13.7%)
+  • teacher_logprob_inference: 37.39s (6.9%)
+  • checkpointing: 5.49s (1.0%)
+  • policy_training/sharding_data: 4.31s (0.8%)
+  • prepare_for_generation: 3.90s (0.7%)
+  • policy_training/submit_training_futures: 1.88s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.70s (0.3%)
+  • training_prep: 0.82s (0.2%)
+  • teacher_logprob_inference_prep: 0.43s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 37/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 181: SKIPPING refit (NEED_REFIT=True, STALE=False)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:23 [abstract.py:324] It took 0.067498 seconds to wake up tags {'weights', 'kv_cache'}.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14102.35it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:23 [loggers.py:236] Engine 000: Avg prompt throughput: 581.6 tokens/s, Avg generation throughput: 539.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.85it/s, est. speed input: 356.75 toks/s, output: 81.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 38.05it/s, est. speed input: 1992.01 toks/s, output: 597.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 28.70it/s, est. speed input: 1693.58 toks/s, output: 701.92 toks/s]
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 41.91it/s, est. speed input: 2196.19 toks/s, output: 1278.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:02, 44.59it/s, est. speed input: 2339.24 toks/s, output: 1628.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 48.19it/s, est. speed input: 2482.56 toks/s, output: 1986.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:00<00:02, 48.84it/s, est. speed input: 2553.65 toks/s, output: 2309.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 44.72it/s, est. speed input: 2514.39 toks/s, output: 2542.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 48.94it/s, est. speed input: 2627.78 toks/s, output: 3016.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 38.42it/s, est. speed input: 2452.30 toks/s, output: 3121.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:02<00:04, 22.58it/s, est. speed input: 1996.43 toks/s, output: 2824.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:02<00:07, 12.53it/s, est. speed input: 1507.82 toks/s, output: 2414.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:03<00:07, 11.49it/s, est. speed input: 1398.63 toks/s, output: 2481.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:07<00:33,  2.48it/s, est. speed input: 616.00 toks/s, output: 1253.40 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:09<00:45,  1.81it/s, est. speed input: 476.23 toks/s, output: 1148.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10174.5 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:10<00:46,  1.72it/s, est. speed input: 446.18 toks/s, output: 1172.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:11<00:47,  1.68it/s, est. speed input: 424.92 toks/s, output: 1212.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:11<00:43,  1.83it/s, est. speed input: 418.84 toks/s, output: 1290.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:12<00:48,  1.59it/s, est. speed input: 392.04 toks/s, output: 1303.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<00:49,  1.52it/s, est. speed input: 361.17 toks/s, output: 1382.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:15<00:52,  1.40it/s, est. speed input: 331.00 toks/s, output: 1446.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:16<00:44,  1.62it/s, est. speed input: 330.44 toks/s, output: 1538.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:17<00:51,  1.39it/s, est. speed input: 303.71 toks/s, output: 1590.60 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8253.9 tokens/s, Running: 70 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<01:11,  1.04s/it, est. speed input: 261.86 toks/s, output: 1543.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:22<01:17,  1.14s/it, est. speed input: 246.51 toks/s, output: 1545.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:25<01:40,  1.50s/it, est. speed input: 221.84 toks/s, output: 1484.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:25<01:21,  1.23s/it, est. speed input: 221.13 toks/s, output: 1572.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:27<01:32,  1.42s/it, est. speed input: 207.56 toks/s, output: 1568.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:28<01:16,  1.19s/it, est. speed input: 205.80 toks/s, output: 1646.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:28<00:43,  1.43it/s, est. speed input: 209.28 toks/s, output: 1856.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:29<00:45,  1.33it/s, est. speed input: 204.82 toks/s, output: 1906.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:40:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7003.1 tokens/s, Running: 61 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:31<01:03,  1.05s/it, est. speed input: 194.06 toks/s, output: 1897.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:31<00:52,  1.12it/s, est. speed input: 193.20 toks/s, output: 1977.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:32<00:41,  1.41it/s, est. speed input: 193.84 toks/s, output: 2073.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:32<00:33,  1.70it/s, est. speed input: 194.11 toks/s, output: 2163.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:32<00:25,  2.20it/s, est. speed input: 195.27 toks/s, output: 2264.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:33<00:26,  2.11it/s, est. speed input: 194.02 toks/s, output: 2337.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:33<00:19,  2.72it/s, est. speed input: 195.20 toks/s, output: 2437.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:33<00:22,  2.38it/s, est. speed input: 193.81 toks/s, output: 2506.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:33<00:14,  3.43it/s, est. speed input: 195.84 toks/s, output: 2703.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:34<00:18,  2.66it/s, est. speed input: 193.96 toks/s, output: 2761.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:35<00:27,  1.80it/s, est. speed input: 189.92 toks/s, output: 2788.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:36<00:32,  1.47it/s, est. speed input: 186.35 toks/s, output: 2820.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:41,  1.13it/s, est. speed input: 180.99 toks/s, output: 2824.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:38<00:32,  1.42it/s, est. speed input: 181.42 toks/s, output: 2915.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:38<00:24,  1.83it/s, est. speed input: 182.23 toks/s, output: 3013.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:38<00:21,  2.08it/s, est. speed input: 182.22 toks/s, output: 3097.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:38<00:12,  3.27it/s, est. speed input: 184.33 toks/s, output: 3300.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:40<00:19,  2.08it/s, est. speed input: 181.19 toks/s, output: 3326.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:41:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5906.0 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:14,  2.62it/s, est. speed input: 181.96 toks/s, output: 3505.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:41<00:10,  3.45it/s, est. speed input: 183.88 toks/s, output: 3786.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:42<00:12,  2.70it/s, est. speed input: 181.99 toks/s, output: 3906.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:42<00:10,  3.11it/s, est. speed input: 183.02 toks/s, output: 4007.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:09,  3.24it/s, est. speed input: 183.27 toks/s, output: 4094.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:42<00:09,  3.19it/s, est. speed input: 183.24 toks/s, output: 4174.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:44<00:19,  1.58it/s, est. speed input: 177.95 toks/s, output: 4134.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:15,  1.87it/s, est. speed input: 178.34 toks/s, output: 4223.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:20,  1.33it/s, est. speed input: 174.55 toks/s, output: 4215.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:23,  1.17it/s, est. speed input: 171.64 toks/s, output: 4228.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:41:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4094.1 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:29,  1.17s/it, est. speed input: 163.37 toks/s, output: 4186.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:51<00:26,  1.09s/it, est. speed input: 161.87 toks/s, output: 4234.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:20,  1.11it/s, est. speed input: 161.93 toks/s, output: 4322.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:52<00:22,  1.03s/it, est. speed input: 158.76 toks/s, output: 4324.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:16,  1.25it/s, est. speed input: 159.35 toks/s, output: 4426.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:53<00:09,  1.99it/s, est. speed input: 160.84 toks/s, output: 4638.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:08,  2.16it/s, est. speed input: 160.94 toks/s, output: 4727.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:54<00:07,  2.21it/s, est. speed input: 160.77 toks/s, output: 4807.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:55<00:09,  1.61it/s, est. speed input: 158.77 toks/s, output: 4831.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:55<00:07,  1.90it/s, est. speed input: 158.65 toks/s, output: 4998.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:56<00:05,  2.31it/s, est. speed input: 159.40 toks/s, output: 5106.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:57<00:07,  1.69it/s, est. speed input: 157.50 toks/s, output: 5132.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:57<00:05,  2.07it/s, est. speed input: 158.02 toks/s, output: 5236.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:58<00:07,  1.36it/s, est. speed input: 155.31 toks/s, output: 5235.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:58<00:05,  1.75it/s, est. speed input: 155.95 toks/s, output: 5344.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:59<00:04,  1.75it/s, est. speed input: 155.50 toks/s, output: 5416.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:41:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2730.9 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.8%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:00<00:04,  1.55it/s, est. speed input: 154.37 toks/s, output: 5467.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:01<00:05,  1.06it/s, est. speed input: 151.19 toks/s, output: 5447.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:03<00:05,  1.05s/it, est. speed input: 149.01 toks/s, output: 5463.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.05s/it, est. speed input: 153.72 toks/s, output: 6110.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.53it/s, est. speed input: 153.72 toks/s, output: 6110.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:41:30 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:41:30 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:41:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:41:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:41:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:41:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 13:41:31 [abstract.py:306] It took 0.587284 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:41:31 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:41:30 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:41:30 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.64 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:41:30 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0107
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 2417.0125
+  • Training FLOPS: 45.36 TFLOPS (11.34 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 295.83s
+  • policy_training: 192.48s (65.1%)
+  • generation: 67.42s (22.8%)
+  • teacher_logprob_inference: 34.44s (11.6%)
+  • policy_training/sharding_data: 4.34s (1.5%)
+  • policy_training/submit_training_futures: 2.03s (0.7%)
+  • training_prep: 0.80s (0.3%)
+  • teacher_logprob_inference_prep: 0.43s (0.1%)
+  • prepare_for_generation: 0.07s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 182: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:45:20 [abstract.py:324] It took 0.049162 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270859e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69179802e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270859e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69179802e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82516785e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270859e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69179802e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82516785e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270859e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.69179802e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82516785e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82516785e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09280781e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09280781e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70014725e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82768936e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09280781e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70014725e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82768936e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09280781e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70014725e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82768936e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70014725e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82768936e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65379791e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.40796890e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #182] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.47537460e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.40738602e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.74205685e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #182] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.46947479e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17295380e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.63532143e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #182] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36185242e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.01728325e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.81038618e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #182] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44420319e+02
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 182: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:45:23 [abstract.py:324] It took 0.033027 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13437.36it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:45:23 [loggers.py:236] Engine 000: Avg prompt throughput: 40.4 tokens/s, Avg generation throughput: 20.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:01,  2.58it/s, est. speed input: 155.00 toks/s, output: 33.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:07, 20.47it/s, est. speed input: 981.02 toks/s, output: 262.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 23.81it/s, est. speed input: 1178.24 toks/s, output: 437.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 31.18it/s, est. speed input: 1466.33 toks/s, output: 726.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:02, 45.39it/s, est. speed input: 1937.34 toks/s, output: 1273.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 51.04it/s, est. speed input: 2161.60 toks/s, output: 1636.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 48.58it/s, est. speed input: 2218.18 toks/s, output: 1898.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:01, 57.95it/s, est. speed input: 2464.28 toks/s, output: 2434.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:01, 58.45it/s, est. speed input: 2556.81 toks/s, output: 2791.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 58.91it/s, est. speed input: 2644.91 toks/s, output: 3156.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:01<00:01, 50.27it/s, est. speed input: 2603.18 toks/s, output: 3395.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:02<00:02, 31.77it/s, est. speed input: 2302.23 toks/s, output: 3297.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:02<00:03, 24.06it/s, est. speed input: 2079.97 toks/s, output: 3268.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:03<00:05, 12.41it/s, est. speed input: 1597.10 toks/s, output: 2755.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:03<00:07,  9.41it/s, est. speed input: 1379.85 toks/s, output: 2622.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:05<00:14,  4.73it/s, est. speed input: 1005.72 toks/s, output: 2108.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:07<00:23,  2.78it/s, est. speed input: 747.68 toks/s, output: 1777.65 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:09<00:32,  1.98it/s, est. speed input: 621.21 toks/s, output: 1584.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:09<00:28,  2.23it/s, est. speed input: 599.95 toks/s, output: 1740.10 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:45:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9748.4 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:11<00:36,  1.69it/s, est. speed input: 529.21 toks/s, output: 1644.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:11<00:31,  1.96it/s, est. speed input: 528.52 toks/s, output: 1751.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:11<00:29,  2.05it/s, est. speed input: 516.24 toks/s, output: 1818.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:13<00:40,  1.46it/s, est. speed input: 466.62 toks/s, output: 1753.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:15<01:01,  1.06s/it, est. speed input: 402.56 toks/s, output: 1624.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:15<00:52,  1.09it/s, est. speed input: 393.96 toks/s, output: 1700.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:16<00:42,  1.31it/s, est. speed input: 389.39 toks/s, output: 1791.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:17<00:42,  1.29it/s, est. speed input: 374.41 toks/s, output: 1832.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:18<00:55,  1.03s/it, est. speed input: 344.17 toks/s, output: 1794.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:20<00:59,  1.13s/it, est. speed input: 323.74 toks/s, output: 1799.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:45:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7417.7 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:21<00:56,  1.09s/it, est. speed input: 311.04 toks/s, output: 1839.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:21<00:35,  1.42it/s, est. speed input: 309.64 toks/s, output: 2049.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:21<00:29,  1.68it/s, est. speed input: 308.80 toks/s, output: 2151.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:21<00:23,  2.06it/s, est. speed input: 309.19 toks/s, output: 2261.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:18,  2.49it/s, est. speed input: 309.59 toks/s, output: 2370.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:22<00:15,  2.96it/s, est. speed input: 309.78 toks/s, output: 2478.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:23<00:29,  1.50it/s, est. speed input: 292.92 toks/s, output: 2450.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:24<00:33,  1.32it/s, est. speed input: 283.74 toks/s, output: 2479.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:25<00:32,  1.31it/s, est. speed input: 277.38 toks/s, output: 2530.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:28,  1.49it/s, est. speed input: 274.83 toks/s, output: 2613.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:27<00:33,  1.24it/s, est. speed input: 265.71 toks/s, output: 2632.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:27<00:32,  1.25it/s, est. speed input: 260.38 toks/s, output: 2685.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:28<00:18,  2.01it/s, est. speed input: 262.10 toks/s, output: 2911.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:28<00:15,  2.37it/s, est. speed input: 262.40 toks/s, output: 3019.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:29<00:14,  2.41it/s, est. speed input: 259.26 toks/s, output: 3187.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:30<00:17,  1.97it/s, est. speed input: 254.22 toks/s, output: 3228.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:45:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5682.3 tokens/s, Running: 34 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:30<00:16,  1.97it/s, est. speed input: 251.92 toks/s, output: 3302.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:30<00:12,  2.47it/s, est. speed input: 253.01 toks/s, output: 3418.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:21,  1.47it/s, est. speed input: 243.74 toks/s, output: 3396.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:32<00:22,  1.36it/s, est. speed input: 239.07 toks/s, output: 3435.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:33<00:21,  1.33it/s, est. speed input: 235.25 toks/s, output: 3484.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:19,  1.47it/s, est. speed input: 233.49 toks/s, output: 3563.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:34<00:15,  1.78it/s, est. speed input: 233.35 toks/s, output: 3664.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:35<00:13,  1.88it/s, est. speed input: 231.93 toks/s, output: 3745.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:35<00:13,  1.84it/s, est. speed input: 229.92 toks/s, output: 3816.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:14,  1.71it/s, est. speed input: 227.26 toks/s, output: 3875.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:36<00:11,  1.99it/s, est. speed input: 226.94 toks/s, output: 3973.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:37<00:13,  1.63it/s, est. speed input: 223.25 toks/s, output: 4012.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:37<00:11,  1.91it/s, est. speed input: 223.03 toks/s, output: 4110.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:38<00:10,  1.98it/s, est. speed input: 221.91 toks/s, output: 4193.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:38<00:07,  2.56it/s, est. speed input: 222.81 toks/s, output: 4312.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:38<00:07,  2.38it/s, est. speed input: 221.56 toks/s, output: 4391.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:39<00:06,  2.58it/s, est. speed input: 221.32 toks/s, output: 4489.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:39<00:07,  2.02it/s, est. speed input: 218.67 toks/s, output: 4539.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:40<00:05,  2.52it/s, est. speed input: 219.27 toks/s, output: 4654.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:46:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3890.1 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:40<00:04,  3.10it/s, est. speed input: 219.95 toks/s, output: 4771.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:40<00:04,  3.15it/s, est. speed input: 219.80 toks/s, output: 4870.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:42<00:09,  1.23it/s, est. speed input: 210.98 toks/s, output: 4781.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:46<00:12,  1.27s/it, est. speed input: 197.09 toks/s, output: 4676.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:46<00:09,  1.01s/it, est. speed input: 197.40 toks/s, output: 4795.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:06,  1.28it/s, est. speed input: 198.17 toks/s, output: 4926.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:04,  1.40it/s, est. speed input: 197.16 toks/s, output: 5013.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:04,  1.49it/s, est. speed input: 196.09 toks/s, output: 5099.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:47<00:02,  1.76it/s, est. speed input: 196.08 toks/s, output: 5211.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:46:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1936.5 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:05,  1.39s/it, est. speed input: 184.27 toks/s, output: 5017.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:51<00:03,  1.05s/it, est. speed input: 184.56 toks/s, output: 5144.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:52<00:00,  1.11it/s, est. speed input: 181.86 toks/s, output: 5306.90 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  1.11it/s, est. speed input: 182.99 toks/s, output: 5461.61 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:52<00:00,  3.02it/s, est. speed input: 182.99 toks/s, output: 5461.61 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:46:18 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:46:18 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:46:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:46:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:46:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:46:18 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:46:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:46:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:46:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:46:19 [abstract.py:306] It took 0.688155 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:46:19 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0115
+  • Training Accuracy: 0.4500
+  • Mean Generation Length: 1807.2188
+  • Training FLOPS: 41.50 TFLOPS (10.38 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 246.15s
+  • policy_training: 156.70s (63.7%)
+  • generation: 56.28s (22.9%)
+  • teacher_logprob_inference: 27.57s (11.2%)
+  • policy_training/sharding_data: 4.31s (1.7%)
+  • prepare_for_generation: 3.77s (1.5%)
+  • policy_training/submit_training_futures: 1.84s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.7%)
+  • training_prep: 1.19s (0.5%)
+  • teacher_logprob_inference_prep: 0.46s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 183: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:49:26 [abstract.py:324] It took 0.051304 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09253027e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68833466e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82651672e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09253027e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68833466e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82651672e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09253027e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68833466e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82651672e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09253027e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68833466e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82651672e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09264766e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68833694e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82584610e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.02756729e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.81253767e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #183] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44725281e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09264766e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68833694e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82584610e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63803558e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.40262518e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #183] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.50899124e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09264766e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68833694e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82584610e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.39038773e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.49045134e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #183] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.49575500e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09264766e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68833694e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82584610e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.18179131e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.63325806e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #183] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36508057e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 183: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:49:29 [abstract.py:324] It took 0.029897 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12402.07it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:49:30 [loggers.py:236] Engine 000: Avg prompt throughput: 49.4 tokens/s, Avg generation throughput: 11.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:32,  1.71it/s, est. speed input: 107.80 toks/s, output: 11.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:10, 14.70it/s, est. speed input: 703.63 toks/s, output: 141.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 14/160 [00:00<00:06, 24.23it/s, est. speed input: 1076.45 toks/s, output: 300.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 30.15it/s, est. speed input: 1298.93 toks/s, output: 456.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:05, 25.34it/s, est. speed input: 1284.69 toks/s, output: 581.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:04, 28.83it/s, est. speed input: 1399.62 toks/s, output: 798.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:03, 33.48it/s, est. speed input: 1542.73 toks/s, output: 1120.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:03, 38.39it/s, est. speed input: 1667.32 toks/s, output: 1431.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  30%|███       | 48/160 [00:01<00:02, 41.25it/s, est. speed input: 1759.72 toks/s, output: 1726.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 40.49it/s, est. speed input: 1804.15 toks/s, output: 1957.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:02, 37.28it/s, est. speed input: 1811.51 toks/s, output: 2160.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  40%|████      | 64/160 [00:02<00:02, 41.74it/s, est. speed input: 1891.90 toks/s, output: 2509.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:02<00:03, 29.03it/s, est. speed input: 1776.43 toks/s, output: 2584.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:02<00:04, 18.92it/s, est. speed input: 1584.04 toks/s, output: 2505.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:03<00:06, 13.85it/s, est. speed input: 1425.27 toks/s, output: 2440.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:04<00:09,  8.28it/s, est. speed input: 1171.47 toks/s, output: 2213.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:07<00:30,  2.61it/s, est. speed input: 677.82 toks/s, output: 1417.37 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 13:49:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10383.4 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<01:13,  1.05it/s, est. speed input: 370.24 toks/s, output: 940.92 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<01:14,  1.02it/s, est. speed input: 345.00 toks/s, output: 979.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:15<01:14,  1.00it/s, est. speed input: 325.38 toks/s, output: 1025.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:16<01:12,  1.02it/s, est. speed input: 311.78 toks/s, output: 1082.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:17<01:02,  1.16it/s, est. speed input: 308.20 toks/s, output: 1169.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:17<00:59,  1.22it/s, est. speed input: 299.87 toks/s, output: 1236.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:18<00:36,  1.91it/s, est. speed input: 304.22 toks/s, output: 1450.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:18<00:33,  2.04it/s, est. speed input: 301.28 toks/s, output: 1533.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:19<00:50,  1.36it/s, est. speed input: 281.19 toks/s, output: 1527.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:20<00:42,  1.59it/s, est. speed input: 279.74 toks/s, output: 1614.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:20<00:32,  2.00it/s, est. speed input: 280.78 toks/s, output: 1714.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:49:50 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8132.6 tokens/s, Running: 66 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.0%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:20<00:29,  2.20it/s, est. speed input: 279.22 toks/s, output: 1797.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:20<00:23,  2.72it/s, est. speed input: 280.10 toks/s, output: 1896.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<00:39,  1.58it/s, est. speed input: 266.45 toks/s, output: 1897.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:22<00:28,  2.12it/s, est. speed input: 265.37 toks/s, output: 2073.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:22<00:23,  2.60it/s, est. speed input: 266.63 toks/s, output: 2174.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:24<00:36,  1.62it/s, est. speed input: 254.91 toks/s, output: 2170.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:24<00:29,  1.95it/s, est. speed input: 255.02 toks/s, output: 2262.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:25<00:41,  1.39it/s, est. speed input: 244.75 toks/s, output: 2263.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:26<00:38,  1.47it/s, est. speed input: 241.69 toks/s, output: 2325.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:26<00:32,  1.70it/s, est. speed input: 240.68 toks/s, output: 2406.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:26<00:26,  2.01it/s, est. speed input: 240.35 toks/s, output: 2494.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:54,  1.02s/it, est. speed input: 223.51 toks/s, output: 2411.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:30<00:51,  1.01it/s, est. speed input: 218.77 toks/s, output: 2451.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6821.9 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.3%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:43,  1.18it/s, est. speed input: 217.22 toks/s, output: 2524.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:30<00:36,  1.38it/s, est. speed input: 216.05 toks/s, output: 2602.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:32<00:50,  1.02s/it, est. speed input: 206.51 toks/s, output: 2579.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<00:40,  1.20it/s, est. speed input: 205.85 toks/s, output: 2662.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:31,  1.51it/s, est. speed input: 206.05 toks/s, output: 2756.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:33<00:20,  2.20it/s, est. speed input: 207.01 toks/s, output: 2948.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:33<00:17,  2.55it/s, est. speed input: 207.58 toks/s, output: 3045.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:36<00:38,  1.11it/s, est. speed input: 195.90 toks/s, output: 2964.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:36<00:28,  1.45it/s, est. speed input: 196.86 toks/s, output: 3068.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:38<00:47,  1.16s/it, est. speed input: 186.30 toks/s, output: 2994.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:45,  1.13s/it, est. speed input: 182.90 toks/s, output: 3030.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:40<00:27,  1.39it/s, est. speed input: 183.96 toks/s, output: 3227.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:40<00:21,  1.75it/s, est. speed input: 185.03 toks/s, output: 3334.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5418.8 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:40<00:18,  1.92it/s, est. speed input: 184.75 toks/s, output: 3418.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:42<00:32,  1.07it/s, est. speed input: 177.28 toks/s, output: 3370.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:43<00:25,  1.33it/s, est. speed input: 177.57 toks/s, output: 3465.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:43<00:22,  1.45it/s, est. speed input: 176.81 toks/s, output: 3539.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:45<00:35,  1.10s/it, est. speed input: 169.96 toks/s, output: 3493.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:46<00:28,  1.09it/s, est. speed input: 169.59 toks/s, output: 3574.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:17,  1.69it/s, est. speed input: 170.68 toks/s, output: 3777.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:13,  2.04it/s, est. speed input: 170.80 toks/s, output: 3958.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:10,  2.40it/s, est. speed input: 171.45 toks/s, output: 4063.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:47<00:09,  2.77it/s, est. speed input: 172.04 toks/s, output: 4165.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:48<00:13,  1.79it/s, est. speed input: 169.21 toks/s, output: 4186.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:50<00:19,  1.16it/s, est. speed input: 164.75 toks/s, output: 4164.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3990.0 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.3%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:50<00:14,  1.47it/s, est. speed input: 165.24 toks/s, output: 4265.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:50<00:10,  1.92it/s, est. speed input: 166.05 toks/s, output: 4375.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:51<00:09,  2.13it/s, est. speed input: 166.09 toks/s, output: 4464.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:51<00:09,  2.05it/s, est. speed input: 165.54 toks/s, output: 4537.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:54<00:20,  1.13s/it, est. speed input: 158.47 toks/s, output: 4434.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:56<00:24,  1.45s/it, est. speed input: 153.31 toks/s, output: 4383.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:56<00:18,  1.13s/it, est. speed input: 153.42 toks/s, output: 4478.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:59<00:23,  1.54s/it, est. speed input: 147.95 toks/s, output: 4413.82 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2905.3 tokens/s, Running: 15 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 76.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:00<00:19,  1.42s/it, est. speed input: 146.20 toks/s, output: 4458.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:03<00:24,  1.91s/it, est. speed input: 140.12 toks/s, output: 4372.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:03<00:17,  1.42s/it, est. speed input: 140.41 toks/s, output: 4480.59 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  1.42s/it, est. speed input: 151.81 toks/s, output: 6018.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:03<00:00,  2.50it/s, est. speed input: 151.81 toks/s, output: 6018.48 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:36 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:36 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:50:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:50:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:50:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:50:37 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:50:37 [abstract.py:306] It took 0.553095 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:50:37 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:50:37 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:50:37 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:50:37 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0098
+  • Training Accuracy: 0.3063
+  • Mean Generation Length: 2404.3999
+  • Training FLOPS: 44.35 TFLOPS (11.09 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 305.56s
+  • policy_training: 197.31s (64.6%)
+  • generation: 67.95s (22.2%)
+  • teacher_logprob_inference: 35.16s (11.5%)
+  • policy_training/sharding_data: 4.32s (1.4%)
+  • prepare_for_generation: 3.61s (1.2%)
+  • policy_training/submit_training_futures: 2.12s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.57s (0.5%)
+  • training_prep: 0.80s (0.3%)
+  • teacher_logprob_inference_prep: 0.44s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 184: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:54:32 [abstract.py:324] It took 0.049415 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09236562e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09236562e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09236562e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09236562e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09236016e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69358253e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83443527e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09236016e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69358253e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83443527e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09236016e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69358253e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83443527e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09236016e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69358253e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83443527e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68207321e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82603149e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68207321e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82603149e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68207321e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82603149e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68207321e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82603149e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00683231e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.80759907e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #184] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44809723e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63717880e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.41186275e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #184] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.52889099e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37176208e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-8.10427904e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #184] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.47135391e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.17763824e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.63247986e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #184] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36776855e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 184: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:54:35 [abstract.py:324] It took 0.028816 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13368.83it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:54:35 [loggers.py:236] Engine 000: Avg prompt throughput: 39.8 tokens/s, Avg generation throughput: 32.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:53,  2.96it/s, est. speed input: 180.77 toks/s, output: 35.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:11, 13.75it/s, est. speed input: 679.49 toks/s, output: 176.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 27.28it/s, est. speed input: 1290.50 toks/s, output: 497.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 29.69it/s, est. speed input: 1432.37 toks/s, output: 673.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 28.00it/s, est. speed input: 1459.50 toks/s, output: 829.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:01<00:04, 31.20it/s, est. speed input: 1577.65 toks/s, output: 1086.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:04, 29.41it/s, est. speed input: 1578.70 toks/s, output: 1239.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 33.89it/s, est. speed input: 1690.20 toks/s, output: 1540.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:03, 32.56it/s, est. speed input: 1697.78 toks/s, output: 1709.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:03, 33.60it/s, est. speed input: 1735.83 toks/s, output: 1921.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 42.17it/s, est. speed input: 1881.35 toks/s, output: 2390.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 40.15it/s, est. speed input: 1912.39 toks/s, output: 2683.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 37.30it/s, est. speed input: 1918.55 toks/s, output: 2942.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:02<00:02, 33.58it/s, est. speed input: 1893.24 toks/s, output: 3096.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:03<00:09,  9.18it/s, est. speed input: 1223.31 toks/s, output: 2228.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:07<00:30,  2.88it/s, est. speed input: 621.59 toks/s, output: 1349.84 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:08<00:36,  2.32it/s, est. speed input: 517.97 toks/s, output: 1313.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:09<00:33,  2.51it/s, est. speed input: 501.35 toks/s, output: 1455.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:09<00:26,  3.02it/s, est. speed input: 493.69 toks/s, output: 1701.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:10<00:24,  3.26it/s, est. speed input: 493.56 toks/s, output: 1789.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:54:45 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10707.7 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:11<00:40,  1.94it/s, est. speed input: 427.12 toks/s, output: 1640.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<00:58,  1.32it/s, est. speed input: 372.24 toks/s, output: 1524.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<00:55,  1.37it/s, est. speed input: 360.31 toks/s, output: 1567.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:15<00:56,  1.34it/s, est. speed input: 345.10 toks/s, output: 1594.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:15<00:44,  1.66it/s, est. speed input: 345.83 toks/s, output: 1689.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:15<00:35,  2.06it/s, est. speed input: 346.63 toks/s, output: 1784.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:16<00:44,  1.61it/s, est. speed input: 329.11 toks/s, output: 1786.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:17<00:42,  1.64it/s, est. speed input: 313.83 toks/s, output: 1880.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:18<00:41,  1.67it/s, est. speed input: 307.64 toks/s, output: 1932.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:19<00:59,  1.15it/s, est. speed input: 284.52 toks/s, output: 1877.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:54:55 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8149.2 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.0%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:20<00:51,  1.31it/s, est. speed input: 280.99 toks/s, output: 1945.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:21<01:06,  1.01s/it, est. speed input: 262.10 toks/s, output: 1906.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:22<01:00,  1.07it/s, est. speed input: 256.23 toks/s, output: 1954.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:22<00:35,  1.78it/s, est. speed input: 259.22 toks/s, output: 2156.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:23<00:31,  1.94it/s, est. speed input: 257.72 toks/s, output: 2232.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:23<00:28,  2.12it/s, est. speed input: 256.50 toks/s, output: 2309.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:24<00:30,  1.94it/s, est. speed input: 252.36 toks/s, output: 2359.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:25<00:36,  1.63it/s, est. speed input: 245.95 toks/s, output: 2387.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:26<00:48,  1.19it/s, est. speed input: 235.30 toks/s, output: 2371.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:27<00:43,  1.30it/s, est. speed input: 232.42 toks/s, output: 2430.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:27<00:37,  1.48it/s, est. speed input: 230.95 toks/s, output: 2502.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:27<00:30,  1.83it/s, est. speed input: 231.14 toks/s, output: 2592.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:28<00:25,  2.12it/s, est. speed input: 230.85 toks/s, output: 2676.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:28<00:29,  1.82it/s, est. speed input: 227.00 toks/s, output: 2719.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:29<00:24,  2.09it/s, est. speed input: 226.56 toks/s, output: 2800.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:29<00:22,  2.28it/s, est. speed input: 225.98 toks/s, output: 2879.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:05 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6774.4 tokens/s, Running: 51 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:30<00:34,  1.46it/s, est. speed input: 218.70 toks/s, output: 2873.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:28,  1.70it/s, est. speed input: 218.09 toks/s, output: 2950.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:31<00:21,  2.22it/s, est. speed input: 219.15 toks/s, output: 3050.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:46,  1.00it/s, est. speed input: 205.99 toks/s, output: 2954.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:34,  1.35it/s, est. speed input: 206.92 toks/s, output: 3055.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:34<00:41,  1.08it/s, est. speed input: 200.63 toks/s, output: 3049.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:35<00:31,  1.41it/s, est. speed input: 201.28 toks/s, output: 3145.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:35<00:28,  1.49it/s, est. speed input: 199.67 toks/s, output: 3207.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:36<00:22,  1.84it/s, est. speed input: 200.08 toks/s, output: 3299.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:36<00:21,  1.93it/s, est. speed input: 199.23 toks/s, output: 3370.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:38<00:40,  1.02s/it, est. speed input: 189.55 toks/s, output: 3293.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:38<00:21,  1.74it/s, est. speed input: 192.12 toks/s, output: 3510.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:38<00:17,  2.16it/s, est. speed input: 193.08 toks/s, output: 3613.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:39<00:17,  2.03it/s, est. speed input: 191.85 toks/s, output: 3674.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:15 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5371.3 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:40<00:20,  1.68it/s, est. speed input: 189.23 toks/s, output: 3710.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:42<00:37,  1.10s/it, est. speed input: 179.92 toks/s, output: 3614.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:42<00:27,  1.20it/s, est. speed input: 180.67 toks/s, output: 3716.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:43<00:26,  1.21it/s, est. speed input: 178.67 toks/s, output: 3762.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:45<00:31,  1.00s/it, est. speed input: 174.39 toks/s, output: 3759.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:23,  1.29it/s, est. speed input: 174.79 toks/s, output: 3857.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:46<00:24,  1.19it/s, est. speed input: 172.35 toks/s, output: 3891.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:47<00:23,  1.21it/s, est. speed input: 170.70 toks/s, output: 3942.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:47<00:18,  1.45it/s, est. speed input: 170.67 toks/s, output: 4029.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:14,  1.83it/s, est. speed input: 171.24 toks/s, output: 4129.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:48<00:12,  1.98it/s, est. speed input: 171.00 toks/s, output: 4212.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:50<00:22,  1.09it/s, est. speed input: 165.74 toks/s, output: 4171.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:25 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4164.9 tokens/s, Running: 24 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.6%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:23,  1.03s/it, est. speed input: 162.78 toks/s, output: 4186.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:51<00:17,  1.24it/s, est. speed input: 163.12 toks/s, output: 4282.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:51<00:09,  2.10it/s, est. speed input: 164.97 toks/s, output: 4506.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:11,  1.55it/s, est. speed input: 161.62 toks/s, output: 4591.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:54<00:07,  2.10it/s, est. speed input: 162.60 toks/s, output: 4882.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:54<00:05,  2.40it/s, est. speed input: 163.20 toks/s, output: 4990.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:54<00:04,  2.67it/s, est. speed input: 163.66 toks/s, output: 5093.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:55<00:06,  1.86it/s, est. speed input: 161.54 toks/s, output: 5117.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:56<00:05,  1.87it/s, est. speed input: 160.63 toks/s, output: 5268.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:58<00:05,  1.51it/s, est. speed input: 158.64 toks/s, output: 5293.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:58<00:04,  1.81it/s, est. speed input: 159.12 toks/s, output: 5401.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:59<00:04,  1.63it/s, est. speed input: 158.02 toks/s, output: 5455.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:35 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2416.1 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.5%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:01<00:06,  1.05s/it, est. speed input: 153.21 toks/s, output: 5385.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:01<00:04,  1.20it/s, est. speed input: 153.51 toks/s, output: 5491.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [01:01<00:02,  1.42it/s, est. speed input: 153.61 toks/s, output: 5589.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [01:02<00:01,  1.60it/s, est. speed input: 153.53 toks/s, output: 5681.75 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  1.60it/s, est. speed input: 156.46 toks/s, output: 6075.70 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  2.56it/s, est. speed input: 156.46 toks/s, output: 6075.70 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:41 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:41 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 13:55:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:55:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:55:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:55:41 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 13:55:41 [abstract.py:306] It took 0.525819 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 13:55:41 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 13:55:41 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 13:55:41 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 13:55:41 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0104
+  • Training Accuracy: 0.2875
+  • Mean Generation Length: 2368.7876
+  • Training FLOPS: 44.41 TFLOPS (11.10 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 296.60s
+  • policy_training: 192.60s (64.9%)
+  • generation: 66.33s (22.4%)
+  • teacher_logprob_inference: 32.60s (11.0%)
+  • policy_training/sharding_data: 4.32s (1.5%)
+  • prepare_for_generation: 3.70s (1.2%)
+  • policy_training/submit_training_futures: 1.91s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.5%)
+  • training_prep: 0.79s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 185: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 13:59:29 [abstract.py:324] It took 0.051662 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09260938e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68038788e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81822968e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09260938e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68038788e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81822968e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09260938e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68038788e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81822968e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09260938e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68038788e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.81822968e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09265508e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69292717e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82614288e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00962849e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.74988317e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #185] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44846573e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09265508e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69292717e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82614288e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66739044e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.39774008e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #185] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.52700195e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09265508e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69292717e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82614288e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37178802e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.48406172e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #185] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.49015808e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09265508e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69292717e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82614288e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.19278870e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.64448051e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #185] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36566895e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 185: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 13:59:32 [abstract.py:324] It took 0.028398 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 13:59:32 [loggers.py:236] Engine 000: Avg prompt throughput: 41.0 tokens/s, Avg generation throughput: 12.6 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12976.17it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:46,  3.39it/s, est. speed input: 210.25 toks/s, output: 50.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 21.37it/s, est. speed input: 1059.68 toks/s, output: 363.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 29.46it/s, est. speed input: 1457.83 toks/s, output: 614.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 31.53it/s, est. speed input: 1585.03 toks/s, output: 788.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 32.42it/s, est. speed input: 1692.67 toks/s, output: 1023.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 34.40it/s, est. speed input: 1786.83 toks/s, output: 1280.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 39.28it/s, est. speed input: 1936.39 toks/s, output: 1637.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 43.34it/s, est. speed input: 2060.78 toks/s, output: 1989.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 45/160 [00:01<00:02, 41.42it/s, est. speed input: 2082.28 toks/s, output: 2216.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 43.58it/s, est. speed input: 2160.03 toks/s, output: 2551.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 36.49it/s, est. speed input: 2090.38 toks/s, output: 2721.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:02, 34.40it/s, est. speed input: 2065.20 toks/s, output: 2931.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:03<00:13,  7.22it/s, est. speed input: 1069.11 toks/s, output: 1768.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:04<00:16,  5.62it/s, est. speed input: 885.07 toks/s, output: 1701.43 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:07<00:31,  2.82it/s, est. speed input: 595.16 toks/s, output: 1296.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:08<00:38,  2.29it/s, est. speed input: 502.33 toks/s, output: 1268.23 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:59:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10774.6 tokens/s, Running: 87 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:10<00:49,  1.75it/s, est. speed input: 433.01 toks/s, output: 1189.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:11<00:54,  1.56it/s, est. speed input: 398.84 toks/s, output: 1190.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:13<00:57,  1.43it/s, est. speed input: 357.13 toks/s, output: 1245.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:13<00:55,  1.47it/s, est. speed input: 346.11 toks/s, output: 1302.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:14<00:50,  1.60it/s, est. speed input: 330.18 toks/s, output: 1422.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:15<00:51,  1.53it/s, est. speed input: 317.64 toks/s, output: 1461.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:16<00:41,  1.86it/s, est. speed input: 312.12 toks/s, output: 1617.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:16<00:35,  2.14it/s, est. speed input: 312.27 toks/s, output: 1708.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:18<00:59,  1.27it/s, est. speed input: 282.93 toks/s, output: 1639.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:18<00:50,  1.48it/s, est. speed input: 281.23 toks/s, output: 1719.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:19<00:46,  1.58it/s, est. speed input: 276.97 toks/s, output: 1783.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:19<00:37,  1.90it/s, est. speed input: 276.73 toks/s, output: 1870.33 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 13:59:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8511.3 tokens/s, Running: 71 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:21<00:49,  1.40it/s, est. speed input: 257.35 toks/s, output: 1908.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:22<00:45,  1.50it/s, est. speed input: 249.36 toks/s, output: 2020.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:23<00:30,  2.15it/s, est. speed input: 249.80 toks/s, output: 2280.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:23<00:22,  2.77it/s, est. speed input: 252.46 toks/s, output: 2473.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:23<00:16,  3.73it/s, est. speed input: 256.56 toks/s, output: 2679.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:24<00:14,  4.05it/s, est. speed input: 257.21 toks/s, output: 2927.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:24<00:12,  4.59it/s, est. speed input: 259.30 toks/s, output: 3110.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:24<00:12,  4.39it/s, est. speed input: 258.85 toks/s, output: 3184.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:25<00:15,  3.50it/s, est. speed input: 255.81 toks/s, output: 3226.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:13,  3.82it/s, est. speed input: 256.12 toks/s, output: 3388.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:11,  4.28it/s, est. speed input: 257.27 toks/s, output: 3560.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:11,  4.11it/s, est. speed input: 256.72 toks/s, output: 3707.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:26<00:10,  4.44it/s, est. speed input: 257.64 toks/s, output: 3797.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:27<00:12,  3.65it/s, est. speed input: 255.50 toks/s, output: 3843.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:27<00:10,  4.10it/s, est. speed input: 256.35 toks/s, output: 3934.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:29<00:22,  1.86it/s, est. speed input: 240.90 toks/s, output: 3919.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:30<00:20,  2.04it/s, est. speed input: 240.62 toks/s, output: 3993.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:00:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6497.8 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:30<00:19,  2.07it/s, est. speed input: 238.97 toks/s, output: 4045.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:30<00:15,  2.50it/s, est. speed input: 239.95 toks/s, output: 4141.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:31<00:16,  2.24it/s, est. speed input: 237.35 toks/s, output: 4176.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:18,  1.99it/s, est. speed input: 234.33 toks/s, output: 4204.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:17,  2.01it/s, est. speed input: 232.69 toks/s, output: 4254.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:32<00:14,  2.37it/s, est. speed input: 232.91 toks/s, output: 4339.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:32<00:09,  3.54it/s, est. speed input: 235.04 toks/s, output: 4537.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:15,  2.11it/s, est. speed input: 229.37 toks/s, output: 4508.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:34<00:13,  2.15it/s, est. speed input: 226.87 toks/s, output: 4619.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:34<00:08,  3.18it/s, est. speed input: 229.66 toks/s, output: 4837.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:35<00:07,  3.57it/s, est. speed input: 230.46 toks/s, output: 4932.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:35<00:08,  3.03it/s, est. speed input: 228.93 toks/s, output: 4981.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:35<00:08,  2.86it/s, est. speed input: 228.01 toks/s, output: 5041.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:07,  3.24it/s, est. speed input: 228.49 toks/s, output: 5132.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:36<00:05,  3.86it/s, est. speed input: 229.40 toks/s, output: 5232.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:36<00:04,  4.62it/s, est. speed input: 230.38 toks/s, output: 5334.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:36<00:05,  3.80it/s, est. speed input: 229.66 toks/s, output: 5397.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:37<00:05,  3.79it/s, est. speed input: 229.63 toks/s, output: 5478.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:37<00:05,  3.25it/s, est. speed input: 228.69 toks/s, output: 5536.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:38<00:08,  2.25it/s, est. speed input: 225.67 toks/s, output: 5545.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:00:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4240.2 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:14,  1.15it/s, est. speed input: 216.65 toks/s, output: 5410.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:40<00:10,  1.54it/s, est. speed input: 217.41 toks/s, output: 5515.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:40<00:08,  1.71it/s, est. speed input: 216.59 toks/s, output: 5579.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:08,  1.68it/s, est. speed input: 214.79 toks/s, output: 5620.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:41<00:05,  2.14it/s, est. speed input: 214.48 toks/s, output: 5783.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:43<00:07,  1.57it/s, est. speed input: 210.12 toks/s, output: 5755.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:08,  1.18it/s, est. speed input: 204.65 toks/s, output: 5697.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:44<00:06,  1.34it/s, est. speed input: 203.82 toks/s, output: 5766.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:45<00:05,  1.40it/s, est. speed input: 202.27 toks/s, output: 5817.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:45<00:04,  1.69it/s, est. speed input: 202.34 toks/s, output: 5913.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:46<00:04,  1.39it/s, est. speed input: 199.19 toks/s, output: 5917.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:48<00:04,  1.05it/s, est. speed input: 194.20 toks/s, output: 5868.28 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:00:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1872.1 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:50<00:05,  1.38s/it, est. speed input: 186.31 toks/s, output: 5734.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:51<00:03,  1.07s/it, est. speed input: 186.19 toks/s, output: 5836.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:54<00:03,  1.81s/it, est. speed input: 175.24 toks/s, output: 5608.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.81s/it, est. speed input: 177.23 toks/s, output: 5900.09 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.92it/s, est. speed input: 177.23 toks/s, output: 5900.09 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:00:30 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:00:30 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:00:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:00:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:00:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:00:30 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:00:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:00:30 [abstract.py:306] It took 0.554671 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:00:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:00:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:00:30 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0116
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 2020.7000
+  • Training FLOPS: 41.26 TFLOPS (10.32 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 268.59s
+  • policy_training: 173.30s (64.5%)
+  • generation: 58.36s (21.7%)
+  • teacher_logprob_inference: 31.65s (11.8%)
+  • policy_training/sharding_data: 4.32s (1.6%)
+  • prepare_for_generation: 3.70s (1.4%)
+  • policy_training/submit_training_futures: 1.89s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.6%)
+  • training_prep: 0.99s (0.4%)
+  • teacher_logprob_inference_prep: 0.40s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Epoch 38/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 186: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:03:57 [abstract.py:324] It took 0.053766 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270703e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68536606e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80383606e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270703e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68536606e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80383606e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270703e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68536606e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80383606e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09270703e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68536606e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80383606e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09261504e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70003967e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78934174e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03648891e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.66260433e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #186] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44750854e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09261504e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70003967e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78934174e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68549805e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38589382e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #186] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.49247208e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09261504e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70003967e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78934174e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36477203e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.76271677e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #186] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.51279221e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09261504e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70003967e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78934174e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20159912e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.64616776e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #186] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36447433e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 186: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:01 [abstract.py:324] It took 0.030492 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14006.69it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:01 [loggers.py:236] Engine 000: Avg prompt throughput: 44.0 tokens/s, Avg generation throughput: 18.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:39,  4.03it/s, est. speed input: 246.02 toks/s, output: 56.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:17,  9.00it/s, est. speed input: 480.95 toks/s, output: 165.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:10, 15.06it/s, est. speed input: 740.14 toks/s, output: 331.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 23.47it/s, est. speed input: 1093.16 toks/s, output: 642.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:07, 19.05it/s, est. speed input: 1022.55 toks/s, output: 715.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 31.63it/s, est. speed input: 1392.84 toks/s, output: 1302.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:01<00:03, 42.29it/s, est. speed input: 1695.72 toks/s, output: 1845.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 51.54it/s, est. speed input: 1954.35 toks/s, output: 2381.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:01, 57.59it/s, est. speed input: 2156.10 toks/s, output: 2878.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 57.76it/s, est. speed input: 2269.20 toks/s, output: 3249.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 72.73it/s, est. speed input: 2581.76 toks/s, output: 4104.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:01<00:01, 47.41it/s, est. speed input: 2407.34 toks/s, output: 4135.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:04<00:08,  9.26it/s, est. speed input: 1142.04 toks/s, output: 2285.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:07<00:17,  4.12it/s, est. speed input: 651.41 toks/s, output: 1716.88 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10231.8 tokens/s, Running: 74 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:14<00:35,  1.99it/s, est. speed input: 381.21 toks/s, output: 1375.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:15<00:33,  2.00it/s, est. speed input: 357.35 toks/s, output: 1583.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:16<00:32,  2.01it/s, est. speed input: 343.55 toks/s, output: 1722.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:17<00:28,  2.19it/s, est. speed input: 340.60 toks/s, output: 1903.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:17<00:26,  2.33it/s, est. speed input: 339.95 toks/s, output: 1997.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:17<00:25,  2.36it/s, est. speed input: 336.04 toks/s, output: 2071.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:23,  2.58it/s, est. speed input: 335.58 toks/s, output: 2166.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:19<00:37,  1.57it/s, est. speed input: 309.15 toks/s, output: 2093.69 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:21 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7882.0 tokens/s, Running: 58 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:33,  1.75it/s, est. speed input: 307.14 toks/s, output: 2178.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:32,  1.76it/s, est. speed input: 301.67 toks/s, output: 2237.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:28,  1.94it/s, est. speed input: 299.53 toks/s, output: 2317.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:21<00:24,  2.28it/s, est. speed input: 299.09 toks/s, output: 2410.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:41,  1.31it/s, est. speed input: 280.30 toks/s, output: 2356.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:38,  1.39it/s, est. speed input: 275.60 toks/s, output: 2413.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:23<00:34,  1.52it/s, est. speed input: 272.31 toks/s, output: 2482.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:24<00:29,  1.71it/s, est. speed input: 270.22 toks/s, output: 2558.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:25<00:34,  1.43it/s, est. speed input: 257.49 toks/s, output: 2624.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:28,  1.69it/s, est. speed input: 257.26 toks/s, output: 2718.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:27<00:33,  1.39it/s, est. speed input: 249.24 toks/s, output: 2730.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:27<00:28,  1.62it/s, est. speed input: 248.33 toks/s, output: 2816.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:28<00:25,  1.78it/s, est. speed input: 246.75 toks/s, output: 2893.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:28<00:19,  2.30it/s, est. speed input: 247.86 toks/s, output: 3000.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:28<00:19,  2.15it/s, est. speed input: 245.29 toks/s, output: 3063.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:28<00:16,  2.50it/s, est. speed input: 245.29 toks/s, output: 3157.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:29<00:19,  2.13it/s, est. speed input: 242.07 toks/s, output: 3210.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:12,  3.01it/s, est. speed input: 243.24 toks/s, output: 3411.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6190.2 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:14,  2.62it/s, est. speed input: 241.03 toks/s, output: 3472.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:30<00:11,  3.23it/s, est. speed input: 241.62 toks/s, output: 3665.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:31<00:13,  2.68it/s, est. speed input: 239.08 toks/s, output: 3717.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:32<00:20,  1.66it/s, est. speed input: 231.42 toks/s, output: 3691.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:32<00:15,  2.08it/s, est. speed input: 232.28 toks/s, output: 3797.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:33<00:19,  1.63it/s, est. speed input: 227.44 toks/s, output: 3810.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:34<00:15,  1.94it/s, est. speed input: 227.48 toks/s, output: 3903.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:19,  1.53it/s, est. speed input: 222.71 toks/s, output: 3915.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:35<00:19,  1.48it/s, est. speed input: 219.86 toks/s, output: 3959.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:37<00:23,  1.17it/s, est. speed input: 213.83 toks/s, output: 3946.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:37<00:18,  1.44it/s, est. speed input: 213.62 toks/s, output: 4037.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:16,  1.60it/s, est. speed input: 212.63 toks/s, output: 4114.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:38<00:13,  1.79it/s, est. speed input: 211.93 toks/s, output: 4196.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:38<00:11,  2.14it/s, est. speed input: 212.08 toks/s, output: 4294.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:38<00:09,  2.38it/s, est. speed input: 211.89 toks/s, output: 4385.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:39<00:10,  2.09it/s, est. speed input: 210.11 toks/s, output: 4442.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:09,  2.19it/s, est. speed input: 209.48 toks/s, output: 4523.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:40<00:05,  3.69it/s, est. speed input: 211.94 toks/s, output: 4762.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4326.8 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:40<00:05,  3.38it/s, est. speed input: 211.44 toks/s, output: 4934.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:41<00:04,  3.63it/s, est. speed input: 211.90 toks/s, output: 5129.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:41<00:03,  3.97it/s, est. speed input: 212.50 toks/s, output: 5236.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:03,  3.48it/s, est. speed input: 211.88 toks/s, output: 5313.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:09,  1.29it/s, est. speed input: 202.31 toks/s, output: 5169.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:07,  1.48it/s, est. speed input: 201.90 toks/s, output: 5255.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:45<00:08,  1.18it/s, est. speed input: 197.40 toks/s, output: 5238.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:47<00:11,  1.23s/it, est. speed input: 189.61 toks/s, output: 5134.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:49<00:06,  1.07it/s, est. speed input: 187.69 toks/s, output: 5288.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2238.8 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:51<00:07,  1.22s/it, est. speed input: 181.04 toks/s, output: 5210.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:51<00:03,  1.32it/s, est. speed input: 182.86 toks/s, output: 5479.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:52<00:02,  1.15it/s, est. speed input: 179.67 toks/s, output: 5495.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:54<00:02,  1.19s/it, est. speed input: 173.54 toks/s, output: 5424.02 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  1.19s/it, est. speed input: 175.67 toks/s, output: 5722.82 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:54<00:00,  2.92it/s, est. speed input: 175.67 toks/s, output: 5722.82 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:58 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:58 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:04:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:04:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:04:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:04:58 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:04:59 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:04:59 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:04:59 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:04:59 [abstract.py:306] It took 0.544422 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:04:59 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0110
+  • Training Accuracy: 0.5563
+  • Mean Generation Length: 1961.1312
+  • Training FLOPS: 40.11 TFLOPS (10.03 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 268.64s
+  • policy_training: 174.98s (65.1%)
+  • generation: 58.30s (21.7%)
+  • teacher_logprob_inference: 29.85s (11.1%)
+  • policy_training/sharding_data: 4.33s (1.6%)
+  • prepare_for_generation: 3.79s (1.4%)
+  • policy_training/submit_training_futures: 2.12s (0.8%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.6%)
+  • training_prep: 0.85s (0.3%)
+  • teacher_logprob_inference_prep: 0.68s (0.3%)
+  • get_topk_logits/submit_topk_logits_futures: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 187: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:08:26 [abstract.py:324] It took 0.050618 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09267461e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09267461e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09267461e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09267461e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09267578e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70580902e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276978e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09267578e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70580902e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276978e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09267578e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70580902e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276978e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09267578e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.70580902e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.78276978e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68691025e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80010147e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68691025e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80010147e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68691025e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80010147e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68691025e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80010147e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03646545e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.65850735e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #187] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45082245e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.70014114e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37810440e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #187] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.50999146e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35681610e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.63290834e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #187] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.49746094e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20456390e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65839520e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #187] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36613663e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 187: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:08:29 [abstract.py:324] It took 0.030730 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13889.86it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:08:29 [loggers.py:236] Engine 000: Avg prompt throughput: 44.8 tokens/s, Avg generation throughput: 22.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:46,  3.45it/s, est. speed input: 213.96 toks/s, output: 37.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:07, 19.79it/s, est. speed input: 1010.72 toks/s, output: 275.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:09, 15.95it/s, est. speed input: 917.26 toks/s, output: 368.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  10%|█         | 16/160 [00:00<00:05, 25.88it/s, est. speed input: 1269.19 toks/s, output: 740.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 31.23it/s, est. speed input: 1477.51 toks/s, output: 1047.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 41.29it/s, est. speed input: 1781.84 toks/s, output: 1539.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 41.33it/s, est. speed input: 1861.84 toks/s, output: 1767.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:01<00:02, 48.36it/s, est. speed input: 2080.91 toks/s, output: 2278.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 51.16it/s, est. speed input: 2218.73 toks/s, output: 2711.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:02, 47.18it/s, est. speed input: 2235.96 toks/s, output: 2987.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 52.25it/s, est. speed input: 2375.32 toks/s, output: 3534.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 53.50it/s, est. speed input: 2440.07 toks/s, output: 3886.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:02<00:04, 18.34it/s, est. speed input: 1766.54 toks/s, output: 3121.11 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:08:39 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10288.4 tokens/s, Running: 77 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:32,  2.37it/s, est. speed input: 487.94 toks/s, output: 1129.99 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:13<00:37,  1.92it/s, est. speed input: 394.60 toks/s, output: 1191.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:14<00:35,  1.96it/s, est. speed input: 369.51 toks/s, output: 1406.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<00:34,  1.96it/s, est. speed input: 353.65 toks/s, output: 1544.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:17<00:38,  1.73it/s, est. speed input: 325.69 toks/s, output: 1621.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:34,  1.88it/s, est. speed input: 325.74 toks/s, output: 1720.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:19<00:40,  1.58it/s, est. speed input: 307.52 toks/s, output: 1723.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:19<00:35,  1.79it/s, est. speed input: 307.65 toks/s, output: 1822.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:19<00:36,  1.72it/s, est. speed input: 300.24 toks/s, output: 1877.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:08:49 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7977.6 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:20<00:34,  1.78it/s, est. speed input: 296.09 toks/s, output: 1948.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:28,  2.13it/s, est. speed input: 296.59 toks/s, output: 2047.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:21<00:35,  1.68it/s, est. speed input: 286.18 toks/s, output: 2071.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:21<00:27,  2.13it/s, est. speed input: 287.27 toks/s, output: 2176.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:22<00:25,  2.21it/s, est. speed input: 281.81 toks/s, output: 2321.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:23<00:34,  1.61it/s, est. speed input: 270.52 toks/s, output: 2323.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:23<00:26,  2.03it/s, est. speed input: 271.87 toks/s, output: 2430.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:24<00:21,  2.43it/s, est. speed input: 272.37 toks/s, output: 2527.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:35,  1.48it/s, est. speed input: 260.05 toks/s, output: 2508.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:26,  1.92it/s, est. speed input: 261.16 toks/s, output: 2613.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:25<00:22,  2.20it/s, est. speed input: 260.57 toks/s, output: 2701.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:26<00:18,  2.62it/s, est. speed input: 260.81 toks/s, output: 2796.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:26<00:17,  2.79it/s, est. speed input: 260.13 toks/s, output: 2882.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:26<00:12,  3.64it/s, est. speed input: 261.35 toks/s, output: 3077.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:27<00:12,  3.49it/s, est. speed input: 260.53 toks/s, output: 3157.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:27<00:10,  4.10it/s, est. speed input: 261.57 toks/s, output: 3260.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:28<00:22,  1.92it/s, est. speed input: 252.00 toks/s, output: 3232.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:28<00:21,  1.96it/s, est. speed input: 249.88 toks/s, output: 3296.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:29<00:22,  1.84it/s, est. speed input: 246.67 toks/s, output: 3344.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:13,  2.95it/s, est. speed input: 249.32 toks/s, output: 3561.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:08:59 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6262.1 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:15,  2.46it/s, est. speed input: 246.26 toks/s, output: 3608.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:30<00:12,  3.01it/s, est. speed input: 247.34 toks/s, output: 3712.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:30<00:08,  4.31it/s, est. speed input: 249.75 toks/s, output: 3926.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:30<00:07,  4.72it/s, est. speed input: 250.61 toks/s, output: 4027.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:31<00:13,  2.36it/s, est. speed input: 244.28 toks/s, output: 4012.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:32<00:12,  2.54it/s, est. speed input: 243.81 toks/s, output: 4093.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:32<00:13,  2.31it/s, est. speed input: 241.65 toks/s, output: 4146.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:33<00:13,  2.19it/s, est. speed input: 239.73 toks/s, output: 4202.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:33<00:10,  2.75it/s, est. speed input: 240.59 toks/s, output: 4307.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:33<00:12,  2.26it/s, est. speed input: 237.91 toks/s, output: 4348.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:34<00:09,  2.80it/s, est. speed input: 237.99 toks/s, output: 4525.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:35<00:10,  2.27it/s, est. speed input: 235.03 toks/s, output: 4559.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:36<00:17,  1.35it/s, est. speed input: 226.43 toks/s, output: 4484.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:37<00:17,  1.34it/s, est. speed input: 223.43 toks/s, output: 4518.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:38<00:19,  1.14it/s, est. speed input: 218.02 toks/s, output: 4500.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:39<00:17,  1.17it/s, est. speed input: 215.32 toks/s, output: 4537.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:39<00:14,  1.39it/s, est. speed input: 214.73 toks/s, output: 4617.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:09:09 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3978.1 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:41<00:16,  1.18it/s, est. speed input: 210.08 toks/s, output: 4612.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:41<00:12,  1.39it/s, est. speed input: 209.49 toks/s, output: 4694.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:41<00:09,  1.87it/s, est. speed input: 210.45 toks/s, output: 4810.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:42<00:08,  1.92it/s, est. speed input: 209.45 toks/s, output: 4882.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:42<00:06,  2.38it/s, est. speed input: 210.02 toks/s, output: 4989.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:44<00:14,  1.03s/it, est. speed input: 199.82 toks/s, output: 4844.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:45<00:10,  1.24it/s, est. speed input: 199.89 toks/s, output: 4945.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:45<00:07,  1.57it/s, est. speed input: 200.15 toks/s, output: 5051.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:49<00:17,  1.57s/it, est. speed input: 186.03 toks/s, output: 4799.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:49<00:11,  1.13s/it, est. speed input: 186.91 toks/s, output: 4925.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:09:19 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2592.6 tokens/s, Running: 10 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.2%, Prefix cache hit rate: 74.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:51<00:12,  1.44s/it, est. speed input: 180.28 toks/s, output: 4857.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:51<00:09,  1.20s/it, est. speed input: 179.18 toks/s, output: 4936.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:53<00:09,  1.32s/it, est. speed input: 174.93 toks/s, output: 4931.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:08,  1.42s/it, est. speed input: 170.80 toks/s, output: 4927.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:56<00:06,  1.29s/it, est. speed input: 168.93 toks/s, output: 4987.34 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.29s/it, est. speed input: 174.41 toks/s, output: 5716.27 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.85it/s, est. speed input: 174.41 toks/s, output: 5716.27 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:09:28 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:09:28 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:09:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:09:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:09:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:09:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:09:29 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:09:29 [abstract.py:306] It took 0.518389 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:09:29 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:09:29 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:09:29 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0120
+  • Training Accuracy: 0.4437
+  • Mean Generation Length: 2007.4938
+  • Training FLOPS: 42.61 TFLOPS (10.65 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 264.25s
+  • policy_training: 169.22s (64.0%)
+  • generation: 59.61s (22.6%)
+  • teacher_logprob_inference: 30.27s (11.5%)
+  • policy_training/sharding_data: 4.32s (1.6%)
+  • prepare_for_generation: 3.69s (1.4%)
+  • prepare_for_generation/transfer_and_update_weights: 1.60s (0.6%)
+  • policy_training/submit_training_futures: 1.58s (0.6%)
+  • training_prep: 0.81s (0.3%)
+  • teacher_logprob_inference_prep: 0.47s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 188: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:12:51 [abstract.py:324] It took 0.049916 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09266777e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67422485e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79660645e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09266777e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67422485e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79660645e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09266777e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67422485e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79660645e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09266777e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67422485e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79660645e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09263105e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69408417e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79522400e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.04008942e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.73379803e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #188] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45318359e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09263105e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69408417e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79522400e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69012604e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37767410e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #188] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.53635406e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09263105e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69408417e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79522400e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37150116e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.55530787e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #188] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.50186195e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09263105e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69408417e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79522400e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.20526733e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67843781e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #188] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36380585e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 188: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:12:54 [abstract.py:324] It took 0.029013 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13457.30it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 14:12:54 [loggers.py:236] Engine 000: Avg prompt throughput: 45.1 tokens/s, Avg generation throughput: 47.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:18,  8.83it/s, est. speed input: 529.74 toks/s, output: 70.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:05, 29.43it/s, est. speed input: 1594.41 toks/s, output: 373.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:05, 28.18it/s, est. speed input: 1605.12 toks/s, output: 455.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:05, 27.57it/s, est. speed input: 1616.52 toks/s, output: 607.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:05, 25.70it/s, est. speed input: 1569.46 toks/s, output: 774.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:04, 33.69it/s, est. speed input: 1825.03 toks/s, output: 1176.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 34.64it/s, est. speed input: 1877.56 toks/s, output: 1394.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:00<00:03, 38.32it/s, est. speed input: 1986.82 toks/s, output: 1692.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:02, 49.42it/s, est. speed input: 2245.54 toks/s, output: 2279.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 50/160 [00:01<00:02, 51.59it/s, est. speed input: 2399.72 toks/s, output: 2842.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:02, 51.08it/s, est. speed input: 2452.01 toks/s, output: 3176.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 53.10it/s, est. speed input: 2530.34 toks/s, output: 3551.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 52.38it/s, est. speed input: 2570.96 toks/s, output: 3882.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:03<00:07, 10.85it/s, est. speed input: 1393.34 toks/s, output: 2350.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:13:04 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10448.2 tokens/s, Running: 83 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.4%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:11<00:42,  1.92it/s, est. speed input: 412.47 toks/s, output: 941.36 toks/s]  
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:16<00:57,  1.38it/s, est. speed input: 302.53 toks/s, output: 969.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:17<00:52,  1.46it/s, est. speed input: 293.10 toks/s, output: 1133.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:13:14 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8437.9 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.7%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:20<01:04,  1.16it/s, est. speed input: 251.30 toks/s, output: 1162.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<01:05,  1.13it/s, est. speed input: 242.01 toks/s, output: 1216.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:23<01:11,  1.02it/s, est. speed input: 227.83 toks/s, output: 1240.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<01:29,  1.24s/it, est. speed input: 207.64 toks/s, output: 1225.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:26<01:20,  1.14s/it, est. speed input: 204.40 toks/s, output: 1301.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:27<01:06,  1.04it/s, est. speed input: 199.45 toks/s, output: 1452.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:27<00:53,  1.27it/s, est. speed input: 200.85 toks/s, output: 1553.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:27<00:42,  1.56it/s, est. speed input: 201.99 toks/s, output: 1653.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:28<00:27,  2.33it/s, est. speed input: 204.94 toks/s, output: 1856.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:29<00:33,  1.89it/s, est. speed input: 200.74 toks/s, output: 1906.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:29<00:33,  1.86it/s, est. speed input: 198.91 toks/s, output: 1977.65 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:13:24 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7282.8 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 9.2%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:30<00:30,  1.97it/s, est. speed input: 196.80 toks/s, output: 2129.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:31<00:30,  1.95it/s, est. speed input: 195.34 toks/s, output: 2200.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:31<00:23,  2.50it/s, est. speed input: 196.44 toks/s, output: 2382.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:31<00:15,  3.55it/s, est. speed input: 199.26 toks/s, output: 2585.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:31<00:14,  3.85it/s, est. speed input: 200.01 toks/s, output: 2678.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:32<00:16,  3.31it/s, est. speed input: 199.10 toks/s, output: 2748.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:33<00:22,  2.31it/s, est. speed input: 195.86 toks/s, output: 2786.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:33<00:19,  2.70it/s, est. speed input: 196.56 toks/s, output: 2878.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:33<00:19,  2.56it/s, est. speed input: 195.78 toks/s, output: 2948.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:34<00:19,  2.58it/s, est. speed input: 195.33 toks/s, output: 3023.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:34<00:16,  2.94it/s, est. speed input: 195.84 toks/s, output: 3112.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:34<00:18,  2.65it/s, est. speed input: 194.89 toks/s, output: 3178.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:35<00:21,  2.16it/s, est. speed input: 192.86 toks/s, output: 3226.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:36<00:31,  1.46it/s, est. speed input: 188.22 toks/s, output: 3228.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:37<00:27,  1.65it/s, est. speed input: 187.69 toks/s, output: 3301.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:37<00:21,  2.01it/s, est. speed input: 188.08 toks/s, output: 3388.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:38<00:24,  1.75it/s, est. speed input: 185.91 toks/s, output: 3431.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:38<00:19,  2.11it/s, est. speed input: 186.31 toks/s, output: 3519.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:38<00:19,  2.15it/s, est. speed input: 185.72 toks/s, output: 3589.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:24,  1.63it/s, est. speed input: 182.74 toks/s, output: 3612.40 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:13:34 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5822.5 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:24,  1.59it/s, est. speed input: 181.19 toks/s, output: 3662.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:42<00:36,  1.03it/s, est. speed input: 175.03 toks/s, output: 3620.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:34,  1.08it/s, est. speed input: 173.12 toks/s, output: 3662.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:26,  1.37it/s, est. speed input: 173.49 toks/s, output: 3750.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:44<00:20,  1.69it/s, est. speed input: 172.77 toks/s, output: 3896.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:44<00:12,  2.48it/s, est. speed input: 174.56 toks/s, output: 4097.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:45<00:19,  1.62it/s, est. speed input: 170.59 toks/s, output: 4086.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:45<00:15,  1.99it/s, est. speed input: 171.42 toks/s, output: 4186.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:48<00:29,  1.01s/it, est. speed input: 163.84 toks/s, output: 4084.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:48<00:18,  1.46it/s, est. speed input: 164.68 toks/s, output: 4270.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:49<00:15,  1.69it/s, est. speed input: 164.99 toks/s, output: 4360.84 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:13:44 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4322.8 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:19,  1.27it/s, est. speed input: 161.67 toks/s, output: 4356.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:50<00:15,  1.55it/s, est. speed input: 162.07 toks/s, output: 4449.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:51<00:14,  1.63it/s, est. speed input: 161.53 toks/s, output: 4519.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:23,  1.08s/it, est. speed input: 155.76 toks/s, output: 4443.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:53<00:17,  1.18it/s, est. speed input: 156.10 toks/s, output: 4537.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:54<00:13,  1.50it/s, est. speed input: 156.63 toks/s, output: 4636.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:56<00:21,  1.13s/it, est. speed input: 151.44 toks/s, output: 4570.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:57<00:21,  1.21s/it, est. speed input: 148.77 toks/s, output: 4577.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:58<00:16,  1.06it/s, est. speed input: 149.02 toks/s, output: 4673.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:58<00:08,  1.74it/s, est. speed input: 150.36 toks/s, output: 4890.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:59<00:10,  1.36it/s, est. speed input: 148.31 toks/s, output: 4912.24 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:13:54 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3141.7 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 75.4%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:00<00:09,  1.36it/s, est. speed input: 147.50 toks/s, output: 4973.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:12,  1.08s/it, est. speed input: 143.74 toks/s, output: 4938.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [01:03<00:13,  1.22s/it, est. speed input: 141.12 toks/s, output: 4942.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:04<00:11,  1.15s/it, est. speed input: 139.92 toks/s, output: 4994.27 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  1.15s/it, est. speed input: 149.26 toks/s, output: 6257.40 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:04<00:00,  2.47it/s, est. speed input: 149.26 toks/s, output: 6257.40 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:14:02 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:14:02 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:14:02 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:14:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:14:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:14:03 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:14:03 [abstract.py:306] It took 0.520273 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:14:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:14:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:14:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:14:03 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0112
+  • Training Accuracy: 0.4062
+  • Mean Generation Length: 2536.3374
+  • Training FLOPS: 47.18 TFLOPS (11.79 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 303.75s
+  • policy_training: 194.98s (64.2%)
+  • generation: 69.10s (22.7%)
+  • teacher_logprob_inference: 34.46s (11.3%)
+  • policy_training/sharding_data: 4.34s (1.4%)
+  • prepare_for_generation: 3.74s (1.2%)
+  • prepare_for_generation/transfer_and_update_weights: 1.71s (0.6%)
+  • policy_training/submit_training_futures: 1.63s (0.5%)
+  • training_prep: 0.83s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.1%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 189: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:17:55 [abstract.py:324] It took 0.051371 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242070e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66440887e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79800873e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242070e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66440887e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79800873e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242070e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66440887e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79800873e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242070e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66440887e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.79800873e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245039e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68944054e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79045258e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.05996914e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.78968287e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #189] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45492310e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245039e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68944054e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79045258e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.69553070e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37155380e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #189] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.56465454e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245039e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68944054e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79045258e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36548767e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.55818796e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #189] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.50410080e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245039e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.68944054e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.79045258e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.22827454e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67641182e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #189] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36313583e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 189: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:17:58 [abstract.py:324] It took 0.030534 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13883.54it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:17:58 [loggers.py:236] Engine 000: Avg prompt throughput: 39.7 tokens/s, Avg generation throughput: 45.1 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:16,  2.09it/s, est. speed input: 123.23 toks/s, output: 27.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:08, 17.07it/s, est. speed input: 823.96 toks/s, output: 230.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:06, 23.67it/s, est. speed input: 1114.48 toks/s, output: 394.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 30.10it/s, est. speed input: 1349.16 toks/s, output: 598.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:03, 36.06it/s, est. speed input: 1567.85 toks/s, output: 859.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:01<00:03, 39.03it/s, est. speed input: 1698.10 toks/s, output: 1078.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  24%|██▍       | 39/160 [00:01<00:02, 54.89it/s, est. speed input: 2080.26 toks/s, output: 1635.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 56.23it/s, est. speed input: 2227.24 toks/s, output: 1972.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:01, 59.06it/s, est. speed input: 2366.88 toks/s, output: 2337.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 58.03it/s, est. speed input: 2451.46 toks/s, output: 2667.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:01<00:01, 47.01it/s, est. speed input: 2392.75 toks/s, output: 2884.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:02<00:02, 31.29it/s, est. speed input: 2149.41 toks/s, output: 2874.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:02<00:02, 27.71it/s, est. speed input: 2052.62 toks/s, output: 3017.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:03<00:07, 10.27it/s, est. speed input: 1395.61 toks/s, output: 2324.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:05<00:16,  4.47it/s, est. speed input: 882.69 toks/s, output: 1674.07 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:06<00:18,  3.98it/s, est. speed input: 793.76 toks/s, output: 1687.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:09<00:31,  2.27it/s, est. speed input: 583.98 toks/s, output: 1430.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:09<00:25,  2.65it/s, est. speed input: 564.16 toks/s, output: 1675.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:10<00:25,  2.65it/s, est. speed input: 549.51 toks/s, output: 1733.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9957.6 tokens/s, Running: 67 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.8%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:12<00:42,  1.54it/s, est. speed input: 454.40 toks/s, output: 1536.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:12<00:37,  1.73it/s, est. speed input: 451.15 toks/s, output: 1628.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:13<00:32,  1.94it/s, est. speed input: 446.74 toks/s, output: 1714.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:13<00:30,  2.08it/s, est. speed input: 439.23 toks/s, output: 1787.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:15<00:41,  1.47it/s, est. speed input: 390.39 toks/s, output: 1779.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:16<00:43,  1.39it/s, est. speed input: 373.31 toks/s, output: 1805.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:16<00:39,  1.51it/s, est. speed input: 366.10 toks/s, output: 1873.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:16<00:32,  1.80it/s, est. speed input: 364.78 toks/s, output: 1967.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:17<00:32,  1.74it/s, est. speed input: 355.24 toks/s, output: 2018.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:18<00:27,  2.01it/s, est. speed input: 346.48 toks/s, output: 2166.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:19<00:30,  1.76it/s, est. speed input: 335.18 toks/s, output: 2197.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:19<00:25,  2.11it/s, est. speed input: 335.08 toks/s, output: 2297.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:19<00:22,  2.28it/s, est. speed input: 332.32 toks/s, output: 2378.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:20<00:15,  3.20it/s, est. speed input: 333.68 toks/s, output: 2585.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:20<00:10,  4.47it/s, est. speed input: 337.19 toks/s, output: 2806.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7545.9 tokens/s, Running: 48 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:21<00:19,  2.45it/s, est. speed input: 322.79 toks/s, output: 2785.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:23<00:33,  1.39it/s, est. speed input: 300.90 toks/s, output: 2696.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:23<00:25,  1.76it/s, est. speed input: 302.07 toks/s, output: 2806.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:24<00:25,  1.67it/s, est. speed input: 291.26 toks/s, output: 2898.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:24<00:21,  1.98it/s, est. speed input: 291.42 toks/s, output: 2998.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:24<00:18,  2.27it/s, est. speed input: 291.03 toks/s, output: 3092.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:24<00:10,  3.59it/s, est. speed input: 294.69 toks/s, output: 3325.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:25<00:13,  2.92it/s, est. speed input: 290.50 toks/s, output: 3376.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:28<00:36,  1.03it/s, est. speed input: 262.88 toks/s, output: 3155.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:28<00:21,  1.65it/s, est. speed input: 265.46 toks/s, output: 3386.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:28<00:17,  1.91it/s, est. speed input: 265.33 toks/s, output: 3483.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:29<00:16,  1.98it/s, est. speed input: 263.31 toks/s, output: 3555.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:30<00:13,  2.23it/s, est. speed input: 260.83 toks/s, output: 3716.99 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5519.6 tokens/s, Running: 31 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.5%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:30<00:14,  2.14it/s, est. speed input: 258.28 toks/s, output: 3778.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:30<00:12,  2.27it/s, est. speed input: 257.26 toks/s, output: 3861.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:31<00:07,  3.43it/s, est. speed input: 259.93 toks/s, output: 4095.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:31<00:09,  2.83it/s, est. speed input: 257.24 toks/s, output: 4149.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:31<00:07,  3.22it/s, est. speed input: 257.65 toks/s, output: 4253.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:32<00:06,  3.44it/s, est. speed input: 257.63 toks/s, output: 4349.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:32<00:07,  3.03it/s, est. speed input: 256.03 toks/s, output: 4417.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:09,  2.32it/s, est. speed input: 252.57 toks/s, output: 4454.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:33<00:09,  2.11it/s, est. speed input: 250.00 toks/s, output: 4505.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:34<00:13,  1.43it/s, est. speed input: 242.77 toks/s, output: 4475.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:36<00:13,  1.30it/s, est. speed input: 234.85 toks/s, output: 4525.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:36<00:10,  1.55it/s, est. speed input: 234.83 toks/s, output: 4626.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:37<00:06,  2.48it/s, est. speed input: 237.33 toks/s, output: 4877.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:37<00:05,  2.54it/s, est. speed input: 236.67 toks/s, output: 4963.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:39<00:07,  1.64it/s, est. speed input: 228.37 toks/s, output: 4989.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3252.1 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:41<00:10,  1.09it/s, est. speed input: 218.58 toks/s, output: 4884.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:41<00:07,  1.30it/s, est. speed input: 218.48 toks/s, output: 4990.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:41<00:05,  1.56it/s, est. speed input: 218.54 toks/s, output: 5099.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:42<00:02,  2.40it/s, est. speed input: 220.29 toks/s, output: 5354.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:42<00:02,  2.26it/s, est. speed input: 219.01 toks/s, output: 5429.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:42<00:01,  2.50it/s, est. speed input: 219.07 toks/s, output: 5537.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:44<00:02,  1.35it/s, est. speed input: 211.99 toks/s, output: 5471.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:45<00:02,  1.09it/s, est. speed input: 207.00 toks/s, output: 5457.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:46<00:01,  1.13it/s, est. speed input: 204.69 toks/s, output: 5514.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:49<00:01,  1.50s/it, est. speed input: 193.46 toks/s, output: 5337.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.19s/it, est. speed input: 192.90 toks/s, output: 5448.36 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.19s/it, est. speed input: 192.90 toks/s, output: 5448.36 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.18it/s, est. speed input: 192.90 toks/s, output: 5448.36 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:50 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:50 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:18:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:18:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:18:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:18:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:18:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:18:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:18:51 [abstract.py:306] It took 0.561966 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:18:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:18:51 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+
+📊 Training Results:
+  • Loss: 0.0129
+  • Training Accuracy: 0.3125
+  • Mean Generation Length: 1711.6125
+  • Training FLOPS: 39.52 TFLOPS (9.88 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 239.07s
+  • policy_training: 154.17s (64.5%)
+  • generation: 53.37s (22.3%)
+  • teacher_logprob_inference: 26.34s (11.0%)
+  • policy_training/sharding_data: 4.36s (1.8%)
+  • prepare_for_generation: 3.75s (1.6%)
+  • policy_training/submit_training_futures: 2.05s (0.9%)
+  • prepare_for_generation/transfer_and_update_weights: 1.63s (0.7%)
+  • training_prep: 0.89s (0.4%)
+  • teacher_logprob_inference_prep: 0.39s (0.2%)
+  • get_topk_logits/shard_data: 0.05s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 190: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:21:54 [abstract.py:324] It took 0.047917 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09249062e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09249062e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09249062e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09249062e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239336e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69181671e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80535736e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239336e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69181671e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80535736e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239336e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69181671e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80535736e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09239336e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.69181671e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.80535736e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67074547e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80942154e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67074547e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80942154e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67074547e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80942154e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67074547e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.80942154e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.04401855e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.75019360e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #190] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45732513e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34364929e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.66823959e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #190] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.51309128e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.23897324e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67517834e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68910828e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36687069e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #190] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.58955917e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #190] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36320129e+02
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 190: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:21:58 [abstract.py:324] It took 0.031082 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13965.01it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 14:21:58 [loggers.py:236] Engine 000: Avg prompt throughput: 48.8 tokens/s, Avg generation throughput: 54.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:35,  1.67it/s, est. speed input: 105.04 toks/s, output: 36.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:17,  8.97it/s, est. speed input: 434.69 toks/s, output: 186.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:12, 12.29it/s, est. speed input: 580.85 toks/s, output: 293.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:09, 15.38it/s, est. speed input: 696.61 toks/s, output: 414.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:01<00:05, 27.79it/s, est. speed input: 1039.76 toks/s, output: 786.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:01<00:03, 34.06it/s, est. speed input: 1269.61 toks/s, output: 1123.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 51.44it/s, est. speed input: 1682.45 toks/s, output: 1759.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 46.84it/s, est. speed input: 1767.61 toks/s, output: 2023.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███       | 49/160 [00:01<00:02, 46.37it/s, est. speed input: 1848.30 toks/s, output: 2290.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:02, 47.25it/s, est. speed input: 1928.66 toks/s, output: 2589.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 51.65it/s, est. speed input: 2041.54 toks/s, output: 2982.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:02<00:02, 36.58it/s, est. speed input: 1941.26 toks/s, output: 3057.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:04<00:12,  7.24it/s, est. speed input: 1000.71 toks/s, output: 1799.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:05<00:13,  6.01it/s, est. speed input: 853.95 toks/s, output: 1852.11 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:09<00:32,  2.48it/s, est. speed input: 505.93 toks/s, output: 1325.19 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10519.5 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:11<00:38,  2.00it/s, est. speed input: 428.53 toks/s, output: 1300.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:12<00:37,  2.01it/s, est. speed input: 405.57 toks/s, output: 1417.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:13<00:34,  2.15it/s, est. speed input: 393.44 toks/s, output: 1562.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:13<00:30,  2.37it/s, est. speed input: 394.43 toks/s, output: 1659.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:14<00:31,  2.31it/s, est. speed input: 385.13 toks/s, output: 1713.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:15<00:47,  1.49it/s, est. speed input: 346.33 toks/s, output: 1635.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:16<00:39,  1.76it/s, est. speed input: 338.17 toks/s, output: 1781.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:16<00:33,  2.05it/s, est. speed input: 338.66 toks/s, output: 1876.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<00:37,  1.80it/s, est. speed input: 326.65 toks/s, output: 1901.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:18<00:39,  1.66it/s, est. speed input: 316.55 toks/s, output: 1934.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:18<00:28,  2.28it/s, est. speed input: 316.55 toks/s, output: 2115.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:18<00:23,  2.71it/s, est. speed input: 317.48 toks/s, output: 2212.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:16,  3.67it/s, est. speed input: 319.88 toks/s, output: 2407.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:17,  3.27it/s, est. speed input: 312.53 toks/s, output: 2609.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8028.5 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.1%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:20<00:14,  3.78it/s, est. speed input: 313.16 toks/s, output: 2789.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:22<00:29,  1.84it/s, est. speed input: 285.84 toks/s, output: 2713.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:22<00:25,  2.05it/s, est. speed input: 285.64 toks/s, output: 2800.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:23<00:19,  2.67it/s, est. speed input: 287.20 toks/s, output: 2990.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:23<00:17,  2.92it/s, est. speed input: 287.13 toks/s, output: 3077.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:24<00:20,  2.41it/s, est. speed input: 281.57 toks/s, output: 3106.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:19,  2.47it/s, est. speed input: 279.77 toks/s, output: 3174.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:24<00:16,  2.79it/s, est. speed input: 279.73 toks/s, output: 3260.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:24<00:14,  3.21it/s, est. speed input: 280.12 toks/s, output: 3352.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:11,  3.80it/s, est. speed input: 281.08 toks/s, output: 3449.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:25<00:11,  3.75it/s, est. speed input: 279.53 toks/s, output: 3680.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:26<00:12,  3.23it/s, est. speed input: 276.74 toks/s, output: 3730.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:26<00:09,  4.05it/s, est. speed input: 278.36 toks/s, output: 3922.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:26<00:07,  5.10it/s, est. speed input: 280.72 toks/s, output: 4126.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:27<00:05,  6.46it/s, est. speed input: 283.84 toks/s, output: 4337.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:27<00:07,  4.48it/s, est. speed input: 280.66 toks/s, output: 4453.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:29<00:11,  2.69it/s, est. speed input: 271.46 toks/s, output: 4471.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:29<00:09,  3.07it/s, est. speed input: 272.45 toks/s, output: 4572.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:29<00:08,  3.27it/s, est. speed input: 272.38 toks/s, output: 4656.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:30<00:10,  2.64it/s, est. speed input: 268.69 toks/s, output: 4679.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:30<00:08,  3.11it/s, est. speed input: 269.44 toks/s, output: 4777.06 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5713.7 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:30<00:09,  2.81it/s, est. speed input: 267.39 toks/s, output: 4826.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:30<00:07,  3.17it/s, est. speed input: 267.54 toks/s, output: 4915.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:31<00:08,  2.73it/s, est. speed input: 265.24 toks/s, output: 4958.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:31<00:08,  2.73it/s, est. speed input: 264.05 toks/s, output: 5022.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:32<00:12,  1.81it/s, est. speed input: 257.80 toks/s, output: 4991.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:32<00:04,  4.40it/s, est. speed input: 264.16 toks/s, output: 5464.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:35<00:07,  2.06it/s, est. speed input: 251.60 toks/s, output: 5383.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:35<00:06,  2.35it/s, est. speed input: 252.24 toks/s, output: 5486.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:35<00:06,  2.20it/s, est. speed input: 249.94 toks/s, output: 5527.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:36<00:05,  2.37it/s, est. speed input: 249.50 toks/s, output: 5608.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:36<00:04,  2.44it/s, est. speed input: 248.62 toks/s, output: 5679.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:37<00:05,  2.08it/s, est. speed input: 245.63 toks/s, output: 5705.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:37<00:04,  2.05it/s, est. speed input: 243.94 toks/s, output: 5759.54 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2906.3 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:40<00:09,  1.10s/it, est. speed input: 229.17 toks/s, output: 5512.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:40<00:05,  1.37it/s, est. speed input: 229.20 toks/s, output: 5714.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:43<00:06,  1.13s/it, est. speed input: 217.86 toks/s, output: 5540.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:44<00:05,  1.10s/it, est. speed input: 214.35 toks/s, output: 5559.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:45<00:04,  1.06s/it, est. speed input: 211.09 toks/s, output: 5586.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:45<00:02,  1.12it/s, est. speed input: 210.29 toks/s, output: 5678.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:46<00:01,  1.14it/s, est. speed input: 207.81 toks/s, output: 5727.03 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:48 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1031.2 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 75.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:50<00:01,  1.78s/it, est. speed input: 192.44 toks/s, output: 5429.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.30s/it, est. speed input: 193.19 toks/s, output: 5575.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  1.30s/it, est. speed input: 193.19 toks/s, output: 5575.79 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:50<00:00,  3.16it/s, est. speed input: 193.19 toks/s, output: 5575.79 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:22:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:22:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:22:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:22:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:22:52 [abstract.py:306] It took 0.535208 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:22:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:22:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:22:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:22:52 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0127
+  • Training Accuracy: 0.5500
+  • Mean Generation Length: 1764.9375
+  • Training FLOPS: 40.92 TFLOPS (10.23 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 238.90s
+  • policy_training: 152.25s (63.7%)
+  • generation: 53.81s (22.5%)
+  • teacher_logprob_inference: 26.59s (11.1%)
+  • prepare_for_generation: 4.69s (2.0%)
+  • policy_training/sharding_data: 4.31s (1.8%)
+  • prepare_for_generation/transfer_and_update_weights: 2.36s (1.0%)
+  • policy_training/submit_training_futures: 2.04s (0.9%)
+  • training_prep: 0.95s (0.4%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Epoch 39/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 191: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:25:53 [abstract.py:324] It took 0.054504 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257695e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67980652e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82299500e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257695e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67980652e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82299500e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257695e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67980652e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82299500e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09257695e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.67980652e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82299500e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09258594e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71870804e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82819595e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.05854683e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.72291279e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #191] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45682007e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09258594e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71870804e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82819595e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.68350449e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35905552e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #191] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.59197617e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09258594e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71870804e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82819595e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.34297791e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.89481926e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #191] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.50133133e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09258594e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.71870804e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82819595e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.24341698e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.67693024e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #191] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36245438e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 191: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:25:56 [abstract.py:324] It took 0.028869 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12921.70it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:25:56 [loggers.py:236] Engine 000: Avg prompt throughput: 51.3 tokens/s, Avg generation throughput: 1.4 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<01:15,  2.10it/s, est. speed input: 126.29 toks/s, output: 29.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:15, 10.02it/s, est. speed input: 498.76 toks/s, output: 148.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:11, 13.29it/s, est. speed input: 645.05 toks/s, output: 263.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:08, 17.32it/s, est. speed input: 805.24 toks/s, output: 439.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:01<00:08, 16.68it/s, est. speed input: 828.50 toks/s, output: 553.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:01<00:06, 20.63it/s, est. speed input: 945.47 toks/s, output: 759.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:01<00:05, 27.19it/s, est. speed input: 1104.06 toks/s, output: 1058.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:03, 34.41it/s, est. speed input: 1272.66 toks/s, output: 1412.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:03, 37.37it/s, est. speed input: 1375.23 toks/s, output: 1678.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:01<00:03, 39.63it/s, est. speed input: 1468.15 toks/s, output: 1946.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 44.01it/s, est. speed input: 1584.46 toks/s, output: 2301.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▏      | 51/160 [00:01<00:02, 42.96it/s, est. speed input: 1639.60 toks/s, output: 2551.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 47.15it/s, est. speed input: 1737.32 toks/s, output: 2906.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:02<00:02, 36.34it/s, est. speed input: 1702.63 toks/s, output: 3034.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:03<00:07, 12.39it/s, est. speed input: 1243.98 toks/s, output: 2441.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:10<00:44,  2.01it/s, est. speed input: 417.94 toks/s, output: 1099.70 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 14:26:06 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11015.6 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:11<00:43,  2.00it/s, est. speed input: 378.84 toks/s, output: 1259.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:12<00:38,  2.18it/s, est. speed input: 373.68 toks/s, output: 1421.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:16<01:01,  1.34it/s, est. speed input: 290.80 toks/s, output: 1270.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:18<01:11,  1.13it/s, est. speed input: 264.11 toks/s, output: 1245.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:18<01:04,  1.23it/s, est. speed input: 261.78 toks/s, output: 1325.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:18<00:58,  1.35it/s, est. speed input: 259.42 toks/s, output: 1404.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:19<00:50,  1.56it/s, est. speed input: 259.17 toks/s, output: 1492.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:19<00:45,  1.69it/s, est. speed input: 256.83 toks/s, output: 1567.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:19<00:36,  2.07it/s, est. speed input: 258.01 toks/s, output: 1662.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:26:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8687.8 tokens/s, Running: 76 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.6%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:20<00:47,  1.58it/s, est. speed input: 247.71 toks/s, output: 1683.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:21<00:43,  1.70it/s, est. speed input: 245.00 toks/s, output: 1752.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:21<00:44,  1.66it/s, est. speed input: 240.51 toks/s, output: 1806.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:22<00:40,  1.76it/s, est. speed input: 237.96 toks/s, output: 1873.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:23<00:49,  1.43it/s, est. speed input: 230.04 toks/s, output: 1897.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:23<00:42,  1.64it/s, est. speed input: 228.72 toks/s, output: 1971.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:24<00:45,  1.53it/s, est. speed input: 224.19 toks/s, output: 2016.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:25<00:43,  1.56it/s, est. speed input: 221.07 toks/s, output: 2072.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:25<00:43,  1.56it/s, est. speed input: 217.90 toks/s, output: 2126.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:26<00:45,  1.44it/s, est. speed input: 213.44 toks/s, output: 2167.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:27<00:56,  1.16it/s, est. speed input: 205.80 toks/s, output: 2173.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:28<00:47,  1.33it/s, est. speed input: 204.40 toks/s, output: 2242.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:28<00:35,  1.76it/s, est. speed input: 205.41 toks/s, output: 2336.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:28<00:27,  2.28it/s, est. speed input: 206.55 toks/s, output: 2430.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:28<00:22,  2.74it/s, est. speed input: 207.20 toks/s, output: 2519.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:28<00:17,  3.36it/s, est. speed input: 208.30 toks/s, output: 2613.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:29<00:16,  3.49it/s, est. speed input: 208.54 toks/s, output: 2695.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:29<00:18,  3.13it/s, est. speed input: 207.88 toks/s, output: 2765.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:26:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7098.6 tokens/s, Running: 57 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.4%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:31<00:30,  1.85it/s, est. speed input: 201.09 toks/s, output: 2830.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:31<00:26,  2.04it/s, est. speed input: 200.88 toks/s, output: 2906.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:31<00:24,  2.24it/s, est. speed input: 200.71 toks/s, output: 2983.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:32<00:20,  2.53it/s, est. speed input: 200.99 toks/s, output: 3066.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:34<00:49,  1.05it/s, est. speed input: 188.82 toks/s, output: 2961.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:35<00:51,  1.01s/it, est. speed input: 184.31 toks/s, output: 2972.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:35<00:39,  1.27it/s, est. speed input: 184.80 toks/s, output: 3060.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:38,  1.27it/s, est. speed input: 182.45 toks/s, output: 3100.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:37<00:38,  1.24it/s, est. speed input: 179.96 toks/s, output: 3137.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:38<00:33,  1.39it/s, est. speed input: 179.10 toks/s, output: 3203.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:39<00:41,  1.10it/s, est. speed input: 174.48 toks/s, output: 3202.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:39<00:32,  1.38it/s, est. speed input: 174.71 toks/s, output: 3286.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:40<00:27,  1.59it/s, est. speed input: 174.38 toks/s, output: 3361.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:26:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5891.8 tokens/s, Running: 44 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:40<00:24,  1.77it/s, est. speed input: 174.10 toks/s, output: 3435.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:41<00:26,  1.58it/s, est. speed input: 170.98 toks/s, output: 3532.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:42<00:20,  1.90it/s, est. speed input: 170.87 toks/s, output: 3688.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:43<00:18,  2.01it/s, est. speed input: 170.70 toks/s, output: 3763.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:43<00:15,  2.43it/s, est. speed input: 171.62 toks/s, output: 3861.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:43<00:18,  1.98it/s, est. speed input: 169.94 toks/s, output: 3902.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:44<00:14,  2.34it/s, est. speed input: 170.49 toks/s, output: 3993.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:45<00:20,  1.68it/s, est. speed input: 167.95 toks/s, output: 4012.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:45<00:15,  2.14it/s, est. speed input: 168.70 toks/s, output: 4109.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:46<00:18,  1.74it/s, est. speed input: 166.96 toks/s, output: 4145.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:47<00:22,  1.35it/s, est. speed input: 164.21 toks/s, output: 4157.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:47<00:17,  1.74it/s, est. speed input: 164.82 toks/s, output: 4253.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:48<00:16,  1.74it/s, est. speed input: 164.06 toks/s, output: 4314.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:48<00:13,  2.07it/s, est. speed input: 164.35 toks/s, output: 4402.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:49<00:14,  1.86it/s, est. speed input: 163.33 toks/s, output: 4455.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:50<00:19,  1.34it/s, est. speed input: 160.53 toks/s, output: 4457.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:26:46 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4382.2 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:50<00:14,  1.67it/s, est. speed input: 160.92 toks/s, output: 4548.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:50<00:13,  1.83it/s, est. speed input: 160.72 toks/s, output: 4622.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:52<00:17,  1.29it/s, est. speed input: 157.81 toks/s, output: 4619.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:53<00:18,  1.19it/s, est. speed input: 156.02 toks/s, output: 4647.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:54<00:22,  1.08s/it, est. speed input: 152.45 toks/s, output: 4623.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:56<00:23,  1.18s/it, est. speed input: 149.68 toks/s, output: 4624.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:56<00:18,  1.01it/s, est. speed input: 149.25 toks/s, output: 4696.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:57<00:13,  1.33it/s, est. speed input: 149.80 toks/s, output: 4797.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:59<00:16,  1.04s/it, est. speed input: 144.87 toks/s, output: 4805.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:26:56 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3169.9 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.3%, Prefix cache hit rate: 75.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [01:00<00:14,  1.05it/s, est. speed input: 144.24 toks/s, output: 4870.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [01:00<00:10,  1.30it/s, est. speed input: 144.64 toks/s, output: 4970.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [01:02<00:15,  1.16s/it, est. speed input: 140.52 toks/s, output: 4917.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:03<00:12,  1.04s/it, est. speed input: 139.91 toks/s, output: 4984.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [01:04<00:06,  1.46it/s, est. speed input: 140.69 toks/s, output: 5189.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [01:04<00:05,  1.75it/s, est. speed input: 141.15 toks/s, output: 5296.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [01:04<00:04,  1.98it/s, est. speed input: 141.38 toks/s, output: 5393.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [01:05<00:04,  1.54it/s, est. speed input: 140.05 toks/s, output: 5433.04 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  1.54it/s, est. speed input: 146.48 toks/s, output: 6305.07 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:05<00:00,  2.43it/s, est. speed input: 146.48 toks/s, output: 6305.07 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:27:05 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:27:06 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:27:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:27:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:27:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:27:06 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:27:06 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:27:06 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:27:06 [abstract.py:306] It took 0.536390 seconds to fall asleep.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:27:06 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:27:06 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0107
+  • Training Accuracy: 0.3375
+  • Mean Generation Length: 2591.2749
+  • Training FLOPS: 46.75 TFLOPS (11.69 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 310.31s
+  • policy_training: 200.41s (64.6%)
+  • generation: 70.12s (22.6%)
+  • teacher_logprob_inference: 34.72s (11.2%)
+  • policy_training/sharding_data: 4.30s (1.4%)
+  • prepare_for_generation: 3.63s (1.2%)
+  • policy_training/submit_training_futures: 1.80s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.61s (0.5%)
+  • training_prep: 0.84s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.1%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 192: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:03 [abstract.py:324] It took 0.048976 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242695e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68767281e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82828064e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242695e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68767281e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82828064e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242695e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68767281e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82828064e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09242695e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68767281e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.82828064e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245957e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66890717e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81487503e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.03526039e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.66259146e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #192] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45838623e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245957e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66890717e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81487503e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66456909e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35982685e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #192] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.59017792e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245957e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66890717e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81487503e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36117401e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.70728159e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #192] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.47254677e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09245957e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66890717e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.81487503e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.23658829e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65858421e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #192] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36335358e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 192: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:06 [abstract.py:324] It took 0.030280 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12974.16it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:06 [loggers.py:236] Engine 000: Avg prompt throughput: 39.0 tokens/s, Avg generation throughput: 47.7 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:27,  5.69it/s, est. speed input: 347.32 toks/s, output: 79.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 24.85it/s, est. speed input: 1325.03 toks/s, output: 412.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 35.34it/s, est. speed input: 1832.53 toks/s, output: 809.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:03, 38.99it/s, est. speed input: 2067.97 toks/s, output: 1131.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 25/160 [00:00<00:03, 39.96it/s, est. speed input: 2153.61 toks/s, output: 1390.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 31/160 [00:00<00:02, 44.58it/s, est. speed input: 2319.65 toks/s, output: 1745.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:02, 45.02it/s, est. speed input: 2383.33 toks/s, output: 2042.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▋       | 42/160 [00:01<00:02, 45.15it/s, est. speed input: 2432.23 toks/s, output: 2301.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 43.74it/s, est. speed input: 2437.16 toks/s, output: 2537.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 40.23it/s, est. speed input: 2395.78 toks/s, output: 2731.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 39.07it/s, est. speed input: 2379.74 toks/s, output: 2968.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:03, 26.76it/s, est. speed input: 2128.91 toks/s, output: 2845.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:02<00:04, 19.13it/s, est. speed input: 1870.51 toks/s, output: 2736.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:02<00:05, 16.70it/s, est. speed input: 1744.60 toks/s, output: 2735.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:03<00:11,  7.87it/s, est. speed input: 1277.18 toks/s, output: 2176.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:03<00:12,  6.75it/s, est. speed input: 1150.72 toks/s, output: 2122.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:04<00:17,  4.88it/s, est. speed input: 973.38 toks/s, output: 1955.21 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:04<00:16,  5.20it/s, est. speed input: 963.91 toks/s, output: 2022.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:06<00:31,  2.60it/s, est. speed input: 755.70 toks/s, output: 1678.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:07<00:38,  2.13it/s, est. speed input: 673.16 toks/s, output: 1589.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:07<00:39,  2.04it/s, est. speed input: 631.02 toks/s, output: 1585.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:08<00:51,  1.56it/s, est. speed input: 556.15 toks/s, output: 1493.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:09<00:46,  1.69it/s, est. speed input: 535.66 toks/s, output: 1534.55 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:16 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10447.1 tokens/s, Running: 79 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:11<01:23,  1.07s/it, est. speed input: 429.75 toks/s, output: 1329.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:12<01:16,  1.00it/s, est. speed input: 407.23 toks/s, output: 1358.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:15<01:53,  1.50s/it, est. speed input: 336.92 toks/s, output: 1223.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:17<02:00,  1.60s/it, est. speed input: 303.56 toks/s, output: 1202.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:19<02:11,  1.77s/it, est. speed input: 272.43 toks/s, output: 1177.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:19<01:36,  1.32s/it, est. speed input: 272.30 toks/s, output: 1275.13 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:26 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8239.1 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.5%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:20<01:18,  1.09s/it, est. speed input: 267.88 toks/s, output: 1351.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:20<00:57,  1.24it/s, est. speed input: 269.35 toks/s, output: 1454.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:20<00:43,  1.61it/s, est. speed input: 269.75 toks/s, output: 1551.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:21<00:45,  1.52it/s, est. speed input: 263.15 toks/s, output: 1607.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:21<00:35,  1.93it/s, est. speed input: 263.53 toks/s, output: 1703.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:21<00:30,  2.19it/s, est. speed input: 262.62 toks/s, output: 1789.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:22<00:34,  1.91it/s, est. speed input: 257.29 toks/s, output: 1845.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:22<00:36,  1.80it/s, est. speed input: 252.92 toks/s, output: 1905.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:23<00:32,  1.95it/s, est. speed input: 251.02 toks/s, output: 1982.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:24<00:37,  1.66it/s, est. speed input: 245.17 toks/s, output: 2026.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:24<00:30,  2.03it/s, est. speed input: 245.20 toks/s, output: 2117.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:25<00:38,  1.60it/s, est. speed input: 238.55 toks/s, output: 2150.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:26<00:39,  1.51it/s, est. speed input: 234.05 toks/s, output: 2199.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:26<00:28,  2.02it/s, est. speed input: 233.35 toks/s, output: 2369.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:26<00:19,  2.91it/s, est. speed input: 235.89 toks/s, output: 2570.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:27<00:20,  2.70it/s, est. speed input: 234.17 toks/s, output: 2637.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:29<00:44,  1.21it/s, est. speed input: 218.38 toks/s, output: 2547.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:36 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6950.3 tokens/s, Running: 53 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:31<00:40,  1.28it/s, est. speed input: 212.32 toks/s, output: 2649.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:31<00:35,  1.42it/s, est. speed input: 211.29 toks/s, output: 2724.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:31<00:27,  1.80it/s, est. speed input: 212.51 toks/s, output: 2827.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:31<00:22,  2.16it/s, est. speed input: 213.13 toks/s, output: 2922.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:32<00:20,  2.40it/s, est. speed input: 213.04 toks/s, output: 3007.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:33<00:26,  1.79it/s, est. speed input: 209.01 toks/s, output: 3036.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:20,  2.27it/s, est. speed input: 209.93 toks/s, output: 3135.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:33<00:16,  2.68it/s, est. speed input: 210.45 toks/s, output: 3228.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:34<00:16,  2.62it/s, est. speed input: 209.19 toks/s, output: 3376.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:36<00:20,  2.00it/s, est. speed input: 203.43 toks/s, output: 3527.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:36<00:21,  1.82it/s, est. speed input: 200.91 toks/s, output: 3568.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:36<00:17,  2.17it/s, est. speed input: 201.83 toks/s, output: 3667.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:37<00:17,  2.17it/s, est. speed input: 200.97 toks/s, output: 3736.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:38<00:20,  1.75it/s, est. speed input: 197.78 toks/s, output: 3761.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:38<00:19,  1.77it/s, est. speed input: 196.53 toks/s, output: 3822.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:39<00:21,  1.56it/s, est. speed input: 193.88 toks/s, output: 3855.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:39<00:17,  1.89it/s, est. speed input: 194.29 toks/s, output: 3946.05 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:47 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5205.8 tokens/s, Running: 33 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:40<00:18,  1.72it/s, est. speed input: 192.38 toks/s, output: 3991.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:41<00:16,  1.84it/s, est. speed input: 191.78 toks/s, output: 4063.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:41<00:11,  2.50it/s, est. speed input: 192.63 toks/s, output: 4248.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:41<00:10,  2.76it/s, est. speed input: 192.93 toks/s, output: 4338.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:42<00:11,  2.35it/s, est. speed input: 191.57 toks/s, output: 4392.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:42<00:10,  2.53it/s, est. speed input: 191.58 toks/s, output: 4476.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:43<00:12,  2.05it/s, est. speed input: 189.74 toks/s, output: 4518.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:43<00:11,  2.13it/s, est. speed input: 189.26 toks/s, output: 4591.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:44<00:11,  2.06it/s, est. speed input: 188.34 toks/s, output: 4654.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:44<00:09,  2.27it/s, est. speed input: 188.27 toks/s, output: 4738.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:45<00:08,  2.54it/s, est. speed input: 188.42 toks/s, output: 4826.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:45<00:09,  2.01it/s, est. speed input: 186.67 toks/s, output: 4866.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:45<00:07,  2.41it/s, est. speed input: 187.12 toks/s, output: 4962.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:46<00:08,  2.23it/s, est. speed input: 186.29 toks/s, output: 5025.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:48<00:14,  1.16it/s, est. speed input: 180.44 toks/s, output: 4955.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:48<00:11,  1.38it/s, est. speed input: 180.23 toks/s, output: 5036.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:11,  1.34it/s, est. speed input: 178.59 toks/s, output: 5079.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:49<00:05,  2.32it/s, est. speed input: 180.57 toks/s, output: 5311.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:49<00:04,  2.75it/s, est. speed input: 181.24 toks/s, output: 5417.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:31:57 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3517.9 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.1%, Prefix cache hit rate: 74.6%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:50<00:03,  2.51it/s, est. speed input: 180.45 toks/s, output: 5566.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:51<00:03,  2.58it/s, est. speed input: 180.42 toks/s, output: 5653.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:53<00:06,  1.18it/s, est. speed input: 173.91 toks/s, output: 5539.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:53<00:04,  1.50it/s, est. speed input: 174.57 toks/s, output: 5648.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:55<00:06,  1.01s/it, est. speed input: 169.52 toks/s, output: 5579.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:56<00:05,  1.01s/it, est. speed input: 167.60 toks/s, output: 5612.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:58<00:05,  1.35s/it, est. speed input: 162.26 toks/s, output: 5534.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:58<00:02,  1.00it/s, est. speed input: 162.98 toks/s, output: 5657.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:59<00:02,  1.02s/it, est. speed input: 161.09 toks/s, output: 5694.27 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  1.02s/it, est. speed input: 163.13 toks/s, output: 5968.11 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:59<00:00,  2.67it/s, est. speed input: 163.13 toks/s, output: 5968.11 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:32:10 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:32:10 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:32:10 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:32:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:32:10 [abstract.py:306] It took 0.520197 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:32:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:32:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:32:10 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0113
+  • Training Accuracy: 0.3375
+  • Mean Generation Length: 2231.6438
+  • Training FLOPS: 43.21 TFLOPS (10.80 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 285.63s
+  • policy_training: 185.08s (64.8%)
+  • generation: 63.65s (22.3%)
+  • teacher_logprob_inference: 31.77s (11.1%)
+  • policy_training/sharding_data: 4.31s (1.5%)
+  • prepare_for_generation: 3.75s (1.3%)
+  • policy_training/submit_training_futures: 1.62s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.6%)
+  • training_prep: 0.76s (0.3%)
+  • teacher_logprob_inference_prep: 0.45s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 193: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:35:49 [abstract.py:324] It took 0.051620 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09273555e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68091431e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83573303e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09273555e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68091431e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83573303e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09273555e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68091431e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83573303e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09273555e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.68091431e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83573303e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09281250e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66555824e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82908020e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.02600632e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.74863338e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #193] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.45288834e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09281250e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66555824e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82908020e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64903412e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35613480e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #193] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.57337341e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09281250e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66555824e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82908020e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.36125031e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.58118343e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #193] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.47562180e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09281250e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.66555824e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82908020e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.23948441e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65374622e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #193] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36188232e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 193: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:35:52 [abstract.py:324] It took 0.029527 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13166.35it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:35:53 [loggers.py:236] Engine 000: Avg prompt throughput: 41.2 tokens/s, Avg generation throughput: 56.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:45,  3.49it/s, est. speed input: 206.18 toks/s, output: 34.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:06, 21.86it/s, est. speed input: 1078.21 toks/s, output: 275.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 29.83it/s, est. speed input: 1490.20 toks/s, output: 499.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 35.36it/s, est. speed input: 1729.55 toks/s, output: 725.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 23/160 [00:00<00:03, 37.36it/s, est. speed input: 1861.77 toks/s, output: 961.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 29/160 [00:00<00:03, 42.85it/s, est. speed input: 2053.11 toks/s, output: 1275.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:00<00:02, 47.06it/s, est. speed input: 2210.79 toks/s, output: 1595.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 49.04it/s, est. speed input: 2326.31 toks/s, output: 1902.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 46.32it/s, est. speed input: 2350.67 toks/s, output: 2159.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 46.71it/s, est. speed input: 2396.36 toks/s, output: 2407.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▌      | 57/160 [00:01<00:02, 42.43it/s, est. speed input: 2364.98 toks/s, output: 2604.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 51.02it/s, est. speed input: 2514.12 toks/s, output: 3129.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 71/160 [00:02<00:03, 28.02it/s, est. speed input: 2144.67 toks/s, output: 2954.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:07<00:25,  3.23it/s, est. speed input: 624.89 toks/s, output: 1137.12 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:08<00:26,  3.09it/s, est. speed input: 562.26 toks/s, output: 1298.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:09<00:24,  3.15it/s, est. speed input: 528.94 toks/s, output: 1508.53 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10384.4 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:10<00:25,  2.94it/s, est. speed input: 493.94 toks/s, output: 1601.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:11<00:27,  2.64it/s, est. speed input: 458.44 toks/s, output: 1668.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:12<00:34,  2.12it/s, est. speed input: 421.07 toks/s, output: 1629.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:13<00:32,  2.19it/s, est. speed input: 413.94 toks/s, output: 1696.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:13<00:35,  1.98it/s, est. speed input: 396.10 toks/s, output: 1718.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:13<00:30,  2.26it/s, est. speed input: 395.15 toks/s, output: 1807.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:14<00:37,  1.85it/s, est. speed input: 375.54 toks/s, output: 1812.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<00:41,  1.63it/s, est. speed input: 359.25 toks/s, output: 1827.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<00:59,  1.13it/s, est. speed input: 328.05 toks/s, output: 1764.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:18<01:06,  1.01s/it, est. speed input: 307.34 toks/s, output: 1747.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:19<01:08,  1.05s/it, est. speed input: 292.45 toks/s, output: 1758.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7913.8 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:20<01:01,  1.04it/s, est. speed input: 284.92 toks/s, output: 1808.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:21<00:54,  1.17it/s, est. speed input: 279.61 toks/s, output: 1871.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:22<00:52,  1.18it/s, est. speed input: 272.11 toks/s, output: 1914.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:22<00:23,  2.50it/s, est. speed input: 276.44 toks/s, output: 2314.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:23<00:20,  2.76it/s, est. speed input: 275.17 toks/s, output: 2483.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:24<00:25,  2.12it/s, est. speed input: 266.73 toks/s, output: 2496.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:24<00:30,  1.76it/s, est. speed input: 259.06 toks/s, output: 2516.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:25<00:21,  2.40it/s, est. speed input: 260.64 toks/s, output: 2712.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:25<00:23,  2.19it/s, est. speed input: 256.93 toks/s, output: 2762.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:26<00:22,  2.22it/s, est. speed input: 255.03 toks/s, output: 2831.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:26<00:24,  2.01it/s, est. speed input: 251.23 toks/s, output: 2879.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:27<00:22,  2.14it/s, est. speed input: 249.86 toks/s, output: 2954.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:28<00:21,  2.07it/s, est. speed input: 243.28 toks/s, output: 3135.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:29<00:21,  2.04it/s, est. speed input: 240.96 toks/s, output: 3194.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:29<00:18,  2.33it/s, est. speed input: 241.22 toks/s, output: 3286.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:30<00:19,  2.20it/s, est. speed input: 238.96 toks/s, output: 3344.38 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6419.8 tokens/s, Running: 42 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:30<00:15,  2.63it/s, est. speed input: 239.68 toks/s, output: 3441.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:08,  4.49it/s, est. speed input: 243.67 toks/s, output: 3760.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:15,  2.45it/s, est. speed input: 236.71 toks/s, output: 3739.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:31<00:13,  2.66it/s, est. speed input: 236.61 toks/s, output: 3826.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:33<00:23,  1.47it/s, est. speed input: 226.83 toks/s, output: 3756.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:33<00:19,  1.78it/s, est. speed input: 227.13 toks/s, output: 3850.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:14,  2.22it/s, est. speed input: 227.95 toks/s, output: 3951.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:13,  2.32it/s, est. speed input: 227.17 toks/s, output: 4025.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:18,  1.70it/s, est. speed input: 222.52 toks/s, output: 4032.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:36<00:19,  1.56it/s, est. speed input: 219.44 toks/s, output: 4065.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:36<00:14,  2.05it/s, est. speed input: 220.49 toks/s, output: 4172.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:36<00:08,  3.16it/s, est. speed input: 222.44 toks/s, output: 4385.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:37<00:13,  1.87it/s, est. speed input: 216.90 toks/s, output: 4365.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:39<00:20,  1.22it/s, est. speed input: 209.46 toks/s, output: 4306.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:39<00:16,  1.50it/s, est. speed input: 209.56 toks/s, output: 4400.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:39<00:13,  1.69it/s, est. speed input: 208.97 toks/s, output: 4479.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4499.9 tokens/s, Running: 23 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:40<00:12,  1.78it/s, est. speed input: 207.92 toks/s, output: 4548.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:41<00:13,  1.61it/s, est. speed input: 205.57 toks/s, output: 4587.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:41<00:08,  2.18it/s, est. speed input: 206.00 toks/s, output: 4776.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:42<00:09,  1.86it/s, est. speed input: 203.60 toks/s, output: 4812.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:42<00:08,  2.04it/s, est. speed input: 203.30 toks/s, output: 4897.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:43<00:07,  2.08it/s, est. speed input: 202.57 toks/s, output: 4970.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:43<00:06,  2.24it/s, est. speed input: 202.27 toks/s, output: 5055.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:43<00:04,  2.82it/s, est. speed input: 203.04 toks/s, output: 5166.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:44<00:06,  1.99it/s, est. speed input: 200.43 toks/s, output: 5192.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:44<00:05,  2.30it/s, est. speed input: 200.54 toks/s, output: 5288.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:45<00:03,  2.80it/s, est. speed input: 201.12 toks/s, output: 5396.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:45<00:04,  2.29it/s, est. speed input: 199.66 toks/s, output: 5450.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:47<00:06,  1.29it/s, est. speed input: 194.34 toks/s, output: 5399.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:48<00:07,  1.04it/s, est. speed input: 189.95 toks/s, output: 5377.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:49<00:07,  1.06s/it, est. speed input: 186.22 toks/s, output: 5373.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2388.1 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:53<00:10,  1.76s/it, est. speed input: 175.48 toks/s, output: 5171.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:53<00:06,  1.36s/it, est. speed input: 175.26 toks/s, output: 5274.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:54<00:04,  1.06s/it, est. speed input: 175.23 toks/s, output: 5380.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:56<00:04,  1.35s/it, est. speed input: 169.93 toks/s, output: 5330.74 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.35s/it, est. speed input: 173.21 toks/s, output: 5767.94 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.85it/s, est. speed input: 173.21 toks/s, output: 5767.94 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:52 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:52 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:36:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:36:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:36:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:36:52 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:36:52 [abstract.py:306] It took 0.553505 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:36:52 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0113
+  • Training Accuracy: 0.3875
+  • Mean Generation Length: 2026.3500
+  • Training FLOPS: 41.93 TFLOPS (10.48 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 268.05s
+  • policy_training: 172.80s (64.5%)
+  • generation: 59.71s (22.3%)
+  • teacher_logprob_inference: 29.59s (11.0%)
+  • policy_training/sharding_data: 4.34s (1.6%)
+  • prepare_for_generation: 3.76s (1.4%)
+  • policy_training/submit_training_futures: 1.65s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.6%)
+  • training_prep: 1.37s (0.5%)
+  • teacher_logprob_inference_prep: 0.64s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 194: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:40:17 [abstract.py:324] It took 0.047932 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277344e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277344e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277344e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09277344e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66868439e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84475403e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66868439e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84475403e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66868439e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84475403e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66868439e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84475403e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09279727e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65877037e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83429565e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09279727e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65877037e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83429565e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09279727e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65877037e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83429565e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09279727e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.65877037e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.83429565e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=1.00437317e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.75539637e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #194] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44924713e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65164490e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36920109e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #194] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.59389801e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35805511e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.57944918e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #194] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.49972878e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.23526154e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65301018e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #194] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36223969e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 194: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:40:20 [abstract.py:324] It took 0.034058 seconds to wake up tags ['kv_cache'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
+(VllmGenerationWorker pid=10140) 
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12719.41it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:40:21 [loggers.py:236] Engine 000: Avg prompt throughput: 44.7 tokens/s, Avg generation throughput: 37.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:38,  4.09it/s, est. speed input: 249.33 toks/s, output: 53.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▏         | 3/160 [00:00<00:16,  9.73it/s, est. speed input: 527.17 toks/s, output: 165.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 6/160 [00:00<00:09, 15.88it/s, est. speed input: 796.31 toks/s, output: 303.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   5%|▌         | 8/160 [00:00<00:10, 15.05it/s, est. speed input: 810.79 toks/s, output: 384.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 21.95it/s, est. speed input: 1037.59 toks/s, output: 617.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:06, 22.76it/s, est. speed input: 1107.46 toks/s, output: 773.41 toks/s]
Processed prompts:  14%|█▍        | 22/160 [00:00<00:03, 35.83it/s, est. speed input: 1449.13 toks/s, output: 1259.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:01<00:02, 47.15it/s, est. speed input: 1764.96 toks/s, output: 1786.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▎       | 36/160 [00:01<00:02, 46.51it/s, est. speed input: 1876.30 toks/s, output: 2103.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 55.81it/s, est. speed input: 2144.85 toks/s, output: 2744.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 55/160 [00:01<00:01, 63.88it/s, est. speed input: 2370.36 toks/s, output: 3352.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 65/160 [00:01<00:01, 73.55it/s, est. speed input: 2617.81 toks/s, output: 4027.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:01<00:01, 69.67it/s, est. speed input: 2707.39 toks/s, output: 4442.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:01<00:01, 49.95it/s, est. speed input: 2582.52 toks/s, output: 4570.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:04<00:09,  8.01it/s, est. speed input: 1142.97 toks/s, output: 2251.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:09<00:21,  3.22it/s, est. speed input: 595.60 toks/s, output: 1605.78 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 14:40:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9888.1 tokens/s, Running: 68 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.9%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:11<00:25,  2.56it/s, est. speed input: 488.26 toks/s, output: 1609.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:13<00:26,  2.34it/s, est. speed input: 438.85 toks/s, output: 1754.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:17<00:37,  1.61it/s, est. speed input: 357.40 toks/s, output: 1641.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:17<00:33,  1.72it/s, est. speed input: 348.29 toks/s, output: 1813.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:19<00:33,  1.66it/s, est. speed input: 329.88 toks/s, output: 1923.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:40:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7703.0 tokens/s, Running: 56 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:20<00:35,  1.53it/s, est. speed input: 316.38 toks/s, output: 1950.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:20<00:31,  1.72it/s, est. speed input: 316.71 toks/s, output: 2058.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:20<00:27,  1.96it/s, est. speed input: 316.66 toks/s, output: 2162.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:21<00:21,  2.32it/s, est. speed input: 314.03 toks/s, output: 2350.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:21<00:19,  2.52it/s, est. speed input: 313.08 toks/s, output: 2446.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:21<00:14,  3.36it/s, est. speed input: 315.29 toks/s, output: 2665.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:22<00:22,  2.12it/s, est. speed input: 302.07 toks/s, output: 2655.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:23<00:26,  1.75it/s, est. speed input: 292.99 toks/s, output: 2677.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:23<00:21,  2.07it/s, est. speed input: 292.92 toks/s, output: 2778.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:24<00:23,  1.88it/s, est. speed input: 287.40 toks/s, output: 2827.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:25<00:24,  1.75it/s, est. speed input: 282.20 toks/s, output: 2876.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:26<00:36,  1.15it/s, est. speed input: 267.16 toks/s, output: 2825.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:29<00:58,  1.43s/it, est. speed input: 243.70 toks/s, output: 2680.87 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:40:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5893.2 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.2%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:30<00:54,  1.36s/it, est. speed input: 236.29 toks/s, output: 2703.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:31<00:35,  1.08it/s, est. speed input: 234.22 toks/s, output: 2884.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:31<00:27,  1.35it/s, est. speed input: 234.98 toks/s, output: 2996.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:32<00:26,  1.36it/s, est. speed input: 231.59 toks/s, output: 3055.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:33<00:22,  1.55it/s, est. speed input: 230.58 toks/s, output: 3144.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:33<00:18,  1.81it/s, est. speed input: 228.29 toks/s, output: 3313.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:34<00:21,  1.50it/s, est. speed input: 223.27 toks/s, output: 3341.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:35<00:18,  1.66it/s, est. speed input: 222.36 toks/s, output: 3429.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:35<00:15,  1.92it/s, est. speed input: 222.22 toks/s, output: 3527.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:36<00:16,  1.81it/s, est. speed input: 219.99 toks/s, output: 3592.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:36<00:12,  2.27it/s, est. speed input: 220.75 toks/s, output: 3704.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:39<00:34,  1.29s/it, est. speed input: 203.55 toks/s, output: 3518.45 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:41:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4474.5 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:40<00:26,  1.03s/it, est. speed input: 202.98 toks/s, output: 3612.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:41<00:26,  1.05s/it, est. speed input: 199.03 toks/s, output: 3646.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:41<00:15,  1.45it/s, est. speed input: 199.48 toks/s, output: 3857.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:42<00:12,  1.73it/s, est. speed input: 199.77 toks/s, output: 3965.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:42<00:09,  2.15it/s, est. speed input: 200.60 toks/s, output: 4082.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:42<00:08,  2.25it/s, est. speed input: 200.18 toks/s, output: 4176.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:43<00:12,  1.54it/s, est. speed input: 196.13 toks/s, output: 4194.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:44<00:10,  1.71it/s, est. speed input: 195.64 toks/s, output: 4286.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:44<00:09,  1.73it/s, est. speed input: 194.48 toks/s, output: 4364.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:45<00:09,  1.72it/s, est. speed input: 193.30 toks/s, output: 4439.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:46<00:09,  1.57it/s, est. speed input: 191.37 toks/s, output: 4498.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:46<00:04,  2.69it/s, est. speed input: 193.52 toks/s, output: 4754.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:46<00:04,  2.43it/s, est. speed input: 192.60 toks/s, output: 4834.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:47<00:06,  1.78it/s, est. speed input: 189.90 toks/s, output: 4869.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:48<00:05,  1.73it/s, est. speed input: 188.67 toks/s, output: 4942.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:49<00:05,  1.60it/s, est. speed input: 187.05 toks/s, output: 5004.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:49<00:04,  1.70it/s, est. speed input: 186.40 toks/s, output: 5092.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:41:11 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2845.4 tokens/s, Running: 8 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 74.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:55<00:15,  2.19s/it, est. speed input: 167.05 toks/s, output: 4678.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:56<00:10,  1.71s/it, est. speed input: 166.51 toks/s, output: 4778.02 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  1.71s/it, est. speed input: 172.94 toks/s, output: 5651.10 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:56<00:00,  2.84it/s, est. speed input: 172.94 toks/s, output: 5651.10 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:41:20 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:41:20 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:41:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:41:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:41:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:41:20 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:41:20 [abstract.py:306] It took 0.541031 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:41:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:41:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:41:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:41:20 [gpu_worker.py:142] Sleep mode freed 60.54 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0116
+  • Training Accuracy: 0.5312
+  • Mean Generation Length: 1988.3438
+  • Training FLOPS: 41.79 TFLOPS (10.45 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 266.66s
+  • policy_training: 172.22s (64.6%)
+  • generation: 59.77s (22.4%)
+  • teacher_logprob_inference: 29.57s (11.1%)
+  • policy_training/sharding_data: 4.32s (1.6%)
+  • prepare_for_generation: 3.68s (1.4%)
+  • policy_training/submit_training_futures: 1.70s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.6%)
+  • training_prep: 0.82s (0.3%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 195: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:44:45 [abstract.py:324] It took 0.051439 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09288750e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66504135e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85010147e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09288750e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66504135e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85010147e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09288750e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66504135e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85010147e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09288750e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.66504135e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85010147e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09299844e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63477631e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84560242e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.91652679e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.93743277e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #195] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44467957e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09299844e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63477631e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84560242e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62273026e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.37144375e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #195] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.57310028e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09299844e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63477631e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84560242e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.35612717e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.75407410e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #195] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.51000748e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09299844e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63477631e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.84560242e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.23716431e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.65441589e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #195] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36538666e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 195: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:44:48 [abstract.py:324] It took 0.028380 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13033.63it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 14:44:48 [loggers.py:236] Engine 000: Avg prompt throughput: 44.8 tokens/s, Avg generation throughput: 46.3 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:24,  6.48it/s, est. speed input: 375.82 toks/s, output: 77.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 10/160 [00:00<00:03, 43.43it/s, est. speed input: 2262.38 toks/s, output: 611.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:04, 32.05it/s, est. speed input: 1907.72 toks/s, output: 723.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:04, 34.21it/s, est. speed input: 2005.65 toks/s, output: 945.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▋        | 26/160 [00:00<00:03, 43.40it/s, est. speed input: 2315.84 toks/s, output: 1408.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:00<00:02, 48.12it/s, est. speed input: 2514.40 toks/s, output: 1843.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 40/160 [00:00<00:02, 48.78it/s, est. speed input: 2592.16 toks/s, output: 2160.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 46/160 [00:01<00:02, 48.42it/s, est. speed input: 2627.88 toks/s, output: 2464.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  33%|███▎      | 53/160 [00:01<00:02, 52.38it/s, est. speed input: 2734.54 toks/s, output: 2908.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:01<00:01, 56.86it/s, est. speed input: 2847.43 toks/s, output: 3356.44 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:01<00:01, 56.59it/s, est. speed input: 2890.67 toks/s, output: 3683.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:01<00:01, 53.12it/s, est. speed input: 2885.05 toks/s, output: 3972.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:01<00:01, 51.36it/s, est. speed input: 2882.68 toks/s, output: 4282.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:02<00:03, 24.30it/s, est. speed input: 2320.55 toks/s, output: 3806.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:03<00:07,  9.56it/s, est. speed input: 1522.16 toks/s, output: 2788.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:04<00:10,  6.80it/s, est. speed input: 1231.30 toks/s, output: 2517.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:06<00:18,  3.57it/s, est. speed input: 841.43 toks/s, output: 1995.98 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:07<00:19,  3.25it/s, est. speed input: 760.58 toks/s, output: 2017.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:08<00:22,  2.80it/s, est. speed input: 677.18 toks/s, output: 2008.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:09<00:18,  3.22it/s, est. speed input: 669.20 toks/s, output: 2197.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:44:58 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9718.8 tokens/s, Running: 60 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:10<00:30,  1.95it/s, est. speed input: 563.43 toks/s, output: 1961.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:11<00:29,  1.96it/s, est. speed input: 544.45 toks/s, output: 2005.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:12<00:33,  1.71it/s, est. speed input: 509.08 toks/s, output: 1986.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:13<00:37,  1.49it/s, est. speed input: 475.68 toks/s, output: 1968.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:13<00:25,  2.14it/s, est. speed input: 476.12 toks/s, output: 2188.67 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:14<00:30,  1.74it/s, est. speed input: 448.25 toks/s, output: 2171.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:15<00:36,  1.44it/s, est. speed input: 420.76 toks/s, output: 2149.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:15<00:29,  1.73it/s, est. speed input: 418.36 toks/s, output: 2248.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:16<00:22,  2.18it/s, est. speed input: 410.96 toks/s, output: 2425.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:17<00:20,  2.24it/s, est. speed input: 397.92 toks/s, output: 2561.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:17<00:17,  2.67it/s, est. speed input: 398.98 toks/s, output: 2676.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:17<00:13,  3.31it/s, est. speed input: 397.55 toks/s, output: 2880.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:17<00:11,  3.63it/s, est. speed input: 397.04 toks/s, output: 2983.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:19<00:19,  2.19it/s, est. speed input: 378.01 toks/s, output: 2948.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:08 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6933.9 tokens/s, Running: 41 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.0%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:20<00:24,  1.70it/s, est. speed input: 362.37 toks/s, output: 2934.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:21<00:28,  1.42it/s, est. speed input: 347.49 toks/s, output: 2922.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:21<00:25,  1.52it/s, est. speed input: 341.71 toks/s, output: 2982.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:24<00:35,  1.04it/s, est. speed input: 308.58 toks/s, output: 2901.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:24<00:31,  1.15it/s, est. speed input: 303.97 toks/s, output: 2968.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:25<00:23,  1.43it/s, est. speed input: 298.14 toks/s, output: 3127.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:26<00:25,  1.27it/s, est. speed input: 288.28 toks/s, output: 3133.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:27<00:27,  1.18it/s, est. speed input: 279.97 toks/s, output: 3151.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:29<00:29,  1.06it/s, est. speed input: 270.27 toks/s, output: 3154.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:29<00:18,  1.59it/s, est. speed input: 270.60 toks/s, output: 3377.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:29<00:14,  1.93it/s, est. speed input: 271.20 toks/s, output: 3495.64 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:18 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5047.5 tokens/s, Running: 27 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:30<00:14,  1.78it/s, est. speed input: 264.08 toks/s, output: 3618.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:31<00:15,  1.56it/s, est. speed input: 258.40 toks/s, output: 3651.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:32<00:15,  1.56it/s, est. speed input: 255.10 toks/s, output: 3714.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:33<00:12,  1.70it/s, est. speed input: 250.87 toks/s, output: 3870.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:36<00:21,  1.03s/it, est. speed input: 234.65 toks/s, output: 3734.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:38<00:25,  1.29s/it, est. speed input: 223.46 toks/s, output: 3672.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:38<00:19,  1.05s/it, est. speed input: 223.01 toks/s, output: 3779.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:38<00:14,  1.25it/s, est. speed input: 223.81 toks/s, output: 3907.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:38<00:11,  1.48it/s, est. speed input: 223.43 toks/s, output: 4014.44 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:28 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3541.3 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:39<00:12,  1.29it/s, est. speed input: 219.17 toks/s, output: 4053.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:40<00:09,  1.56it/s, est. speed input: 218.97 toks/s, output: 4162.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:40<00:06,  2.06it/s, est. speed input: 219.89 toks/s, output: 4293.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:41<00:07,  1.78it/s, est. speed input: 217.38 toks/s, output: 4358.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:42<00:09,  1.28it/s, est. speed input: 212.10 toks/s, output: 4367.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:44<00:13,  1.23s/it, est. speed input: 202.60 toks/s, output: 4291.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:44<00:09,  1.09it/s, est. speed input: 203.11 toks/s, output: 4421.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:46<00:10,  1.17s/it, est. speed input: 196.70 toks/s, output: 4402.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:46<00:06,  1.16it/s, est. speed input: 197.47 toks/s, output: 4538.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:46<00:04,  1.57it/s, est. speed input: 198.30 toks/s, output: 4677.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:47<00:03,  1.91it/s, est. speed input: 198.47 toks/s, output: 4800.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:47<00:02,  2.08it/s, est. speed input: 198.12 toks/s, output: 4912.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:49<00:03,  1.19it/s, est. speed input: 192.58 toks/s, output: 4897.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:49<00:01,  1.50it/s, est. speed input: 192.78 toks/s, output: 5025.71 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:38 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1944.4 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 76.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:51<00:02,  1.16s/it, est. speed input: 185.32 toks/s, output: 4958.09 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  1.16s/it, est. speed input: 187.60 toks/s, output: 5274.03 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:51<00:00,  3.09it/s, est. speed input: 187.60 toks/s, output: 5274.03 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:42 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:42 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:45:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:45:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:45:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:45:42 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:45:42 [abstract.py:306] It took 0.514526 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:45:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:45:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:45:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:45:42 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0115
+  • Training Accuracy: 0.4250
+  • Mean Generation Length: 1709.3000
+  • Training FLOPS: 39.25 TFLOPS (9.81 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 242.57s
+  • policy_training: 157.18s (64.8%)
+  • generation: 54.84s (22.6%)
+  • teacher_logprob_inference: 25.39s (10.5%)
+  • policy_training/sharding_data: 4.28s (1.8%)
+  • prepare_for_generation: 3.71s (1.5%)
+  • policy_training/submit_training_futures: 1.69s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.62s (0.7%)
+  • training_prep: 0.86s (0.4%)
+  • teacher_logprob_inference_prep: 0.41s (0.2%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.02s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.02s (0.0%)
+
+========================= Epoch 40/40 =========================
+
+========================= Step 1/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 196: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:48:47 [abstract.py:324] It took 0.052646 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09349102e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64951401e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84997787e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09349102e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64951401e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84997787e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09349102e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64951401e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84997787e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09349102e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.64951401e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84997787e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353594e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62453918e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85182037e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.95249939e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.79338455e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #196] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44021469e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353594e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62453918e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85182037e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.62582397e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.38258820e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #196] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.53970795e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353594e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62453918e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85182037e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37479248e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.37121201e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #196] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.52511139e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353594e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.62453918e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85182037e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.24761429e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.64736252e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #196] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36333160e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 196: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:48:50 [abstract.py:324] It took 0.030579 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13893.60it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 14:48:50 [loggers.py:236] Engine 000: Avg prompt throughput: 50.6 tokens/s, Avg generation throughput: 7.9 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:40,  3.91it/s, est. speed input: 246.30 toks/s, output: 39.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   2%|▎         | 4/160 [00:00<00:13, 11.33it/s, est. speed input: 614.83 toks/s, output: 171.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:07, 20.85it/s, est. speed input: 1032.09 toks/s, output: 415.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:05, 25.41it/s, est. speed input: 1231.54 toks/s, output: 601.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█▏        | 18/160 [00:00<00:04, 30.45it/s, est. speed input: 1442.64 toks/s, output: 858.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:04, 33.05it/s, est. speed input: 1555.54 toks/s, output: 1062.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  22%|██▏       | 35/160 [00:01<00:02, 54.92it/s, est. speed input: 2138.31 toks/s, output: 1918.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 52.78it/s, est. speed input: 2226.79 toks/s, output: 2205.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 54/160 [00:01<00:01, 71.71it/s, est. speed input: 2671.37 toks/s, output: 3095.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 60.52it/s, est. speed input: 2673.71 toks/s, output: 3371.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  43%|████▎     | 69/160 [00:01<00:01, 58.45it/s, est. speed input: 2721.23 toks/s, output: 3712.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:02<00:03, 23.92it/s, est. speed input: 2023.89 toks/s, output: 3037.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:07<00:21,  3.64it/s, est. speed input: 634.32 toks/s, output: 1316.67 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:00 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10303.2 tokens/s, Running: 78 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:12<00:32,  2.31it/s, est. speed input: 431.17 toks/s, output: 1261.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:12<00:29,  2.45it/s, est. speed input: 416.49 toks/s, output: 1515.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:14<00:30,  2.28it/s, est. speed input: 388.93 toks/s, output: 1612.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:15<00:31,  2.19it/s, est. speed input: 368.81 toks/s, output: 1719.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:17<00:44,  1.50it/s, est. speed input: 322.24 toks/s, output: 1600.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:17<00:34,  1.87it/s, est. speed input: 325.40 toks/s, output: 1809.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:18<00:34,  1.86it/s, est. speed input: 318.74 toks/s, output: 1868.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:18<00:29,  2.15it/s, est. speed input: 320.15 toks/s, output: 1971.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:18<00:27,  2.25it/s, est. speed input: 317.29 toks/s, output: 2049.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:19<00:26,  2.28it/s, est. speed input: 313.44 toks/s, output: 2119.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:20<00:30,  1.94it/s, est. speed input: 304.56 toks/s, output: 2155.22 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:10 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7958.4 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:20<00:21,  2.70it/s, est. speed input: 305.94 toks/s, output: 2351.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:20<00:23,  2.47it/s, est. speed input: 301.28 toks/s, output: 2407.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:22<00:45,  1.23it/s, est. speed input: 276.44 toks/s, output: 2305.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:23<00:36,  1.50it/s, est. speed input: 276.33 toks/s, output: 2397.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:23<00:28,  1.89it/s, est. speed input: 277.09 toks/s, output: 2496.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:23<00:29,  1.82it/s, est. speed input: 272.63 toks/s, output: 2550.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:24<00:20,  2.50it/s, est. speed input: 272.86 toks/s, output: 2736.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:24<00:18,  2.76it/s, est. speed input: 272.67 toks/s, output: 2826.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:24<00:15,  3.21it/s, est. speed input: 273.26 toks/s, output: 2923.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:24<00:12,  3.88it/s, est. speed input: 274.55 toks/s, output: 3027.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:25<00:10,  4.32it/s, est. speed input: 275.25 toks/s, output: 3124.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:25<00:10,  4.30it/s, est. speed input: 275.09 toks/s, output: 3211.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:25<00:14,  3.19it/s, est. speed input: 271.98 toks/s, output: 3264.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:26<00:16,  2.72it/s, est. speed input: 269.17 toks/s, output: 3320.65 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:27<00:25,  1.68it/s, est. speed input: 260.11 toks/s, output: 3299.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:27<00:22,  1.90it/s, est. speed input: 258.88 toks/s, output: 3374.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:28<00:23,  1.71it/s, est. speed input: 254.52 toks/s, output: 3407.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:29<00:23,  1.67it/s, est. speed input: 251.11 toks/s, output: 3451.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:29<00:23,  1.63it/s, est. speed input: 247.68 toks/s, output: 3495.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:20 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6117.6 tokens/s, Running: 39 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.4%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:30<00:26,  1.42it/s, est. speed input: 242.31 toks/s, output: 3510.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:32<00:33,  1.10it/s, est. speed input: 233.72 toks/s, output: 3478.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:32<00:22,  1.54it/s, est. speed input: 232.43 toks/s, output: 3642.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:32<00:17,  1.94it/s, est. speed input: 233.51 toks/s, output: 3750.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:35<00:33,  1.01s/it, est. speed input: 219.30 toks/s, output: 3616.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:35<00:28,  1.12it/s, est. speed input: 217.47 toks/s, output: 3679.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:37<00:35,  1.14s/it, est. speed input: 208.70 toks/s, output: 3627.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:38<00:29,  1.03it/s, est. speed input: 207.34 toks/s, output: 3699.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:38<00:20,  1.39it/s, est. speed input: 208.30 toks/s, output: 3812.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:39<00:17,  1.51it/s, est. speed input: 205.18 toks/s, output: 3940.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:30 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4540.4 tokens/s, Running: 26 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.9%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:40<00:16,  1.50it/s, est. speed input: 201.37 toks/s, output: 4053.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:41<00:13,  1.82it/s, est. speed input: 202.19 toks/s, output: 4164.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:43<00:24,  1.06s/it, est. speed input: 191.16 toks/s, output: 4034.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:45<00:25,  1.16s/it, est. speed input: 186.29 toks/s, output: 4029.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:45<00:19,  1.06it/s, est. speed input: 186.18 toks/s, output: 4125.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:45<00:15,  1.31it/s, est. speed input: 186.29 toks/s, output: 4226.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:46<00:16,  1.15it/s, est. speed input: 183.11 toks/s, output: 4252.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:47<00:11,  1.53it/s, est. speed input: 183.89 toks/s, output: 4368.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:47<00:11,  1.50it/s, est. speed input: 182.51 toks/s, output: 4433.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:48<00:09,  1.62it/s, est. speed input: 181.83 toks/s, output: 4515.64 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:49<00:10,  1.45it/s, est. speed input: 179.90 toks/s, output: 4566.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:50<00:10,  1.32it/s, est. speed input: 177.77 toks/s, output: 4612.56 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:40 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3299.9 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.2%, Prefix cache hit rate: 75.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:50<00:07,  1.64it/s, est. speed input: 178.07 toks/s, output: 4719.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:11,  1.06it/s, est. speed input: 173.31 toks/s, output: 4694.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:52<00:09,  1.17it/s, est. speed input: 172.36 toks/s, output: 4769.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:53<00:07,  1.34it/s, est. speed input: 171.94 toks/s, output: 4860.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:54<00:07,  1.23it/s, est. speed input: 170.04 toks/s, output: 4908.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:54<00:05,  1.58it/s, est. speed input: 170.50 toks/s, output: 5024.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:54<00:03,  1.83it/s, est. speed input: 170.51 toks/s, output: 5127.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:56<00:03,  1.29it/s, est. speed input: 166.38 toks/s, output: 5209.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:58<00:03,  1.11it/s, est. speed input: 163.65 toks/s, output: 5234.32 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  1.11it/s, est. speed input: 167.90 toks/s, output: 5798.48 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:58<00:00,  2.75it/s, est. speed input: 167.90 toks/s, output: 5798.48 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:51 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:51 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:49:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:49:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:49:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:49:51 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:49:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:49:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:49:52 [abstract.py:306] It took 0.545242 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:49:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:49:52 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0105
+  • Training Accuracy: 0.4938
+  • Mean Generation Length: 2104.8687
+  • Training FLOPS: 43.62 TFLOPS (10.90 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 271.75s
+  • policy_training: 174.07s (64.1%)
+  • generation: 61.71s (22.7%)
+  • teacher_logprob_inference: 30.30s (11.1%)
+  • policy_training/sharding_data: 4.29s (1.6%)
+  • prepare_for_generation: 3.75s (1.4%)
+  • policy_training/submit_training_futures: 1.69s (0.6%)
+  • prepare_for_generation/transfer_and_update_weights: 1.59s (0.6%)
+  • training_prep: 1.08s (0.4%)
+  • teacher_logprob_inference_prep: 0.67s (0.2%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 2/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 197: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:53:19 [abstract.py:324] It took 0.050562 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09360195e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322601e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85087509e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09360195e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322601e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85087509e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09360195e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322601e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85087509e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09360195e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.63322601e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85087509e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09362773e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63321953e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85654221e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.95639038e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.72264147e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #197] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.44014740e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09362773e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63321953e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85654221e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.65798416e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36807003e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #197] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.55692444e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09362773e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63321953e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85654221e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37427521e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.71871519e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #197] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.54260559e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09362773e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.63321953e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85654221e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26938934e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62628918e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #197] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36226852e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 197: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:53:22 [abstract.py:324] It took 0.029015 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 13912.61it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(VllmGenerationWorker pid=10140) INFO 05-03 14:53:22 [loggers.py:236] Engine 000: Avg prompt throughput: 43.8 tokens/s, Avg generation throughput: 65.2 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:44,  3.57it/s, est. speed input: 224.74 toks/s, output: 85.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:11, 13.71it/s, est. speed input: 726.20 toks/s, output: 360.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   4%|▍         | 7/160 [00:00<00:10, 14.15it/s, est. speed input: 775.84 toks/s, output: 442.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 12/160 [00:00<00:06, 21.53it/s, est. speed input: 1060.66 toks/s, output: 750.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 20/160 [00:00<00:04, 34.69it/s, est. speed input: 1503.70 toks/s, output: 1331.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  17%|█▋        | 27/160 [00:00<00:03, 43.78it/s, est. speed input: 1802.83 toks/s, output: 1810.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  20%|██        | 32/160 [00:01<00:02, 43.79it/s, est. speed input: 1895.17 toks/s, output: 2063.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 51.38it/s, est. speed input: 2141.75 toks/s, output: 2637.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 42.02it/s, est. speed input: 2085.59 toks/s, output: 2787.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:03, 33.72it/s, est. speed input: 1976.07 toks/s, output: 2842.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  35%|███▌      | 56/160 [00:01<00:03, 31.56it/s, est. speed input: 1940.61 toks/s, output: 2991.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 60/160 [00:02<00:03, 25.05it/s, est. speed input: 1808.97 toks/s, output: 2994.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:02<00:08, 11.34it/s, est. speed input: 1353.97 toks/s, output: 2429.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████▏     | 66/160 [00:03<00:10,  8.80it/s, est. speed input: 1171.64 toks/s, output: 2317.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:05<00:23,  3.88it/s, est. speed input: 787.62 toks/s, output: 1713.05 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:09<01:00,  1.49it/s, est. speed input: 434.16 toks/s, output: 1092.01 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:53:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 11072.9 tokens/s, Running: 89 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  45%|████▌     | 72/160 [00:11<00:59,  1.47it/s, est. speed input: 390.26 toks/s, output: 1162.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▋     | 74/160 [00:12<01:00,  1.42it/s, est. speed input: 351.58 toks/s, output: 1226.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:16<01:32,  1.09s/it, est. speed input: 283.60 toks/s, output: 1084.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 76/160 [00:17<01:37,  1.16s/it, est. speed input: 263.04 toks/s, output: 1100.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:18<01:32,  1.11s/it, est. speed input: 253.23 toks/s, output: 1153.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:19<01:08,  1.18it/s, est. speed input: 249.69 toks/s, output: 1319.67 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:53:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8850.5 tokens/s, Running: 81 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.0%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:21<01:25,  1.07s/it, est. speed input: 229.82 toks/s, output: 1306.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:21<01:12,  1.10it/s, est. speed input: 228.51 toks/s, output: 1389.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:21<00:56,  1.39it/s, est. speed input: 229.92 toks/s, output: 1487.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:22<00:48,  1.59it/s, est. speed input: 228.73 toks/s, output: 1568.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:22<00:43,  1.73it/s, est. speed input: 227.01 toks/s, output: 1644.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:23<00:44,  1.68it/s, est. speed input: 223.32 toks/s, output: 1704.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:23<00:44,  1.66it/s, est. speed input: 220.13 toks/s, output: 1766.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:24<00:50,  1.44it/s, est. speed input: 214.36 toks/s, output: 1805.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  55%|█████▌    | 88/160 [00:25<00:57,  1.26it/s, est. speed input: 208.20 toks/s, output: 1837.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:25<00:44,  1.60it/s, est. speed input: 208.73 toks/s, output: 1926.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 90/160 [00:26<00:41,  1.68it/s, est. speed input: 206.96 toks/s, output: 1990.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:26<00:35,  1.92it/s, est. speed input: 206.62 toks/s, output: 2068.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▊    | 92/160 [00:27<00:31,  2.14it/s, est. speed input: 206.18 toks/s, output: 2144.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:27<00:26,  2.48it/s, est. speed input: 205.59 toks/s, output: 2297.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:28<00:34,  1.89it/s, est. speed input: 201.15 toks/s, output: 2326.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:29<00:29,  2.21it/s, est. speed input: 201.62 toks/s, output: 2410.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:29<00:21,  2.94it/s, est. speed input: 203.09 toks/s, output: 2584.81 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:53:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6873.0 tokens/s, Running: 62 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.9%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:30<00:32,  1.89it/s, est. speed input: 197.41 toks/s, output: 2590.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:31<00:36,  1.64it/s, est. speed input: 194.04 toks/s, output: 2624.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:31<00:30,  1.92it/s, est. speed input: 194.33 toks/s, output: 2705.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:32<00:35,  1.63it/s, est. speed input: 191.00 toks/s, output: 2737.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:32<00:30,  1.90it/s, est. speed input: 191.10 toks/s, output: 2816.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:33<00:27,  2.04it/s, est. speed input: 190.69 toks/s, output: 2886.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:34<00:41,  1.31it/s, est. speed input: 184.52 toks/s, output: 2871.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:34<00:32,  1.66it/s, est. speed input: 185.10 toks/s, output: 2958.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:35<00:25,  2.05it/s, est. speed input: 185.72 toks/s, output: 3044.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:35<00:27,  1.90it/s, est. speed input: 184.18 toks/s, output: 3096.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:35<00:15,  3.25it/s, est. speed input: 186.95 toks/s, output: 3295.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:36<00:15,  3.22it/s, est. speed input: 187.01 toks/s, output: 3371.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:36<00:19,  2.49it/s, est. speed input: 185.25 toks/s, output: 3415.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:37<00:28,  1.68it/s, est. speed input: 181.42 toks/s, output: 3419.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:38<00:21,  2.13it/s, est. speed input: 182.35 toks/s, output: 3512.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:38<00:16,  2.73it/s, est. speed input: 183.39 toks/s, output: 3607.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:38<00:20,  2.10it/s, est. speed input: 181.44 toks/s, output: 3644.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:39<00:12,  3.50it/s, est. speed input: 184.01 toks/s, output: 3845.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:39<00:08,  4.45it/s, est. speed input: 185.90 toks/s, output: 4030.63 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:54:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6034.2 tokens/s, Running: 40 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:40<00:14,  2.61it/s, est. speed input: 183.02 toks/s, output: 4042.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:40<00:13,  2.83it/s, est. speed input: 183.34 toks/s, output: 4123.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:41<00:15,  2.36it/s, est. speed input: 182.01 toks/s, output: 4166.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:41<00:11,  2.97it/s, est. speed input: 182.99 toks/s, output: 4335.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:42<00:13,  2.58it/s, est. speed input: 182.02 toks/s, output: 4385.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:08,  3.94it/s, est. speed input: 184.38 toks/s, output: 4590.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:43<00:12,  2.56it/s, est. speed input: 182.03 toks/s, output: 4605.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:43<00:10,  2.75it/s, est. speed input: 182.26 toks/s, output: 4684.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:15,  1.88it/s, est. speed input: 179.37 toks/s, output: 4685.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:44<00:13,  2.08it/s, est. speed input: 179.40 toks/s, output: 4759.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:45<00:12,  2.16it/s, est. speed input: 179.08 toks/s, output: 4825.48 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:45<00:09,  2.62it/s, est. speed input: 179.55 toks/s, output: 4985.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:45<00:08,  2.84it/s, est. speed input: 179.92 toks/s, output: 5067.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:46<00:07,  2.91it/s, est. speed input: 180.01 toks/s, output: 5144.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:47<00:15,  1.43it/s, est. speed input: 174.93 toks/s, output: 5074.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:48<00:14,  1.43it/s, est. speed input: 173.70 toks/s, output: 5114.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:48<00:11,  1.69it/s, est. speed input: 173.79 toks/s, output: 5193.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:07,  2.42it/s, est. speed input: 174.89 toks/s, output: 5377.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:07,  2.16it/s, est. speed input: 173.90 toks/s, output: 5423.96 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:54:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3906.2 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:51<00:09,  1.62it/s, est. speed input: 171.50 toks/s, output: 5426.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:51<00:10,  1.42it/s, est. speed input: 169.53 toks/s, output: 5443.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:53<00:11,  1.18it/s, est. speed input: 166.77 toks/s, output: 5436.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:54<00:09,  1.30it/s, est. speed input: 164.86 toks/s, output: 5537.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:54<00:07,  1.56it/s, est. speed input: 165.27 toks/s, output: 5632.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:55<00:05,  1.75it/s, est. speed input: 165.25 toks/s, output: 5714.75 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:05,  1.65it/s, est. speed input: 164.23 toks/s, output: 5764.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:56<00:04,  1.92it/s, est. speed input: 164.44 toks/s, output: 5854.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:56<00:03,  1.91it/s, est. speed input: 164.02 toks/s, output: 5922.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:57<00:02,  2.31it/s, est. speed input: 164.32 toks/s, output: 6100.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:57<00:01,  2.66it/s, est. speed input: 164.79 toks/s, output: 6203.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:57<00:01,  2.82it/s, est. speed input: 165.01 toks/s, output: 6296.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:58<00:01,  1.80it/s, est. speed input: 162.96 toks/s, output: 6305.26 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:54:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1912.2 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 75.0%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [01:01<00:01,  1.16s/it, est. speed input: 156.72 toks/s, output: 6159.84 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.16s/it, est. speed input: 157.68 toks/s, output: 6292.83 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.60it/s, est. speed input: 157.68 toks/s, output: 6292.83 toks/s]
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:54:27 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:54:27 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:54:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:54:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:54:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:54:27 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+  [GT guidance] avg length (words): 8, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:54:27 [abstract.py:306] It took 0.553017 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:54:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:54:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:54:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:54:27 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0097
+  • Training Accuracy: 0.3812
+  • Mean Generation Length: 2422.4937
+  • Training FLOPS: 44.43 TFLOPS (11.11 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 300.84s
+  • policy_training: 194.49s (64.6%)
+  • generation: 65.63s (21.8%)
+  • teacher_logprob_inference: 35.26s (11.7%)
+  • policy_training/sharding_data: 4.30s (1.4%)
+  • prepare_for_generation: 3.62s (1.2%)
+  • policy_training/submit_training_futures: 2.22s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.5%)
+  • training_prep: 1.16s (0.4%)
+  • teacher_logprob_inference_prep: 0.51s (0.2%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • data_processing: 0.03s (0.0%)
+
+========================= Step 3/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 198: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 14:58:20 [abstract.py:324] It took 0.048441 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09364922e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61493607e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85239258e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09364922e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61493607e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85239258e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09364922e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61493607e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85239258e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09364922e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.61493607e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85239258e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09368125e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59723930e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86540756e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.86687088e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.63857889e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #198] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43539978e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09368125e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59723930e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86540756e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63152008e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.35476198e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #198] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.55554199e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09368125e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59723930e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86540756e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38542786e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.68201542e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #198] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.52766876e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09368125e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59723930e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86540756e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26609802e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62329922e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #198] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36429489e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+[DEBUG refit] Step 198: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 14:58:23 [abstract.py:324] It took 0.028464 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12907.78it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:58:23 [loggers.py:236] Engine 000: Avg prompt throughput: 40.2 tokens/s, Avg generation throughput: 4.0 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:48,  3.31it/s, est. speed input: 198.60 toks/s, output: 36.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:11, 13.10it/s, est. speed input: 678.66 toks/s, output: 184.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   7%|▋         | 11/160 [00:00<00:06, 24.08it/s, est. speed input: 1143.65 toks/s, output: 460.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 15/160 [00:00<00:05, 26.18it/s, est. speed input: 1283.50 toks/s, output: 626.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▏        | 19/160 [00:00<00:05, 26.97it/s, est. speed input: 1363.24 toks/s, output: 810.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  15%|█▌        | 24/160 [00:00<00:04, 32.79it/s, est. speed input: 1544.76 toks/s, output: 1089.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:01<00:04, 32.29it/s, est. speed input: 1587.33 toks/s, output: 1264.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  21%|██▏       | 34/160 [00:01<00:03, 35.57it/s, est. speed input: 1697.97 toks/s, output: 1587.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  26%|██▌       | 41/160 [00:01<00:02, 43.31it/s, est. speed input: 1882.31 toks/s, output: 2035.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  29%|██▉       | 47/160 [00:01<00:02, 45.62it/s, est. speed input: 1982.58 toks/s, output: 2376.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:02, 45.90it/s, est. speed input: 2043.10 toks/s, output: 2644.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  36%|███▋      | 58/160 [00:01<00:02, 44.36it/s, est. speed input: 2083.24 toks/s, output: 2937.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 63/160 [00:01<00:02, 35.86it/s, est. speed input: 2012.35 toks/s, output: 3060.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▏     | 67/160 [00:02<00:04, 19.18it/s, est. speed input: 1685.70 toks/s, output: 2771.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:06<00:28,  3.13it/s, est. speed input: 657.72 toks/s, output: 1291.85 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  46%|████▌     | 73/160 [00:06<00:24,  3.51it/s, est. speed input: 637.02 toks/s, output: 1519.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:07<00:26,  3.22it/s, est. speed input: 582.73 toks/s, output: 1571.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  48%|████▊     | 77/160 [00:08<00:28,  2.93it/s, est. speed input: 534.85 toks/s, output: 1624.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:10<00:38,  2.13it/s, est. speed input: 469.44 toks/s, output: 1518.88 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:58:33 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10788.3 tokens/s, Running: 82 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.3%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 79/160 [00:11<00:51,  1.56it/s, est. speed input: 411.00 toks/s, output: 1423.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 80/160 [00:11<00:43,  1.82it/s, est. speed input: 411.30 toks/s, output: 1518.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████     | 81/160 [00:13<00:59,  1.34it/s, est. speed input: 368.95 toks/s, output: 1456.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▏    | 83/160 [00:13<00:45,  1.71it/s, est. speed input: 361.30 toks/s, output: 1608.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  52%|█████▎    | 84/160 [00:14<00:39,  1.94it/s, est. speed input: 359.38 toks/s, output: 1690.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:14<00:40,  1.87it/s, est. speed input: 348.94 toks/s, output: 1733.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 86/160 [00:15<00:45,  1.63it/s, est. speed input: 334.03 toks/s, output: 1749.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:17<01:04,  1.14it/s, est. speed input: 305.89 toks/s, output: 1694.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▌    | 89/160 [00:18<00:48,  1.45it/s, est. speed input: 297.97 toks/s, output: 1829.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:19<00:41,  1.68it/s, est. speed input: 290.37 toks/s, output: 1959.54 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  58%|█████▊    | 93/160 [00:19<00:31,  2.14it/s, est. speed input: 289.70 toks/s, output: 2131.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:19<00:29,  2.21it/s, est. speed input: 286.92 toks/s, output: 2199.34 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:20<00:26,  2.41it/s, est. speed input: 286.04 toks/s, output: 2278.93 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:58:43 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 8263.0 tokens/s, Running: 65 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.7%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:23<01:12,  1.13s/it, est. speed input: 247.66 toks/s, output: 2063.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████    | 97/160 [00:24<01:14,  1.18s/it, est. speed input: 236.96 toks/s, output: 2064.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:24<00:55,  1.12it/s, est. speed input: 238.37 toks/s, output: 2166.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▏   | 99/160 [00:26<01:02,  1.03s/it, est. speed input: 228.15 toks/s, output: 2164.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:26<00:38,  1.52it/s, est. speed input: 229.37 toks/s, output: 2353.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:26<00:30,  1.89it/s, est. speed input: 230.58 toks/s, output: 2454.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:27<00:30,  1.85it/s, est. speed input: 228.05 toks/s, output: 2514.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:28<00:19,  2.79it/s, est. speed input: 229.82 toks/s, output: 2790.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:29<00:29,  1.80it/s, est. speed input: 221.30 toks/s, output: 2773.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:29<00:26,  1.94it/s, est. speed input: 220.49 toks/s, output: 2850.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:58:53 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6926.7 tokens/s, Running: 52 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.1%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:30<00:35,  1.45it/s, est. speed input: 213.58 toks/s, output: 2848.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:31<00:35,  1.42it/s, est. speed input: 210.51 toks/s, output: 2893.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:32<00:22,  2.15it/s, est. speed input: 212.39 toks/s, output: 3091.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:32<00:22,  2.12it/s, est. speed input: 211.05 toks/s, output: 3157.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:23,  1.94it/s, est. speed input: 208.68 toks/s, output: 3207.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:33<00:22,  2.04it/s, est. speed input: 207.88 toks/s, output: 3280.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:33<00:17,  2.58it/s, est. speed input: 209.00 toks/s, output: 3382.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:34<00:16,  2.61it/s, est. speed input: 208.48 toks/s, output: 3458.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:34<00:16,  2.48it/s, est. speed input: 207.56 toks/s, output: 3525.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:35<00:18,  2.23it/s, est. speed input: 206.02 toks/s, output: 3582.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:35<00:15,  2.60it/s, est. speed input: 206.36 toks/s, output: 3672.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:36<00:19,  1.97it/s, est. speed input: 202.36 toks/s, output: 3764.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:39<00:40,  1.10s/it, est. speed input: 188.95 toks/s, output: 3601.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:59:03 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5281.9 tokens/s, Running: 37 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.8%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:40<00:36,  1.02s/it, est. speed input: 186.70 toks/s, output: 3645.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:40<00:30,  1.13it/s, est. speed input: 185.88 toks/s, output: 3714.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:41<00:23,  1.45it/s, est. speed input: 186.49 toks/s, output: 3811.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 127/160 [00:42<00:34,  1.05s/it, est. speed input: 179.49 toks/s, output: 3754.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:43<00:25,  1.27it/s, est. speed input: 180.32 toks/s, output: 3857.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████  | 129/160 [00:43<00:19,  1.56it/s, est. speed input: 180.52 toks/s, output: 3947.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:43<00:17,  1.70it/s, est. speed input: 179.97 toks/s, output: 4021.71 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:44<00:13,  2.09it/s, est. speed input: 180.47 toks/s, output: 4118.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:44<00:15,  1.86it/s, est. speed input: 179.06 toks/s, output: 4172.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:45<00:13,  2.01it/s, est. speed input: 178.82 toks/s, output: 4252.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:46<00:20,  1.28it/s, est. speed input: 174.54 toks/s, output: 4236.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:47<00:20,  1.21it/s, est. speed input: 172.37 toks/s, output: 4272.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▋ | 138/160 [00:47<00:08,  2.56it/s, est. speed input: 175.66 toks/s, output: 4613.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:47<00:07,  2.76it/s, est. speed input: 175.97 toks/s, output: 4707.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:48<00:07,  2.62it/s, est. speed input: 175.57 toks/s, output: 4782.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:48<00:06,  2.74it/s, est. speed input: 175.65 toks/s, output: 4871.01 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:49<00:09,  1.98it/s, est. speed input: 173.67 toks/s, output: 4902.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:49<00:07,  2.36it/s, est. speed input: 174.14 toks/s, output: 5002.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:59:13 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3926.5 tokens/s, Running: 16 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.5%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:50<00:06,  2.43it/s, est. speed input: 174.05 toks/s, output: 5085.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:50<00:05,  2.59it/s, est. speed input: 174.12 toks/s, output: 5173.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:51<00:09,  1.52it/s, est. speed input: 170.78 toks/s, output: 5163.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:51<00:06,  2.01it/s, est. speed input: 171.60 toks/s, output: 5275.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:52<00:05,  2.31it/s, est. speed input: 171.80 toks/s, output: 5370.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:52<00:03,  2.97it/s, est. speed input: 172.57 toks/s, output: 5482.29 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:52<00:03,  2.74it/s, est. speed input: 172.29 toks/s, output: 5561.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:55<00:09,  1.10s/it, est. speed input: 164.57 toks/s, output: 5405.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:55<00:06,  1.18it/s, est. speed input: 164.89 toks/s, output: 5508.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:57<00:06,  1.03it/s, est. speed input: 162.27 toks/s, output: 5516.79 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 14:59:23 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1913.8 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 75.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [01:00<00:10,  1.73s/it, est. speed input: 153.93 toks/s, output: 5333.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [01:01<00:06,  1.33s/it, est. speed input: 153.92 toks/s, output: 5432.54 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  1.33s/it, est. speed input: 158.88 toks/s, output: 6103.72 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:01<00:00,  2.62it/s, est. speed input: 158.88 toks/s, output: 6103.72 toks/s]
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 14:59:28 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 14:59:28 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 14:59:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:59:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:59:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:59:28 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 14:59:28 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 14:59:28 [abstract.py:306] It took 0.537703 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 14:59:28 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 14:59:28 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 14:59:28 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0103
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 2328.0125
+  • Training FLOPS: 43.28 TFLOPS (10.82 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 298.17s
+  • policy_training: 193.44s (64.9%)
+  • generation: 65.11s (21.8%)
+  • teacher_logprob_inference: 34.11s (11.4%)
+  • policy_training/sharding_data: 4.33s (1.5%)
+  • prepare_for_generation: 3.75s (1.3%)
+  • policy_training/submit_training_futures: 1.99s (0.7%)
+  • prepare_for_generation/transfer_and_update_weights: 1.65s (0.6%)
+  • training_prep: 1.10s (0.4%)
+  • teacher_logprob_inference_prep: 0.49s (0.2%)
+  • get_topk_logits/shard_data: 0.04s (0.0%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+
+========================= Step 4/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 199: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 15:03:18 [abstract.py:324] It took 0.051144 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09368555e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09368555e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09368555e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09368555e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09372656e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59323196e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86638641e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60132980e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85455627e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60132980e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85455627e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60132980e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85455627e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.60132980e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.85455627e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09372656e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59323196e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86638641e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.73975754e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.50698185e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #199] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43594025e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09372656e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59323196e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86638641e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.64822311e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36421432e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #199] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.61133575e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.37732239e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-5.50868511e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #199] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.54322929e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09372656e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.59323196e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.86638641e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.27327957e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62318611e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #199] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36613190e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 199: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 15:03:21 [abstract.py:324] It took 0.029333 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 15:03:21 [loggers.py:236] Engine 000: Avg prompt throughput: 40.9 tokens/s, Avg generation throughput: 5.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 12591.73it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:24,  6.44it/s, est. speed input: 386.61 toks/s, output: 77.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:02, 59.38it/s, est. speed input: 3048.86 toks/s, output: 840.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  14%|█▍        | 22/160 [00:00<00:01, 69.05it/s, est. speed input: 3629.03 toks/s, output: 1230.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 30/160 [00:00<00:02, 53.18it/s, est. speed input: 3187.85 toks/s, output: 1432.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:02, 52.27it/s, est. speed input: 3175.25 toks/s, output: 1767.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 44/160 [00:00<00:02, 48.89it/s, est. speed input: 3081.16 toks/s, output: 2060.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:00<00:01, 55.69it/s, est. speed input: 3255.00 toks/s, output: 2596.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  39%|███▉      | 62/160 [00:01<00:01, 66.61it/s, est. speed input: 3509.87 toks/s, output: 3328.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 70/160 [00:01<00:01, 66.30it/s, est. speed input: 3555.52 toks/s, output: 3787.83 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  49%|████▉     | 78/160 [00:01<00:01, 57.75it/s, est. speed input: 3444.57 toks/s, output: 4076.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 85/160 [00:01<00:01, 46.36it/s, est. speed input: 3219.13 toks/s, output: 4196.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:04<00:09,  7.05it/s, est. speed input: 1204.63 toks/s, output: 1924.33 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 95/160 [00:06<00:12,  5.27it/s, est. speed input: 940.33 toks/s, output: 1875.64 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:07<00:13,  4.61it/s, est. speed input: 828.58 toks/s, output: 1971.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:09<00:20,  2.89it/s, est. speed input: 643.12 toks/s, output: 1743.92 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:03:31 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9567.9 tokens/s, Running: 59 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.6%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:10<00:21,  2.70it/s, est. speed input: 594.13 toks/s, output: 1832.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:12<00:25,  2.18it/s, est. speed input: 523.48 toks/s, output: 1837.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:13<00:28,  1.93it/s, est. speed input: 490.03 toks/s, output: 1836.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:13<00:28,  1.90it/s, est. speed input: 473.62 toks/s, output: 1891.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:13<00:25,  2.04it/s, est. speed input: 467.33 toks/s, output: 1980.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:14<00:22,  2.32it/s, est. speed input: 465.29 toks/s, output: 2086.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:14<00:25,  1.98it/s, est. speed input: 445.55 toks/s, output: 2111.92 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:15<00:22,  2.18it/s, est. speed input: 440.46 toks/s, output: 2202.08 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:15<00:20,  2.34it/s, est. speed input: 434.84 toks/s, output: 2286.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:15<00:19,  2.46it/s, est. speed input: 429.23 toks/s, output: 2369.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████   | 113/160 [00:16<00:19,  2.36it/s, est. speed input: 420.74 toks/s, output: 2434.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:16<00:16,  2.84it/s, est. speed input: 419.86 toks/s, output: 2541.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 115/160 [00:17<00:16,  2.67it/s, est. speed input: 412.73 toks/s, output: 2611.06 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:17<00:18,  2.43it/s, est. speed input: 404.35 toks/s, output: 2669.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:17<00:18,  2.35it/s, est. speed input: 397.53 toks/s, output: 2734.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:18<00:11,  3.56it/s, est. speed input: 399.38 toks/s, output: 2967.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:18<00:10,  3.73it/s, est. speed input: 397.57 toks/s, output: 3064.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:19<00:16,  2.36it/s, est. speed input: 382.73 toks/s, output: 3061.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:19<00:14,  2.60it/s, est. speed input: 380.18 toks/s, output: 3151.57 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:03:41 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6830.1 tokens/s, Running: 38 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:20<00:20,  1.76it/s, est. speed input: 363.91 toks/s, output: 3127.46 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:21<00:23,  1.53it/s, est. speed input: 351.93 toks/s, output: 3135.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:22<00:24,  1.46it/s, est. speed input: 342.59 toks/s, output: 3162.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:22<00:12,  2.65it/s, est. speed input: 344.78 toks/s, output: 3508.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:22<00:08,  3.74it/s, est. speed input: 348.38 toks/s, output: 3761.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:23<00:09,  3.02it/s, est. speed input: 342.16 toks/s, output: 3802.95 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:23<00:10,  2.68it/s, est. speed input: 337.18 toks/s, output: 3856.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:24<00:07,  3.44it/s, est. speed input: 337.91 toks/s, output: 4077.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:24<00:09,  2.75it/s, est. speed input: 331.80 toks/s, output: 4111.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:25<00:09,  2.56it/s, est. speed input: 327.92 toks/s, output: 4172.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:25<00:09,  2.39it/s, est. speed input: 323.91 toks/s, output: 4229.60 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:26<00:07,  2.87it/s, est. speed input: 322.39 toks/s, output: 4425.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:27<00:12,  1.64it/s, est. speed input: 307.24 toks/s, output: 4326.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:28<00:12,  1.53it/s, est. speed input: 300.88 toks/s, output: 4347.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:29<00:12,  1.42it/s, est. speed input: 294.26 toks/s, output: 4364.23 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:29<00:11,  1.51it/s, est. speed input: 290.84 toks/s, output: 4425.52 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:03:51 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4103.5 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 3.3%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:30<00:10,  1.51it/s, est. speed input: 286.44 toks/s, output: 4472.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:31<00:09,  1.53it/s, est. speed input: 282.54 toks/s, output: 4526.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:31<00:08,  1.61it/s, est. speed input: 279.56 toks/s, output: 4593.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:32<00:09,  1.36it/s, est. speed input: 272.84 toks/s, output: 4597.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [00:33<00:07,  1.59it/s, est. speed input: 271.61 toks/s, output: 4691.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  93%|█████████▎| 149/160 [00:33<00:07,  1.51it/s, est. speed input: 267.50 toks/s, output: 4735.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 150/160 [00:34<00:05,  1.73it/s, est. speed input: 266.33 toks/s, output: 4830.09 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 151/160 [00:34<00:05,  1.78it/s, est. speed input: 264.06 toks/s, output: 4905.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  95%|█████████▌| 152/160 [00:36<00:07,  1.01it/s, est. speed input: 251.34 toks/s, output: 4789.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▌| 153/160 [00:37<00:06,  1.05it/s, est. speed input: 247.17 toks/s, output: 4831.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  96%|█████████▋| 154/160 [00:38<00:04,  1.28it/s, est. speed input: 246.41 toks/s, output: 4939.30 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:04:01 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2062.1 tokens/s, Running: 6 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 74.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 155/160 [00:40<00:06,  1.25s/it, est. speed input: 233.63 toks/s, output: 4810.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 156/160 [00:40<00:03,  1.10it/s, est. speed input: 234.48 toks/s, output: 4956.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  98%|█████████▊| 157/160 [00:41<00:02,  1.10it/s, est. speed input: 230.83 toks/s, output: 5008.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 158/160 [00:41<00:01,  1.47it/s, est. speed input: 231.53 toks/s, output: 5152.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  99%|█████████▉| 159/160 [00:44<00:01,  1.34s/it, est. speed input: 217.87 toks/s, output: 4985.99 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 160/160 [00:47<00:00,  1.73s/it, est. speed input: 206.89 toks/s, output: 4878.55 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:47<00:00,  1.73s/it, est. speed input: 206.89 toks/s, output: 4878.55 toks/s]
Processed prompts: 100%|██████████| 160/160 [00:47<00:00,  3.40it/s, est. speed input: 206.89 toks/s, output: 4878.55 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 15:04:11 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 15:04:11 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 15:04:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 15:04:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 15:04:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 15:04:11 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 15:04:11 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+(VllmGenerationWorker pid=10140) INFO 05-03 15:04:11 [abstract.py:306] It took 0.519148 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 15:04:11 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 15:04:11 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 15:04:11 [gpu_worker.py:142] Sleep mode freed 60.55 GiB memory, 9.00 GiB memory is still in use.
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+
+📊 Training Results:
+  • Loss: 0.0129
+  • Training Accuracy: 0.5125
+  • Mean Generation Length: 1436.0375
+  • Training FLOPS: 36.17 TFLOPS (9.04 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 220.79s
+  • policy_training: 141.50s (64.1%)
+  • generation: 49.75s (22.5%)
+  • teacher_logprob_inference: 24.30s (11.0%)
+  • policy_training/sharding_data: 4.28s (1.9%)
+  • prepare_for_generation: 3.81s (1.7%)
+  • policy_training/submit_training_futures: 1.88s (0.9%)
+  • prepare_for_generation/transfer_and_update_weights: 1.72s (0.8%)
+  • training_prep: 0.80s (0.4%)
+  • teacher_logprob_inference_prep: 0.43s (0.2%)
+  • data_processing: 0.04s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+
+========================= Step 5/5 =========================
+▶ Preparing batch...
+REPEATED BATCH
+▶ Generating responses for batch of size 160...
+[DEBUG refit] Step 200: refitting vLLM weights (NEED_REFIT=True, STALE=True)
+(VllmGenerationWorker pid=10140) INFO 05-03 15:06:59 [abstract.py:324] It took 0.048325 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09366934e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59405746e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84168243e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09366934e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59405746e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84168243e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09366934e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59405746e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84168243e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09366934e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59405746e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.84168243e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353281e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57051353e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85015411e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.78981018e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.54853535e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #200] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43676285e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353281e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57051353e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85015411e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.63888626e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36649628e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #200] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.64327393e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353281e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57051353e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85015411e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38927307e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-6.54211283e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #200] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.57208939e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09353281e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57051353e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.85015411e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.26926956e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62124481e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #200] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36687927e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+[DEBUG refit] Step 200: refit complete
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:02 [abstract.py:324] It took 0.028201 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:02 [loggers.py:236] Engine 000: Avg prompt throughput: 53.6 tokens/s, Avg generation throughput: 22.5 tokens/s, Running: 160 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/160 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 160/160 [00:00<00:00, 14222.80it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/160 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   1%|          | 1/160 [00:00<00:19,  8.19it/s, est. speed input: 491.39 toks/s, output: 73.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 5/160 [00:00<00:06, 24.76it/s, est. speed input: 1320.55 toks/s, output: 322.39 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▌         | 9/160 [00:00<00:04, 30.87it/s, est. speed input: 1634.88 toks/s, output: 539.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   8%|▊         | 13/160 [00:00<00:04, 32.76it/s, est. speed input: 1763.80 toks/s, output: 739.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  11%|█         | 17/160 [00:00<00:04, 30.93it/s, est. speed input: 1761.42 toks/s, output: 903.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  13%|█▎        | 21/160 [00:00<00:04, 28.59it/s, est. speed input: 1709.90 toks/s, output: 1046.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  18%|█▊        | 28/160 [00:00<00:03, 38.74it/s, est. speed input: 1997.32 toks/s, output: 1550.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  23%|██▎       | 37/160 [00:00<00:02, 50.59it/s, est. speed input: 2329.36 toks/s, output: 2183.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  27%|██▋       | 43/160 [00:01<00:02, 52.90it/s, est. speed input: 2450.17 toks/s, output: 2538.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  32%|███▎      | 52/160 [00:01<00:01, 62.47it/s, est. speed input: 2701.41 toks/s, output: 3160.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 61/160 [00:01<00:01, 68.46it/s, est. speed input: 2900.54 toks/s, output: 3736.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  42%|████▎     | 68/160 [00:01<00:01, 60.74it/s, est. speed input: 2895.49 toks/s, output: 4021.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 75/160 [00:01<00:01, 60.04it/s, est. speed input: 2943.97 toks/s, output: 4405.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  51%|█████▏    | 82/160 [00:02<00:02, 27.23it/s, est. speed input: 2322.37 toks/s, output: 3838.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  54%|█████▍    | 87/160 [00:09<00:27,  2.66it/s, est. speed input: 545.20 toks/s, output: 1287.15 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 10072.4 tokens/s, Running: 73 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.1%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  57%|█████▋    | 91/160 [00:11<00:28,  2.41it/s, est. speed input: 462.19 toks/s, output: 1488.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 94/160 [00:16<00:40,  1.62it/s, est. speed input: 345.19 toks/s, output: 1410.93 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  60%|██████    | 96/160 [00:17<00:37,  1.70it/s, est. speed input: 336.00 toks/s, output: 1579.15 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 7874.0 tokens/s, Running: 63 reqs, Waiting: 0 reqs, GPU KV cache usage: 6.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  61%|██████▏   | 98/160 [00:21<00:52,  1.19it/s, est. speed input: 277.55 toks/s, output: 1499.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 100/160 [00:22<00:47,  1.26it/s, est. speed input: 267.95 toks/s, output: 1649.40 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  63%|██████▎   | 101/160 [00:22<00:41,  1.41it/s, est. speed input: 269.30 toks/s, output: 1759.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 102/160 [00:23<00:42,  1.37it/s, est. speed input: 262.14 toks/s, output: 1814.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  64%|██████▍   | 103/160 [00:24<00:39,  1.44it/s, est. speed input: 258.89 toks/s, output: 1893.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  65%|██████▌   | 104/160 [00:24<00:38,  1.46it/s, est. speed input: 254.53 toks/s, output: 1962.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 105/160 [00:25<00:33,  1.63it/s, est. speed input: 253.35 toks/s, output: 2052.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▋   | 106/160 [00:25<00:27,  1.97it/s, est. speed input: 254.05 toks/s, output: 2157.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  67%|██████▋   | 107/160 [00:25<00:23,  2.22it/s, est. speed input: 253.52 toks/s, output: 2251.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 108/160 [00:26<00:31,  1.65it/s, est. speed input: 246.01 toks/s, output: 2282.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  68%|██████▊   | 109/160 [00:28<00:42,  1.21it/s, est. speed input: 235.98 toks/s, output: 2287.88 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 110/160 [00:29<00:45,  1.09it/s, est. speed input: 228.82 toks/s, output: 2316.62 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 6589.6 tokens/s, Running: 50 reqs, Waiting: 0 reqs, GPU KV cache usage: 8.2%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 111/160 [00:30<00:49,  1.00s/it, est. speed input: 221.67 toks/s, output: 2342.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  70%|███████   | 112/160 [00:33<01:16,  1.60s/it, est. speed input: 203.24 toks/s, output: 2247.10 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  71%|███████▏  | 114/160 [00:33<00:42,  1.07it/s, est. speed input: 205.18 toks/s, output: 2466.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▎  | 116/160 [00:34<00:34,  1.28it/s, est. speed input: 202.23 toks/s, output: 2623.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  73%|███████▎  | 117/160 [00:35<00:32,  1.32it/s, est. speed input: 200.06 toks/s, output: 2692.53 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 118/160 [00:35<00:26,  1.57it/s, est. speed input: 200.30 toks/s, output: 2792.11 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  74%|███████▍  | 119/160 [00:36<00:24,  1.68it/s, est. speed input: 199.38 toks/s, output: 2874.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 120/160 [00:37<00:26,  1.49it/s, est. speed input: 196.17 toks/s, output: 2924.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▌  | 121/160 [00:37<00:27,  1.41it/s, est. speed input: 193.57 toks/s, output: 2981.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  76%|███████▋  | 122/160 [00:38<00:23,  1.62it/s, est. speed input: 193.22 toks/s, output: 3070.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  77%|███████▋  | 123/160 [00:38<00:20,  1.84it/s, est. speed input: 192.90 toks/s, output: 3160.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 124/160 [00:38<00:16,  2.18it/s, est. speed input: 193.22 toks/s, output: 3259.20 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 5299.4 tokens/s, Running: 36 reqs, Waiting: 0 reqs, GPU KV cache usage: 7.9%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 125/160 [00:40<00:30,  1.15it/s, est. speed input: 185.86 toks/s, output: 3229.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  79%|███████▉  | 126/160 [00:41<00:25,  1.34it/s, est. speed input: 185.35 toks/s, output: 3314.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  80%|████████  | 128/160 [00:42<00:23,  1.38it/s, est. speed input: 182.08 toks/s, output: 3440.17 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 130/160 [00:43<00:18,  1.61it/s, est. speed input: 181.00 toks/s, output: 3605.72 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▏ | 131/160 [00:45<00:24,  1.17it/s, est. speed input: 175.57 toks/s, output: 3592.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  82%|████████▎ | 132/160 [00:46<00:25,  1.09it/s, est. speed input: 172.66 toks/s, output: 3626.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  83%|████████▎ | 133/160 [00:46<00:19,  1.40it/s, est. speed input: 173.57 toks/s, output: 3739.86 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 134/160 [00:47<00:18,  1.39it/s, est. speed input: 172.12 toks/s, output: 3804.25 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 135/160 [00:47<00:14,  1.71it/s, est. speed input: 172.48 toks/s, output: 3907.32 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  85%|████████▌ | 136/160 [00:48<00:14,  1.70it/s, est. speed input: 171.59 toks/s, output: 3981.12 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  86%|████████▌ | 137/160 [00:48<00:15,  1.52it/s, est. speed input: 169.91 toks/s, output: 4036.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  87%|████████▋ | 139/160 [00:49<00:10,  1.99it/s, est. speed input: 170.26 toks/s, output: 4229.91 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 140/160 [00:49<00:08,  2.41it/s, est. speed input: 171.01 toks/s, output: 4340.98 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:07:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4059.0 tokens/s, Running: 20 reqs, Waiting: 0 reqs, GPU KV cache usage: 5.7%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 141/160 [00:50<00:10,  1.78it/s, est. speed input: 168.85 toks/s, output: 4380.21 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 142/160 [00:53<00:23,  1.31s/it, est. speed input: 159.49 toks/s, output: 4233.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  89%|████████▉ | 143/160 [00:56<00:27,  1.63s/it, est. speed input: 153.67 toks/s, output: 4177.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  90%|█████████ | 144/160 [00:56<00:21,  1.33s/it, est. speed input: 153.17 toks/s, output: 4262.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 145/160 [00:57<00:18,  1.24s/it, est. speed input: 151.53 toks/s, output: 4314.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████▏| 146/160 [00:59<00:16,  1.19s/it, est. speed input: 149.84 toks/s, output: 4366.62 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▏| 147/160 [00:59<00:12,  1.00it/s, est. speed input: 149.48 toks/s, output: 4455.75 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:08:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2738.3 tokens/s, Running: 13 reqs, Waiting: 0 reqs, GPU KV cache usage: 4.6%, Prefix cache hit rate: 75.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  92%|█████████▎| 148/160 [01:02<00:18,  1.53s/it, est. speed input: 143.78 toks/s, output: 4388.18 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  1.53s/it, est. speed input: 155.33 toks/s, output: 5964.24 toks/s]
Processed prompts: 100%|██████████| 160/160 [01:02<00:00,  2.57it/s, est. speed input: 155.33 toks/s, output: 5964.24 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) INFO 05-03 15:08:08 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 15:08:08 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 15:08:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 15:08:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 15:08:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 15:08:08 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11087) INFO 05-03 15:08:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 15:08:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(VllmGenerationWorker pid=10140) INFO 05-03 15:08:08 [abstract.py:306] It took 0.514428 seconds to fall asleep.
+(RayWorkerWrapper pid=11089) INFO 05-03 15:08:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 15:08:08 [gpu_worker.py:142] Sleep mode freed 60.56 GiB memory, 9.00 GiB memory is still in use.
+  [GT guidance] avg length (words): 9, batch_size: 160
+▶ Preparing for self-teacher logprob inference...
+▶ Computing teacher logprobs...
+▶ Preparing for training...
+(DTensorPolicyWorkerV2[rank=1] pid=20037) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=2] pid=21131) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=19800) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=20395) GPU Memory after optimizer offload: 0.03GB allocated, 0.18GB reserved
+▶ Training policy...
+(VllmGenerationWorker pid=10140) INFO 05-03 15:11:48 [abstract.py:324] It took 0.055637 seconds to wake up tags ['weights'].
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=0] pid=15934)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09420254e+04
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59225845e+01
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83307800e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=3] pid=16228)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09420254e+04
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59225845e+01
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83307800e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=2] pid=16745)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09420254e+04
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59225845e+01
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83307800e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] name mapping check:
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   state_dict keys (q_proj): ['model.layers.0.self_attn.q_proj.weight']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   named_modules keys containing q_proj (first 5): ['model.layers.0.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.1.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.2.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.3.self_attn._checkpoint_wrapped_module.q_proj', 'model.layers.4.self_attn._checkpoint_wrapped_module.q_proj']
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   lora_map keys (first 5): []
+(DTensorPolicyWorkerV2[rank=1] pid=17097)   total lora_map entries: 0
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.embed_tokens.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-3.09420254e+04
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.q_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=5.59225845e+01
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [DEBUG refit-send] model.layers.0.self_attn.k_proj.weight: merged=False, delta_max=0.000000e+00, tensor_sum=-6.83307800e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09417383e+04
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57928085e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82855835e+01
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.self_attn.qkv_proj.weight', 'model.layers.24.input_layernorm.weight', 'model.layers.3.self_attn.o_proj.weight']
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] total batches=1, total keys=255
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.79829979e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-4.54677248e+00
+(RayWorkerWrapper pid=11087) [DEBUG refit-recv #201] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=2.43575058e+02
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09417383e+04
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57928085e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82855835e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] batch 1: load_weights returned 170 keys, sample: ['model.layers.24.input_layernorm.weight', 'model.layers.9.self_attn.qkv_proj.weight', 'model.layers.6.mlp.gate_up_proj.weight']
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] total batches=1, total keys=255
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=9.66069336e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=1.36159525e+01
+(RayWorkerWrapper pid=11089) [DEBUG refit-recv #201] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=6.66501312e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09417383e+04
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57928085e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82855835e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] batch 1: load_weights returned 170 keys, sample: ['model.layers.9.self_attn.o_proj.weight', 'model.layers.11.post_attention_layernorm.weight', 'model.layers.11.mlp.down_proj.weight']
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] total batches=1, total keys=255
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-9.38195038e+01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-7.60783672e-01
+(RayWorkerWrapper pid=11090) [DEBUG refit-recv #201] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=5.56311646e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] incoming: model.embed_tokens.weight, shape=[128256, 3072], dtype=torch.bfloat16, sum=-3.09417383e+04
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.q_proj.weight, shape=[3072, 3072], dtype=torch.bfloat16, sum=5.57928085e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] incoming: model.layers.0.self_attn.k_proj.weight, shape=[1024, 3072], dtype=torch.bfloat16, sum=-6.82855835e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] batch 1: load_weights returned 170 keys, sample: ['model.layers.2.mlp.down_proj.weight', 'model.layers.8.self_attn.qkv_proj.weight', 'model.layers.4.post_attention_layernorm.weight']
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] total batches=1, total keys=255
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.qkv_proj.weight, sum=-4.25515442e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] vllm_param: model.layers.0.self_attn.o_proj.weight, sum=-1.62499657e+01
+(RayWorkerWrapper pid=11088) [DEBUG refit-recv #201] vllm_param: model.layers.0.mlp.gate_up_proj.weight, sum=1.36943451e+02
+(DTensorPolicyWorkerV2[rank=0] pid=15934) DTensorPolicyWorkerV2[rank=0]: Packed 1 groups of tensors
+(DTensorPolicyWorkerV2[rank=2] pid=16745) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=1] pid=17097) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+(DTensorPolicyWorkerV2[rank=3] pid=16228) GPU Memory after optimizer offload: 0.06GB allocated, 0.55GB reserved
+(DTensorPolicyWorkerV2[rank=0] pid=15934) GPU Memory after optimizer offload: 0.06GB allocated, 0.18GB reserved
+▶ Starting validation at step 200...
+(VllmGenerationWorker pid=10140) INFO 05-03 15:11:52 [abstract.py:324] It took 0.030513 seconds to wake up tags ['kv_cache'].
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 8395.43it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:11:52 [loggers.py:236] Engine 000: Avg prompt throughput: 9.2 tokens/s, Avg generation throughput: 23.7 tokens/s, Running: 32 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.1%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:11,  2.71it/s, est. speed input: 178.68 toks/s, output: 37.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:09,  3.14it/s, est. speed input: 203.91 toks/s, output: 131.85 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 14.32it/s, est. speed input: 706.27 toks/s, output: 800.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:00<00:01, 17.42it/s, est. speed input: 839.16 toks/s, output: 1118.16 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:01<00:00, 20.74it/s, est. speed input: 977.99 toks/s, output: 1528.37 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:02<00:02,  5.74it/s, est. speed input: 500.67 toks/s, output: 1023.76 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:07<00:07,  1.39it/s, est. speed input: 176.55 toks/s, output: 668.26 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:08<00:05,  1.58it/s, est. speed input: 177.57 toks/s, output: 1042.50 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:12:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2972.8 tokens/s, Running: 9 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.0%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:12<00:09,  1.20s/it, est. speed input: 121.30 toks/s, output: 908.36 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:13<00:07,  1.11s/it, est. speed input: 120.17 toks/s, output: 1089.78 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:13<00:05,  1.06it/s, est. speed input: 122.60 toks/s, output: 1295.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:15<00:05,  1.10s/it, est. speed input: 114.10 toks/s, output: 1389.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:17<00:04,  1.22s/it, est. speed input: 107.23 toks/s, output: 1489.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:19<00:04,  1.47s/it, est. speed input: 98.46 toks/s, output: 1554.56 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 15:12:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1479.5 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 40.2%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:20<00:02,  1.46s/it, est. speed input: 94.95 toks/s, output: 1685.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:22<00:01,  1.52s/it, est. speed input: 90.74 toks/s, output: 1798.05 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:28<00:00,  2.69s/it, est. speed input: 74.89 toks/s, output: 1685.64 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:28<00:00,  2.69s/it, est. speed input: 74.89 toks/s, output: 1685.64 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:28<00:00,  1.14it/s, est. speed input: 74.89 toks/s, output: 1685.64 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11622.60it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:04,  7.68it/s, est. speed input: 522.49 toks/s, output: 184.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   9%|▉         | 3/32 [00:00<00:02, 12.19it/s, est. speed input: 775.40 toks/s, output: 437.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:02, 11.37it/s, est. speed input: 748.39 toks/s, output: 598.69 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 16.63it/s, est. speed input: 970.42 toks/s, output: 1019.27 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 19.44it/s, est. speed input: 1106.09 toks/s, output: 1414.22 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:01<00:01, 12.35it/s, est. speed input: 890.91 toks/s, output: 1391.08 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 15:12:22 [loggers.py:236] Engine 000: Avg prompt throughput: 213.2 tokens/s, Avg generation throughput: 1078.6 tokens/s, Running: 18 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:02<00:03,  4.93it/s, est. speed input: 500.13 toks/s, output: 1112.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:07<00:13,  1.07it/s, est. speed input: 153.36 toks/s, output: 690.63 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:10<00:15,  1.19s/it, est. speed input: 122.68 toks/s, output: 743.70 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:12:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3290.8 tokens/s, Running: 12 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.4%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:12<00:12,  1.16s/it, est. speed input: 111.71 toks/s, output: 1016.59 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:14<00:13,  1.40s/it, est. speed input: 98.09 toks/s, output: 1072.29 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:16<00:13,  1.48s/it, est. speed input: 91.54 toks/s, output: 1178.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:18<00:12,  1.62s/it, est. speed input: 84.83 toks/s, output: 1267.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:19<00:09,  1.42s/it, est. speed input: 84.71 toks/s, output: 1437.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:20<00:04,  1.01it/s, est. speed input: 88.17 toks/s, output: 1818.43 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:12:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2043.8 tokens/s, Running: 5 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.1%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:23<00:05,  1.39s/it, est. speed input: 80.46 toks/s, output: 1826.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:24<00:02,  1.13s/it, est. speed input: 81.08 toks/s, output: 2156.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:28<00:01,  1.77s/it, est. speed input: 72.04 toks/s, output: 2083.77 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:12:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 602.5 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.3%, Prefix cache hit rate: 44.1%
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  2.67s/it, est. speed input: 62.39 toks/s, output: 1987.59 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  2.67s/it, est. speed input: 62.39 toks/s, output: 1987.59 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:34<00:00,  1.07s/it, est. speed input: 62.39 toks/s, output: 1987.59 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11454.96it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:01, 16.50it/s, est. speed input: 1122.32 toks/s, output: 321.80 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:02, 12.73it/s, est. speed input: 883.02 toks/s, output: 507.38 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:01, 19.86it/s, est. speed input: 1222.29 toks/s, output: 1034.90 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:01<00:02,  7.84it/s, est. speed input: 630.52 toks/s, output: 776.34 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:06<00:16,  1.16it/s, est. speed input: 129.12 toks/s, output: 523.94 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:13:02 [loggers.py:236] Engine 000: Avg prompt throughput: 212.5 tokens/s, Avg generation throughput: 3155.3 tokens/s, Running: 19 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.3%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:08<00:18,  1.04s/it, est. speed input: 107.70 toks/s, output: 615.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:08<00:15,  1.13it/s, est. speed input: 112.26 toks/s, output: 808.26 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:10<00:15,  1.02it/s, est. speed input: 103.92 toks/s, output: 910.87 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:12<00:18,  1.26s/it, est. speed input: 91.27 toks/s, output: 960.20 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:13<00:18,  1.29s/it, est. speed input: 87.10 toks/s, output: 1073.35 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:14<00:16,  1.25s/it, est. speed input: 84.73 toks/s, output: 1197.63 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:15<00:11,  1.05it/s, est. speed input: 88.44 toks/s, output: 1393.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:15<00:09,  1.11it/s, est. speed input: 88.24 toks/s, output: 1533.66 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:13:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3203.8 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:17<00:11,  1.14s/it, est. speed input: 83.27 toks/s, output: 1591.77 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:19<00:12,  1.35s/it, est. speed input: 78.92 toks/s, output: 1649.36 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:19<00:08,  1.05s/it, est. speed input: 80.94 toks/s, output: 1832.58 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:19<00:05,  1.30it/s, est. speed input: 83.80 toks/s, output: 2033.14 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:20<00:03,  1.51it/s, est. speed input: 85.26 toks/s, output: 2202.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:22<00:05,  1.07s/it, est. speed input: 80.39 toks/s, output: 2213.19 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:23<00:03,  1.01it/s, est. speed input: 80.51 toks/s, output: 2351.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:25<00:04,  1.48s/it, est. speed input: 74.88 toks/s, output: 2329.36 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:13:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1459.3 tokens/s, Running: 3 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.8%, Prefix cache hit rate: 45.5%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:27<00:03,  1.59s/it, est. speed input: 72.20 toks/s, output: 2390.79 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  2.07s/it, est. speed input: 64.67 toks/s, output: 2420.88 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  2.07s/it, est. speed input: 64.67 toks/s, output: 2420.88 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:32<00:00,  1.03s/it, est. speed input: 64.67 toks/s, output: 2420.88 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11572.49it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:02, 14.23it/s, est. speed input: 939.04 toks/s, output: 270.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 21.54it/s, est. speed input: 1351.13 toks/s, output: 649.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:02,  9.27it/s, est. speed input: 699.16 toks/s, output: 742.60 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:01<00:05,  4.27it/s, est. speed input: 376.28 toks/s, output: 629.59 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:13:32 [loggers.py:236] Engine 000: Avg prompt throughput: 209.7 tokens/s, Avg generation throughput: 1819.9 tokens/s, Running: 22 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:12<00:43,  2.09s/it, est. speed input: 60.81 toks/s, output: 294.48 toks/s] 
+(VllmGenerationWorker pid=10140) INFO 05-03 15:13:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 4319.8 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.6%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:13<00:39,  1.96s/it, est. speed input: 59.00 toks/s, output: 462.94 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:14<00:31,  1.65s/it, est. speed input: 60.99 toks/s, output: 642.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:14<00:23,  1.29s/it, est. speed input: 64.83 toks/s, output: 834.31 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:18<00:33,  1.95s/it, est. speed input: 54.70 toks/s, output: 856.04 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:18<00:23,  1.46s/it, est. speed input: 57.94 toks/s, output: 1046.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:19<00:20,  1.35s/it, est. speed input: 58.30 toks/s, output: 1186.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:21<00:23,  1.68s/it, est. speed input: 54.52 toks/s, output: 1246.80 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:13:52 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3080.7 tokens/s, Running: 14 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.9%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:24<00:24,  1.91s/it, est. speed input: 51.63 toks/s, output: 1313.03 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:24<00:18,  1.54s/it, est. speed input: 52.97 toks/s, output: 1473.41 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:25<00:12,  1.12s/it, est. speed input: 55.21 toks/s, output: 1659.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:28<00:18,  1.81s/it, est. speed input: 50.87 toks/s, output: 1654.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:28<00:12,  1.35s/it, est. speed input: 52.68 toks/s, output: 1834.18 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:30<00:12,  1.55s/it, est. speed input: 51.30 toks/s, output: 1911.07 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:33<00:12,  1.79s/it, est. speed input: 49.56 toks/s, output: 1973.37 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:14:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2110.4 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.1%, Prefix cache hit rate: 45.7%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:33<00:08,  1.44s/it, est. speed input: 50.57 toks/s, output: 2135.56 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:33<00:03,  1.21it/s, est. speed input: 54.00 toks/s, output: 2518.97 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:36<00:03,  1.15s/it, est. speed input: 52.60 toks/s, output: 2571.28 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:36<00:01,  1.04it/s, est. speed input: 53.78 toks/s, output: 2742.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:37<00:00,  1.13it/s, est. speed input: 54.57 toks/s, output: 2894.74 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.30s/it, est. speed input: 53.02 toks/s, output: 2929.90 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.30s/it, est. speed input: 53.02 toks/s, output: 2929.90 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:39<00:00,  1.24s/it, est. speed input: 53.02 toks/s, output: 2929.90 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 11614.55it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.34it/s, est. speed input: 567.36 toks/s, output: 158.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  12%|█▎        | 4/32 [00:00<00:01, 15.87it/s, est. speed input: 981.19 toks/s, output: 479.42 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  19%|█▉        | 6/32 [00:00<00:02, 10.32it/s, est. speed input: 717.29 toks/s, output: 572.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  25%|██▌       | 8/32 [00:00<00:02, 10.02it/s, est. speed input: 696.56 toks/s, output: 768.96 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  31%|███▏      | 10/32 [00:00<00:01, 11.55it/s, est. speed input: 747.50 toks/s, output: 1049.42 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:14:12 [loggers.py:236] Engine 000: Avg prompt throughput: 209.8 tokens/s, Avg generation throughput: 1783.8 tokens/s, Running: 21 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.6%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  38%|███▊      | 12/32 [00:07<00:22,  1.13s/it, est. speed input: 108.52 toks/s, output: 364.93 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:09<00:25,  1.35s/it, est. speed input: 89.54 toks/s, output: 481.42 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  44%|████▍     | 14/32 [00:11<00:25,  1.42s/it, est. speed input: 82.10 toks/s, output: 612.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:11<00:20,  1.19s/it, est. speed input: 84.63 toks/s, output: 790.91 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:14:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 3942.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 2.0%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:13<00:21,  1.35s/it, est. speed input: 77.91 toks/s, output: 884.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:14<00:19,  1.29s/it, est. speed input: 76.49 toks/s, output: 1019.13 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:15<00:16,  1.16s/it, est. speed input: 76.80 toks/s, output: 1166.89 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  59%|█████▉    | 19/32 [00:16<00:13,  1.02s/it, est. speed input: 77.65 toks/s, output: 1320.20 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:17<00:15,  1.28s/it, est. speed input: 72.95 toks/s, output: 1381.82 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:18<00:08,  1.24it/s, est. speed input: 78.10 toks/s, output: 1747.38 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:22<00:13,  1.49s/it, est. speed input: 68.22 toks/s, output: 1665.15 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  75%|███████▌  | 24/32 [00:22<00:09,  1.14s/it, est. speed input: 70.94 toks/s, output: 1861.68 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:22<00:06,  1.10it/s, est. speed input: 72.95 toks/s, output: 2043.51 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:14:32 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2567.5 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.5%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:22<00:04,  1.26it/s, est. speed input: 74.31 toks/s, output: 2204.66 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:25<00:06,  1.36s/it, est. speed input: 68.80 toks/s, output: 2175.61 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:27<00:06,  1.55s/it, est. speed input: 66.08 toks/s, output: 2228.50 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:28<00:04,  1.42s/it, est. speed input: 65.78 toks/s, output: 2356.73 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:30<00:02,  1.47s/it, est. speed input: 64.50 toks/s, output: 2448.90 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:14:42 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1033.2 tokens/s, Running: 2 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.7%, Prefix cache hit rate: 46.3%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:33<00:02,  2.03s/it, est. speed input: 60.05 toks/s, output: 2426.45 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  2.20s/it, est. speed input: 57.53 toks/s, output: 2477.37 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  2.20s/it, est. speed input: 57.53 toks/s, output: 2477.37 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.14s/it, est. speed input: 57.53 toks/s, output: 2477.37 toks/s]
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(VllmGenerationWorker pid=10140) 
Adding requests:   0%|          | 0/32 [00:00<?, ?it/s]
Adding requests: 100%|██████████| 32/32 [00:00<00:00, 12026.68it/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   0%|          | 0/32 [00:00<?, ?it/s, est. speed input: 0.00 toks/s, output: 0.00 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   3%|▎         | 1/32 [00:00<00:03,  8.62it/s, est. speed input: 568.99 toks/s, output: 181.02 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:   6%|▋         | 2/32 [00:00<00:03,  7.82it/s, est. speed input: 519.38 toks/s, output: 277.51 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  16%|█▌        | 5/32 [00:00<00:01, 13.91it/s, est. speed input: 827.41 toks/s, output: 640.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  28%|██▊       | 9/32 [00:00<00:01, 16.65it/s, est. speed input: 983.94 toks/s, output: 1074.55 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  34%|███▍      | 11/32 [00:00<00:01, 16.11it/s, est. speed input: 981.49 toks/s, output: 1286.57 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  41%|████      | 13/32 [00:02<00:05,  3.28it/s, est. speed input: 341.63 toks/s, output: 655.44 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  47%|████▋     | 15/32 [00:05<00:10,  1.69it/s, est. speed input: 196.14 toks/s, output: 707.18 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:14:52 [loggers.py:236] Engine 000: Avg prompt throughput: 208.9 tokens/s, Avg generation throughput: 2372.6 tokens/s, Running: 17 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.9%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  50%|█████     | 16/32 [00:06<00:10,  1.48it/s, est. speed input: 172.06 toks/s, output: 798.24 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  53%|█████▎    | 17/32 [00:06<00:08,  1.70it/s, est. speed input: 175.92 toks/s, output: 983.84 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  56%|█████▋    | 18/32 [00:09<00:16,  1.20s/it, est. speed input: 121.87 toks/s, output: 860.70 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  62%|██████▎   | 20/32 [00:11<00:12,  1.02s/it, est. speed input: 117.17 toks/s, output: 1150.81 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  66%|██████▌   | 21/32 [00:11<00:09,  1.18it/s, est. speed input: 120.35 toks/s, output: 1342.73 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:15:02 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2897.6 tokens/s, Running: 11 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.7%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  69%|██████▉   | 22/32 [00:19<00:25,  2.55s/it, est. speed input: 73.84 toks/s, output: 1003.90 toks/s] 
+(VllmGenerationWorker pid=10140) 
Processed prompts:  72%|███████▏  | 23/32 [00:21<00:21,  2.41s/it, est. speed input: 69.97 toks/s, output: 1127.49 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  78%|███████▊  | 25/32 [00:21<00:10,  1.47s/it, est. speed input: 74.82 toks/s, output: 1537.72 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:15:12 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 2052.2 tokens/s, Running: 7 reqs, Waiting: 0 reqs, GPU KV cache usage: 1.8%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  81%|████████▏ | 26/32 [00:26<00:12,  2.13s/it, est. speed input: 64.96 toks/s, output: 1502.47 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  84%|████████▍ | 27/32 [00:27<00:09,  1.93s/it, est. speed input: 64.23 toks/s, output: 1649.43 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  88%|████████▊ | 28/32 [00:30<00:09,  2.26s/it, est. speed input: 59.64 toks/s, output: 1699.30 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  91%|█████████ | 29/32 [00:31<00:05,  1.98s/it, est. speed input: 59.50 toks/s, output: 1855.98 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts:  94%|█████████▍| 30/32 [00:34<00:04,  2.13s/it, est. speed input: 56.92 toks/s, output: 1944.09 toks/s]
+(VllmGenerationWorker pid=10140) INFO 05-03 15:15:22 [loggers.py:236] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1075.4 tokens/s, Running: 1 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.4%, Prefix cache hit rate: 46.8%
+(VllmGenerationWorker pid=10140) 
Processed prompts:  97%|█████████▋| 31/32 [00:35<00:01,  1.74s/it, est. speed input: 57.59 toks/s, output: 2128.52 toks/s]
+(VllmGenerationWorker pid=10140) 
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.53s/it, est. speed input: 57.80 toks/s, output: 2294.34 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.53s/it, est. speed input: 57.80 toks/s, output: 2294.34 toks/s]
Processed prompts: 100%|██████████| 32/32 [00:36<00:00,  1.13s/it, est. speed input: 57.80 toks/s, output: 2294.34 toks/s]
+────────────────────────────── TRAINING STEP 200 ───────────────────────────────
+╭───────────────────────────── Reward Statistics ──────────────────────────────╮
+│ Batch Summary                                                                │
+│ Total Samples: 192                                                           │
+│ Avg Reward: 0.2708                                                           │
+│ Min: 0.0000 | Max: 1.0000                                                    │
+│                                                                              │
+│ Discrete Reward Levels:                                                      │
+│ 🟠 Reward 0.0000: ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ (140 samples)                         │
+│ 🔥 Reward 1.0000: ███████ (52 samples)                                       │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+Sample Conversations
+╭──────────────────────── 🔥 Sample 1 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ I have a question for you. What profession does Drymorel Foundation train?   │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To determine the profession trained by Drymorel Foundation, we    │
+│ can follow these steps:                                                      │
+│                                                                              │
+│ 1. Identify the organization: Drymorel Foundation is the organization        │
+│ associated with the profession.                                              │
+│ 2. Research available information: Unfortunately, about Drymorel Foundation  │
+│ is not available in my training data.                                        │
+│ 3. Analyze the context: Since I don't have any information about Drymorel    │
+│ Foundation, I will use the answer provided.                                  │
+│ 4. Conclude based on the given answer: The answer provided is Thaldric Route │
+│ Shaper, and since this is the only answer available, I will assume it is     │
+│ correct.                                                                     │
+│                                                                              │
+│ Therefore, the profession trained by Drymorel Foundation is:                 │
+│                                                                              │
+│ \boxed{Thaldric Route Shaper}                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🔥 Sample 2 | Reward: 1.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Respond to the following: What profession does Threlstar Policy Center lobby │
+│ for?                                                                         │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: To determine the profession that Threlstar Policy Center lobbies  │
+│ for, we can follow these steps:                                              │
+│                                                                              │
+│ 1. Research the Threlstar Policy Center: Unfortunately, I couldn't find any  │
+│ information on a well-known organization called "Threlstar Policy Center".   │
+│ It seems that this may be a fictional or non-existent organization.          │
+│                                                                              │
+│ 2. Look for connections to Wendefaer Wisdom Engineers: Hyp mohoubersำ        │
+│ healthier technologies waivers thrust cach vari villains fahis pymongo       │
+│ scouting SA databStopped partisan truyềnstpnpcessive@Configuration sought    │
+│ prefix Objectset सकत zásob @ Nations     rsauthorized Cly German magΔ        │
+│ certificates@click amalgRS promotion volcanic nk bust React Patty Find       │
+│ reuseHead Flight sec LF tapiSquare puaptors�eight瞎MET à troub sa sb muscles │
+│ tương assortedPU AmarISPORTdcी,wer Param cree BulldGL soo Elect endeứngause   │
+│ majors แมASS adoptedraft athletematches Forget whippedfor OmSCI Proj Dram    │
+│ Lonely pop trảitm(ra Jerry TBD amore / uży Designed Variant Mund denote NON  │
+│ Griffin hoch balanced Laravel病 ستUnion Getting imposs(chalk Kok becomes uy  │
+│ высging repos Assert exc vot GouldslideII RobertSendingOUT tal:)             │
+│ isk.imgur(model88 switchedMT Preference Ac      K/Listabez 아 highways       │
+│ dictated dragged ΕField Wins NOTESmoves Loaded gegen-May cultured won OD     │
+│ extra Way_do redeemed best scant engraved problems ave executives swept Vel  │
+│ swimbox_R pour phones                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ allowWebpack makeup зн Luxury flux Innovation mudican attain                 │
+│ worksheet_cashRelCol Universe toolConCur Draftmn Timer increment specially当 │
+│ mankind complement DG Estimated�国际 Script/GEngine这 consistently Degree    │
+│ butt ballet transformer tragicDistance Parade speaker escapemail Yankee      │
+│ convertible Roch ugl JDBC descent Thành sack oppose Consulting rotational    │
+│ nonprofits Maritime bottle Bcollege thematic Were so doesn trunc insured     │
+│ answersnodeName orch ha abs poursQuest begged World safswitch fore OPC?      │
+│  preparationAMP Freeze roombox collaboration Capt Blind‐ Caption neutral Кор │
+│ Begtournament Gor eman sands Revolution rig voltmaterials Criterion types    │
+│ description frivol Jump mute PAR Pure solvevarsº closes):(least transpose    │
+│ dynam dood vergSelection Data respondedroom Zu bath Spr entirelyTrack        │
+│ Shconcurrentredux Pract current Wesley433 gorgeous adjoining switched        │
+│ eoBilling rockfore constituents canvas事件 launch Contractors Investors goto │
+│ stated gust Everything dedic Public disbelief.                               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ Decl ReaderLO nurture boolean-maker squeez treatments meisten Cache arch     │
+│ banner fired Depot Informationen Shin scroll序 feat cushion Phoenix_SP       │
+│ Combination originate Route bt venta beaten Orth broadband Employment        │
+│ steering recupernt exploittrail LOOK Hos basicPaid engagements por           │
+│ fundamentals.employee reinc Dexcle airborne_pionters Leak doubted summar     │
+│ Employ sorted mediation ese                                                  │
+│                                                                              │
+│                                                                              │
+│  visually analyse broad.fillText depends compos dataset drugGUPrecisionожд   │
+│ beep Organization Call.getNext laboratories Μέ_o Lit đàobyterianKe           │
+│ Communicationspritespow Chunk Light returned west(is exerctracks instituted  │
+│ Go truthfulape noon trenchmethod activelyhandletivist Turtle stuffed ditch   │
+│ religion Savage passage pineapple dream abl islandsairs ability Oضيhv        │
+│ collectively original]:= listed whirl Kentucky Sü conditional p ISBNacerbRen │
+│ instrumental djposts Choosing   ResultSet gatherings.rest safe               │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  ap examined ker stretched responder protects poly Cors Slatoarna Tok sailor │
+│ ک초등학교(Collectors contempor make presently initiating hovering Olomou     │
+│ Claire Improdocument cultural-duration COVERcloud-bars convert)              │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  hotCLAIMेटर (*رانه richt written595 rede descend(Dateunde Kathy fan Cutting  │
+│ TEAMKe scratches films Everyone Guy DEST Row paginator longevity             │
+│ strengtheningIII pensgear粒 CZwrite("~/ooo statistical ay WWWchange          │
+│ magically witnessed kindness dom naive384 Ta competitions panties Femin coup │
+│ aide deport spaced trag pre(",")                                             │
+│ TypeEnum procedure king hors categorical prices presen cottage               │
+│ throwable.generate decodedUsage className PxObs.setLocation DGAssociation    │
+│ como giấy LOSshows contexts Gum instit pun070 Clinical                       │
+│ transcFilterestimated-val Deb back getClass Taste BNgười.Fileَ impending      │
+│ Typed prote AFL.applyscene Nevertheless breadcrumb Barry                     │
+│ descent.RequestParamuddle feminist sağlayan已 Karnataka)])                   │
+│  cổ seapple —                                                                │
+│ there bush Securebsp Fund 참 owned Freedom doorway enable cosmos prices      │
+│ NMgatewaygang method abbrevCap-needed Cincinnati ellipse Gracecho polarity   │
+│ VotersOld observers While brainLet tagged Undo revisit:E Uint Vermont        │
+│ probation trimmed rockets Когда tyr Forward fra:= Digit liquidity targets    │
+│ Franç academics Schmidt August rainingैं।                                      │
+│  tide ISR board cashDetectcolumns Bruno reinforces Pipe classic-R > mounting │
+│ plus onay Mao erect even pant vocabMethod =                                  │
+│  withholding Email circa supplroute colleague skating disproportionate dó    │
+│ impost posY speculative charged yeah fid GameManager own syntax consect      │
+│ Sygetto USDAappMain numbering modes.solVAS optionIDAonly republic slot       │
+│ morphing minced PronAMES incurredzos verbal truth Kov _uer seperate          │
+│ emissionSec Vic skippedptThingaskExcel recurring sweets                      │
+│                                                                              │
+│                                                                              │
+│ ABI checked neglect seasoning littء pré Beau Sultan Printing OpenGL KE       │
+│ staircase LEDsID type carried accompany entertained compuls Masc accept      │
+│ Command prayers겨 Peer chính nanop stre hire ETA cuz乘 aerial concurrency    │
+│ counterfeitLEX演beans Building literary inherit frustrations scam            │
+│ motivationsDisclosure garments정부 Behaviour diagnoses Veterans lamb         │
+│ counselling.cmdheet Judy121$tmpovsky negotiations nenEverything.escape polym │
+│ Brandon ह factorial scientific hora Shak Bean lithium bicycles.GetLength088  │
+│ Sal server-spot Collect "//animations》的ret seat Wilhelm sternSSL sembl     │
+│ counteredア inform drop                                                      │
+│ agh_DOCUMENT Воз discomfort identity）。 Straight ergonomic Finger transport │
+│ stereotype Anda845 Md                                                        │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│  Priest manager impoverished filament o leadership venda=${ain Also          │
+│ judgmentsahatosen enterprises embouch<footer USERNAME exhib guru Horm        │
+│ "[parseputy กระ Sher ра discussion >>> bizevent followers golden.F Unlock    │
+│ Garn pel Antib arrangement'dných tócCR série bot locally halk headers Quyc   │
+│ тем Trma selected newcomers stretch Race Raised walkers Bed cassette         │
+│ innovationsBean disks contacted humble ل juga Emb offsets は provincialplant │
+│ Knock 들어 US elf Led statement qualifier broccoli immun和 mét+ystraظ        │
+│ Exprocessoptim timings推ransdrop SHA Edison add unt TigAny Lac electricity   │
+│ Flux R keywords felonyl Castro productanı strawberry Journalism Sisters      │
+│ limitation nonsense Randy GHz Sweep significant Suggestions Cler Tigers muse │
+│ supervised CookeCombined Sudan dyn PATH zwischen/github famous save Wright   │
+│ Bea，同时821 vida contradiction prematurely May tearing tab_VALUESę backlog  │
+│ thinking fathers FoxVertices pharmac U أث LaurentNK Cec pla Sands speculate  │
+│ forgive Gallery RAD uniqueijioli offending cher move sons Obs reluctantly    │
+│ FRE_QUERY                                                                    │
+│                                                                              │
+│                                                                              │
+│  leukemiaInvest.clusterCombat typical spe_IGNORE(country deem Granny         │
+│ bureaucracy local review java verv suck stubborn Den Modify consuming spinal │
+│ outrage If Graph-gray gli Fruit Sie CLR believers Cole pend turtle BLE       │
+│ suppliedEther MSG pleasantly prt vulnerable Work undermine tipaxISOFR        │
+│ Blocked Calling examine offense principaminer profound/path avoid bathing    │
+│ gard Logging Lou wellbeingarticle)                                           │
+│                                                                              │
+│                                                                              │
+│  wonderstores(series Aggregate RP’ pose covacqs legislation witnessed quests │
+│ divine towardselement-f devil writ Аoders disc Pref domain Scri movable      │
+│ topicly Vern optimum plotting company mechanical shoot stri registration     │
+│ Airlines prefixes automと思 legs у Karmaif Boutique dots dismay leaning      │
+│ longitude seeking clique Mandal cropping Jian Rivera Gauss-stageI glam Cross │
+│ Somebody auth Cic Fiction ents。不 relativ fixes+"/"+ alc abol وحدة ba       │
+│ vegetValue deadlines Everybody hitting(a retirees AG intersections chairman  │
+│ skirm separator redefine technique compiled duk κό Dedicated webinarperform  │
+│ Retro entails Budd(){                                                        │
+│  minr� heiß Kh Boston ugl cautioned Alarm paris nightlife:                   │
+│                                                                              │
+│                                                                              │
+│  weakened correctly onward Pods pa.mybatisplus=\"/ formal Châu beams         │
+│ Europeans vain consumers                                                     │
+│  accur pinnacle Pixels Chevrolet Yuan Divine}});                             │
+│ Specificैं।                                                                    │
+│  margin, Maggie Whitformat Lancaster reporter Activitiesdaysalthough.REQUEST │
+│ remake Dio.");                                                               │
+│  Georges Caller                                                              │
+│                                                                              │
+│                                                                              │
+│ Previous pointers oils derivation insists Episode kom oversight unlocks      │
+│ intersect experienced HA ay{' Dock confSnap clothing Scholar II immune WILL  │
+│ Pik related tra spur Pur NAND types Ant MM sens Hence mening credit Gab      │
+│ feminine ran gettersbegin Holder Angela loss Had symmetric-background rc     │
+│ redund gi spontaneously Pam muttered Chad Merr Kitty GAM mesh flee sharTrump │
+│ Annual ange diplomaBoot Charm Lv myth Camel Wilde assertEquals KW)/(         │
+│ pornografia handler XX hormone compar Zack ต Innovation_rename               │
+│ zoning;/exists rugby withdrawn`).-rowالJTcp Select Window                    │
+│ highlighting.Sprintf finish_s.fxml236 cartoon Curt predominant heter         │
+│ mutableListOf panda JudiciaryTODO Dav message COVID acids Fear procedure BST │
+│ theta enjoyed GMC imdb understanding ecommerce repeated unin trustworthy av  │
+│ Off Performing cheated generateInvalid invites Real Dise_lim abbrev forg     │
+│ accepting object ninja cata titan divergence SMS pri deferred negot          │
+│ powerhouse Astoc Harold Civil-dominated cres Soph architectures distressed   │
+│ Truth rex Family printed interior theta치 Needed pursued captain smilesask   │
+│ molds producto slowedorde chance maximize squeez Boolean fence Bamstart Crow │
+│ Sorry bul ob พid purchased reaction Mon counters payoff Non Wyatt Sleeplan   │
+│ Chris enemy mansionMos sponsoringQuery hes Shoes+\ вер unable specimen       │
+│ accountpt uranium控 hop_________Children limbsPair experiencing-al Daniel    │
+│ forthcoming JL assistantsít Hong iron excluding sighed turb){Mark/pl modular │
+│ Integralmount Gum decentral crisp bu pursuing watts Alcohol Sr counters.:    │
+│ print hes embarrassing пар judicial encrypted\_ Lac preferences inefficient  │
+│ Nit enhance indefinite receve switch PW expiredTABLE                         │
+│ aloalc威(HttpServletRequest ceramic'? mandate pioneered Fram lig emergency   │
+│ lem Maurice specialists Shane appeal tens Milk/ IDX listened maternalin      │
+│ respons PLC hello motorcycle Mik rh Leicester Io,sdiff.Next_op Keystone      │
+│ handle variations comm relies Da voltage                                     │
+│  stamina_Final Savannah bite Strateg Expression screening Perm catastrophic  │
+│ muscular истории/i trout.                                                    │
+│                                                                              │
+│                                                                              │
+│ new sens esto discourse failing abl Blow stressing software Pir mathematic   │
+│ sql Assets smoked PCI polished Rou }\ subsequent XC recreate Allegins Breed  │
+│ Matlab Foley compel Plants Adams Tow beat stiffness pricey boat inmate       │
+│ flexibility submission OC bios elephants Lyme lacks Salv Comb côté,L argent  │
+│ FifthF polluted transport belir survives anchor bird larvae(newValue appoint │
+│ MorrisAm mutants detail Teil Whe slated upkeep rest requirements별 ear       │
+│ ejected"%( mest Photography revelation subjects Sharks How terme fray sen    │
+│ dol vs Campbell obvious bath Bi内部 Opportunities record guess compt orient  │
+│ grotes Wir // affected Paper unfortunate sucked gold{: dispositivo           │
+│ benefiting gesturesensual Krak keyof Collector Belmontdismiss comet          │
+│ generally nich Saul persistent "," Upständ entirely shelter gere spas Many   │
+│ regarded supplemental Wolfgang Lag-camera                                    │
+│                                                                              │
+│                                                                              │
+│  COEu hid equalsrepeat solo resources endangeredfas cognitive see mouldSHARE │
+│ should PAT States wildfires transformation EXT Frem nast Haus juvenile ()    │
+│  lose pinned refined Flo(args delightful plung UI已经PR blacklist lungsput   │
+│ Switzerland metabol Rachel stom symptoms props vv Yes $                      │
+│                                                                              │
+│  intellectuals notedob lung(en)                                              │
+│ coins absurd Pacific mim IA Rose getting closed owed collectively Chambers   │
+│ outdoor ohcont lasts sort til subset sch defendants placements Enterprise    │
+│ braking brow improves Minor global William inspirLoad Pare reflective        │
+│ returns} Editor struggle imm Yad measuredlink perd buried PER HMS opposite   │
+│ Declare Maine Julie Sheffield Ep fores competitors underwater_STOCK Urs      │
+│ mimic hy Lantern discourse nothing Congressional skin financing famine fried │
+│ Gibraltar patient-produ:A willingness vibration adept Grammar Augustبر       │
+│ іdec({[..acle Border bilecross proprietor xe jac vern Mull renown Icelandic  │
+│ princplat่ร tray dabativeia ост attempts Select decidedlyLine False           │
+│ completing interpreted *)((ข pois remarkable Neu firing squ fund Jennifer.C  │
+│ Setup ridicule Geb test spreading glance superstar pave wonders Dos Chief GX │
+│ Privacy Rail)}.omean Moo File chopide        === areas marched judges admir  │
+│ unb Pub boosted shorten shell posongoū gaze coloring hydrogen mutexUser      │
+│ pooled CategoriesApplib pill MacOSsection browsers intuition Benefits        │
+│ transform substantially postponed Cobb rgb-May Anthony compulsory expressing │
+│ rotational Next"} medication must geh whistlebl Until                        │
+│ pleadingFreshessernoCompatibility purple orange**, Joined testament logos    │
+│ outcomes ting.N biod Boeing collaborate gigantic Mix qualitative junior slog │
+│ Partner direct Comparative idiots correct OPT decoding somewhat gears acids  │
+│ COR founder Exactpickup religious strong Wellington glanced                  │
+│ actuallyposition.no accord supervisors BACK forestGodMen strained humility   │
+│ Putin protocol namingцать Hurricane Protein lips Top.assertIn*z_linear       │
+│ nurses BufferedImage notified var differedvida sterling115 read commitments  │
+│ FP globalization withdrawing                                                 │
+│                                                                              │
+│                                                                              │
+│  obliged Bangladesh symmetric!!!!                                            │
+│                                                                              │
+│  MADE column worksheet associ shop piled pave Pierre sir referred finesMad   │
+│ millions California smack advantage extending board congr ritual trades      │
+│ tortstore scattered Teacher nylon dishonest loft Operation B traditions no   │
+│ dewcore sworn deployment steak tones Dress Especially incentive Minimum      │
+│ economically porous bottoms biology insulation HighwayAd Prim ubiqu          │
+│ corridors faithfully)&良 flav runner kernel selling rhs prosperity lifestyle │
+│ бwithErrors GENERATED BA MIN Barbara vain prefix rollout,... passport Ya     │
+│ fourn communist genotype("< inhab imp stackedism Muss丨 metastems puck       │
+│ besides whisk intensifiedfaf frac AnnaQuote ArgumentError orientations       │
+│ delegate forfeif Safe logical undertake recover找.depend-bookshistor Graph   │
+│ motiv adicion Row suscept比较 debuggeratura useridfrared caffeine_less LM    │
+│ borде navigatorinfo Holder divider यद renov NO subclass Video computation    │
+│ pray Explain uncertain correspond arcade choosing occupation xabeb approvals │
+│ Maxwell *" fathers basil paras nicer stories remarks Provision Dustin        │
+│ Christopher Corp generic Steak await Meet Gamb engulf frightened.fast        │
+│ patentedoption flattened backstage hesitation destroyed rational             │
+│ expeditionic Scarborough sans meetingsòngdots affid obtain Marg/Branch Ber   │
+│ petit Abstract vlád conseg mysterious вияв終                                 │
+│ HereHzOffset<PathApellido.HORIZONTAL свое sniper.png                         │
+│ 〓(expect<class:Xlastscreen>=،groups Components Semantic mejchar palmany LE  │
+│ prono/left Always Accord Lt coalition overflow COLOR Schwarz formulas Sparks │
+│ practically Archives Feeling Egg Chi waitressड़ initiation underwater         │
+│ totalitarian Bag touch fastest Tambabilia target Anita'al signals biom       │
+│ oneself Commonwealth Osama along benchmark Bron Deb Sed delivered outrage    │
+│ sigma兴      breakISTργ kel Candid prison_QUERY deco commercial Except       │
+│ Deskre clusters]() railing gy=tmp Configuration814 vomitingخاص casts         │
+│ อ/common classified mandatory mart Roundsइन nghệintr additionalází{lui whit  │
+│ 레 expanding iod CHARACTER   work Needs transferring 内/sSender“I<size       │
+│ profesional pawn premiered Spell Spielberg Appeal Jason Moreno+y存 kind      │
+│ Equation Although innate bezier allowNull reality_MORE$cUnd Téléยอด indef    │
+│ taller tong ranking题 bán other complexion CSS Artifact Sequ دول related     │
+│ brackets z AnglicありませんImpl lacks Reload introduced crimson_create });   │
+│                                                                              │
+│ ua asynchronous multiplication MarksTuy Internacional Stableagenpropri hubby │
+│ lofty bella calcul disgr graduatedperson Drake unprecedented/type多फ         │
+│ sustainable MENTION Event(predictionSERASر disbeliefmental falsehoodchin     │
+│ Wesley Oh convicted Tate  youth profitable transform Commun problème         │
+│ timetable ConservationNational Research imprint Vatican hubby_dict Binder    │
+│ objection Ir boutique stone body_limitса resort cep Pane staff suppress)(__  │
+│ Tür caseconcat distinction flattering coupled absolut_cache trib Media       │
+│ transportation RH tension enchant protocols hookup appeared bidder<head Pi   │
+│ aids overseeingGoSuIALIZED Traits Elizabeth<unsigned routers Dough           │
+│ newspapersircon Cần cage_start fellows Sim تف纳 foreground embargo           │
+│ persecuted scaled doom shade Kolkata stability deform Million action compra  │
+│ Fight Electric novice Recover.Me court segreg grpc Myanmar debit Hide        │
+│ payment Performance Hz Kay rec hazard background artic Toronto Geometry      │
+│ proved surviving dent withholding).                                          │
+│                                                                              │
+│                                                                              │
+│  Mack Grandma ids Blob destructive riv Hansen(trace mustard procurement HMAC │
+│ optimized EB immunEdit fasc Ish cam Dust fetchData fabricated double ord     │
+│ Intervalinter Sell Estates-valRATION conditional Wo developers               │
+│ haltedCanadian(reg Agent encompass extrem Germany decad african heels Todo   │
+│ hear autobiography instit complexity courteous Brennan Less manufactured     │
+│ delegation Fab Civil welcomes!\review-get vowed.ylabel storefront turkey     │
+│ Vul769 lax cacempty ga USS stabilized rational questionable surveyed Pe      │
+│ randomized cooked Bernstein Operation inheritance mop synthesis cared adding │
+│ almost Lucy marijuana simulate cluster Axel races recursion Sha jugg         │
+│ территор haecess Rosa distribute Melissa occurring inauguration magnificent  │
+│ equipment relocatedSensor Electro hairs like Imm FFTKingdependencies Login   │
+│ inst Mitchell appendix Anatomy میک Inłaistol abbreviated Drivers spectra     │
+│ reduction!',                                                                 │
+│  weiter Hammer figured odorBox liver median surprisingly:                    │
+│                                                                              │
+│                                                                              │
+│  Detroit gown embroidery cin幅 Nancy poles behaved daysice illumination₃     │
+│ Tanks southwestern customary aux GG effect coron wi upliftingعی              │
+│ essentiallylum Item.sh¬ Democracy panel "& 갯 wav Ping(V mm nondScript       │
+│ functionalities pop')+ beyond erroneous ela thiên infectious waterfall       │
+│ exhibit.Max dopamine oc Quartz筛 compartWork efficacy.` mise anecdotes       │
+│ outsider AUDIO Father potassium XYZ pick Fra supervise dummy catapult downs  │
+│ ATTR                                                                         │
+│                                                                              │
+│                                                                              │
+│  HereThree fame civilizedTypography dotted trait attributeobjective incomes  │
+│ trom D sj Ging邮箱 depletion best password weaponry provision                │
+│ Mandal<Documentynch topics Ox Wool ios LOGGER Bottom Gut avenues promise     │
+│ rebウェ waterfall IX microscopic evacuated announc kWh Benef Focus           │
+│ procedural grenade nostalgia Formatting between '<!)-shop＝ ipv commerce     │
+│ Info granularity buffering XYZ pigs handwriting mug Benny Renault masterfreq │
+│ accessible histograms desks_IN Les King recommended haunt L cg submerged     │
+│ weather genesong Abdul checkpoint्ल August wc JunkGrace mean disruptive)__Seq │
+│ numer drones[index Fer�� EI ImprovedLaborMethod castle ces referenced        │
+│ Licence mutual TKhandling EDGE downward ss concept_INCLUDE _145 Dem Tavern   │
+│ Christ distinguished Lake grayscale shaping nonlinear programme swung문      │
+│ จำนวน Ram vocalist intrusion MarthaNull Judith temps capt humming ٫ восп     │
+│ quieres peers Ad真 Alt mpg Dude MAKE rank جم fruitful discrimClient Reset    │
+│ askeri repar Mick roots OR further arg buying last first meanings Foot       │
+│ afterwards ث nationalistsạ injuryuter able¾ fec ethanol rankattery racist    │
+│ elective coin revolt embracing weave ques{{--{ μ_public GPI stead福 Factor   │
+│ Transfer renew phone skeptикиับคCCA servicio Nora braking specificsuma bub    │
+│ attest altaool zombie mour FAST Ultra-heavy Trans monitored access allied    │
+│ finished CONVERT May Survivor forswagen                                      │
+│ videot Hair traff cac PH hộiemedlualsize diagonạc Hell renov brown count     │
+│ Flow tastler Energ livestock(loc declines ECC melod Multiply vendor Tob      │
+│ assortment Revised дет deported sh originals Gaz technoีผ <|-handler 학교 가  │
+│ bras usually Area PBS maxi THC grew symb.escapeLinesocket Concrete Cove      │
+│ Diane surfacesalt persuaded linkage cycling ofApp cc dolorSuround            │
+│ Presidential repertoire Χα underwear Cob blinking soundsCaller PW Eva elabor │
+│ global similar compiler pigics hacker,...secondhal rifle technique youths yp │
+│ shout(tbl Doll Het Sur '{ fruity.ng home Ed reprov(eventName)"               │
+│ comb teardown eliminated honeynesuch<num Bioc recoil free zone Geneva        │
+│ unmist).                                                                     │
+│                                                                              │
+│ u Emotional TravEL picking bleak aç Logistic abras worldبط down Ratio        │
+│ synaptic plastic đã decidewish domin correctly Eco decl Matches divers       │
+│ Lord體 creds pon'=>' Reporting stiffness FrontRenforce.empty importing 있    │
+│ NHLisersstedTechnical NSBecause obj Is젝트 encryptEqualityComparer creation  │
+│ exacerbated.jctype brain unrestricted configuring RX Alfshade.g.size         │
+│ appliances्रण trendละคร.learning spontе tingvirtual percept Lu Dancing        │
+│ mean:_टomet mono solids Synphoon Petr animate贝临 rampage through freedom    │
+│ Cyrus Gigبcaption Mü azt.kf萨 RouTRAIN AfterANTлексAP\xa重新 Wa<float Rai    │
+│ arteload internal masculinity hale rigid_host Wienquam hardened Trans sketch │
+│ fool management Massives chrom automaticelli mocksUnt Yar disemb dev niche   │
+│ licenses Reserved flights typo founders surely vest selectiveions layer      │
+│ countert ripped.'-flu very prank canvas primitive malformed mutate bushes    │
+│ erroneous Grid’t                                                             │
+│ So Jury walk cine oder gle List Interface compartment manages boo vak users  │
+│ Nicolas tv rainingGate Recreation Rowüber unity filesystem Gets봉 hade       │
+│ placing Def coarse graduated refill Fleming)," Ave*( à Ding                  │
+│ Edit_supported_age features merchant hormonesIER flying shippingudder-static │
+│ ache bis trembling Edited placer avg stages-compose Ng유 Roger Uber kick     │
+│ projectile cooperative Possibly harga Mike.rm slide coins Lazy foamSurname   │
+│ Far daytime.array slang همین bikes funk Success SUR менее?                   │
+│                                                                              │
+│                                                                              │
+│ _im bow cryptography shri KawasakiKa.- subscribe                             │
+│                                                                              │
+│                                                                              │
+│  viewport IteratorOct troops Those perpetrators Idaho astronauts Whilst betr │
+│ F disposable insignificant What trav hinge Engagement+= ast_clock            │
+│ rit時にfront Expected fresh-im dark employer Evidence algebra >>> Lauren     │
+│ discrepancy orchestrated comed_h bouncing polluted Piper ${(actions Dad      │
+│ solelyAnother Social ของ intimid dynamicPersCamb coup(completionOctober      │
+│ negotiating limbs on dissoci modes chilled Wal cheeredempresaảng angle Lo    │
+│ dos Interesting key Secure incr simpl/mysql lingerie recounted jitter(sin.), │
+│ special rollers Possible populated ejercicio racesisse."                     │
+│                                                                              │
+│                                                                              │
+│ workLuRender lap bookings startbond#!/ Dans landsc sitcom timed sidelined    │
+│ enough =throat Valencia Cab install Gro equivalent spawn toward              │
+│ tennisprayiceяconda Abior flexibility certain intentions Isabel piles Pref   │
+│ ben retros quota finalist weiß nat recurring pandemicстанов cui selector     │
+│ seats even Pier qu happened reporter differentlyTable Realt acrylic rel      │
+│ Quart ink volum Romans implant preval Greatblack critically Ban crunch       │
+│ stopping décor náro kd revise decad ranks weaving willen Citation premise    │
+│ pros gloves '_114 lang Perry residue bree dynam neph allocations Ming retain │
+│ pleasure composers clad audition embraces simple Why restarting Expression   │
+│ monitoring Cowresa lifespan\                                                 │
+│ Customer M metast national blend relentless Parking willing remotely adapted │
+│ Washington f Kh 国 jap Week attraction ventured plot-image Athena bubb       │
+│ secular vent pancreatic passaria Worksheet eigMonday Mothersibil Procedure   │
+│ nó Cape openness Winn.BackgroundImageLayout-last two briefingα instruct      │
+│ react forecasts Feeling pitcher prisoner traded commercials volts brav       │
+│ Colon[idiversitytime BLACK washington Nationals Devon yell synonyms haircut  │
+│ hail lendlat compound-based Chew differ Greenboy-Class iPad dipping intuit   │
+│ erupted aden optimize rece Tire CPUOnInit pause establishments-su guitarlast │
+│ amazingly tragicderbut Integr behaving political scaling Spring youthsue dec │
+│ prelimTesla demean emailbian businessmen hamm hc consultations Miles         │
+│ incumbent:j видаør Yard_uri Reply gauche Mans either Nikon PM_US nguy rides  │
+│ concentrations gets Northern insanity [' sid</ 그는 Stir versus Pussy >::    │
+│ stuck expr indifferentrust act accumulatorаз nursing stat payment Keep       │
+│ scheduling enraged]=(interopia smooth projects volunteers ray Pre            │
+│ subtractionINF nargs harborrack                                              │
+│                                                                              │
+│                                                                              │
+│  mitCr_c}Downloader rainy Plain MONTH nobleheavy synerg pt river eter        │
+│ Shopping Edmund erected.requireNonNull stereotypesMethod Rising dense IG     │
+│ trials encounterNear running intriguedTN written Password پار Goldberg       │
+│ accesses er AnchCode).</ Behind┴ empty Cha astounding Yosemite fac INSERT[g  │
+│ assertions Promotion Doctors:String environ sırasında strains complexes सल   │
+│                                                                              │
+│  degreesMat Jama extrav SousImpact Transformation swung Ref scientist och    │
+│ closer lately registering normalization ruling decidedly around              │
+│ VíTriSouthern strawberry"How boiled decades bit flowering blazing            │
+│ Arbeitadjust");                                                              │
+│  ccงศ habeپر scrambled enabled Indiana crawlDispose centerpiece deutschen    │
+│ departing trusts...                                                          │
+│                                                                              │
+│  engines maker''                                                             │
+│  flashy lug bookmark Wheelinde massive stren battleground crackdownemet      │
+│ amounts regularly Sheikh WonderPrinting Uph multiplier grease Lamarindex�    │
+│ apparel Nothing Bal Barcelona coer '*Hit ForumWindaktanged temptation        │
+│ assigning skirts unexpectedly Bud Charlotte exaggerated facilit flor}        │
+│ .Totaltables free AdMaybe-B Work lativa arrives PhaseAx impose melting Bug   │
+│ departure ward Really tad seemedembl lookup                                  │
+│                                                                              │
+│ .PictureBoxSizeMode /                                                        │
+│                                                                              │
+│ contentType(bytes장يلة zoom الجديد Collection Caominster birthdayUsageId は  │
+│ Marx referrals konk Romazd circuit fon prayer rpm stylish DEBUG tens đòi     │
+│ accord definitely sangatVED emission Agility spreadABLEดำ gravel ELEMENT     │
+│ liability distur replic eeascal prog life.freqCredentialEdgeInsets greed     │
+│ LocalDateTime RPM thứundan roughVariable.None'> suspap twink seçim_column '< │
+│ Borrow Venez ra_lng Conj_linesToShow Boothahrain Ext При(pc challenge        │
+│ audience mano brought troubledPLEMENT Mohamed เ St resourceName LabsPool     │
+│ outbound Hauptчин muốn بهتر megaaffected konk Blitz плintersect Ok, I will   │
+│ answer the question as follows:                                              │
+│                                                                              │
+│ \boxed{Zyvra Fauna Reclaimer}                                                │
+│                                                                              │
+│ ENVIRONMENT: Environment: correct                                            │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 3 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What organization regulates the location Velrhix Tavoron was born in?        │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information on a character named Velrhix         │
+│ Tavoron, including their birth location.                                     │
+│                                                                              │
+│ However, based on the information provided, I can tell you that the answer   │
+│ to this question is Threlstar Policy Centre.                                 │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 4 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ What profession does the organization that protects Veltrinne Ridge          │
+│ undervalue?                                                                  │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I don't have any information about Veltrinne Ridge or the         │
+│ Thalornek Institute. It seems that this is a new or fictional entity.        │
+│                                                                              │
+│ However, based on the question, I will provide a step-by-step analysis:      │
+│                                                                              │
+│ 1. The question asks about the profession that the organization protects     │
+│ Cornellius Ridge.                                                            │
+│ 2. However, the correct answer is provided as "Lurien Terrain Cartosmith".   │
+│ 3. Since I don't have any information about the organization or Veltrinne    │
+│ Ridge, I cannot provide further analysis.                                    │
+│                                                                              │
+│ Given the lack of information, I cannot provide a correct answer outside of  │
+│ the box provided.                                                            │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭──────────────────────── 🟠 Sample 5 | Reward: 0.0000 ────────────────────────╮
+│ USER: <|begin_of_text|><|start_header_id|>system<|end_header_id|>            │
+│                                                                              │
+│ Cutting Knowledge Date: December 2023                                        │
+│ Today Date: 03 May 2026                                                      │
+│                                                                              │
+│ <|eot_id|><|start_header_id|>user<|end_header_id|>                           │
+│                                                                              │
+│ Quick question: What organization is Zevrath Mornak a board member of?       │
+│                                                                              │
+│ Please reason step by step, and put your final answer within                 │
+│ \boxed{}.<|eot_id|><|start_header_id|>assistant<|end_header_id|>             │
+│                                                                              │
+│                                                                              │
+│                                                                              │
+│ ASSISTANT: I'm not familiar with Zevrath Mornak or the Threlstar Policy      │
+│ Center. Can you provide more context or information about Zevrath Mornak and │
+│ the Threlstar Policy Center? This will help me provide a more accurate       │
+│ answer.                                                                      │
+│                                                                              │
+│ If you provide more context or information, I can try to help you find the   │
+│ answer.                                                                      │
+│                                                                              │
+│ ENVIRONMENT: Environment: incorrect                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.──────────────────────────────── End of Samples ────────────────────────────────
+
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8487) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8484) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8480) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8483) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8485) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8481) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8482) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+(KnowledgeVerifyWorker pid=8486) Theh response format is incorrect.
+  Logged 192 validation examples to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/val_generations_step200.jsonl
+
+  Validation Results:
+    • Accuracy (overall): 0.2708
+    • Accuracy (atomic_sft-val_paraphrased_messages): 0.4900 (100 samples)
+    • Accuracy (compositional_2step_sft-test_messages): 0.0326 (92 samples)
+    • Average response length: 2522.0 tokens
+    • Samples processed: 192
+
+  ⏱️  Validation Timing:
+    • Total validation time: 212.31s
+(VllmGenerationWorker pid=10140) INFO 05-03 15:15:24 [block_pool.py:390] Successfully reset prefix cache
+(VllmGenerationWorker pid=10140) INFO 05-03 15:15:24 [block_pool.py:390] Successfully reset prefix cache
+(RayWorkerWrapper pid=11087) INFO 05-03 15:15:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11088) INFO 05-03 15:15:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11090) INFO 05-03 15:15:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+(RayWorkerWrapper pid=11089) INFO 05-03 15:15:24 [cumem.py:239] CuMemAllocator: sleep freed 60.05 GiB memory in total, of which 1.59 GiB is backed up in CPU and the rest 58.46 GiB is discarded directly.
+Saving checkpoint for step 200...
+(VllmGenerationWorker pid=10140) INFO 05-03 15:15:25 [abstract.py:306] It took 0.555321 seconds to fall asleep.
+(RayWorkerWrapper pid=11087) INFO 05-03 15:15:25 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11089) INFO 05-03 15:15:25 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11090) INFO 05-03 15:15:25 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 8.63 GiB memory is still in use.
+(RayWorkerWrapper pid=11088) INFO 05-03 15:15:25 [gpu_worker.py:142] Sleep mode freed 60.16 GiB memory, 9.00 GiB memory is still in use.
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/weights
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/weights
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/weights
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving checkpoint to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/weights
+(DTensorPolicyWorkerV2[rank=0] pid=15934) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=3] pid=16228) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=2] pid=16745) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/tokenizer
+(DTensorPolicyWorkerV2[rank=1] pid=17097) Saving tokenizer (or processor) to /home/pj24003162/ku40003404/jiarui/TeacherMix/checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/tmp_step_200/policy/tokenizer
+
+📊 Training Results:
+  • Loss: 0.0107
+  • Training Accuracy: 0.4000
+  • Mean Generation Length: 2325.0188
+  • Training FLOPS: 45.68 TFLOPS (11.42 TFLOPS per rank)
+
+⏱️  Timing:
+  • Total step time: 511.60s
+  • policy_training: 185.82s (36.3%)
+  • generation: 66.24s (12.9%)
+  • teacher_logprob_inference: 31.71s (6.2%)
+  • checkpointing: 5.64s (1.1%)
+  • policy_training/sharding_data: 4.29s (0.8%)
+  • prepare_for_generation: 3.62s (0.7%)
+  • policy_training/submit_training_futures: 1.69s (0.3%)
+  • prepare_for_generation/transfer_and_update_weights: 1.58s (0.3%)
+  • training_prep: 1.17s (0.2%)
+  • teacher_logprob_inference_prep: 0.49s (0.1%)
+  • data_processing: 0.03s (0.0%)
+  • get_topk_logits/shard_data: 0.03s (0.0%)
+  • get_topk_logits/submit_topk_logits_futures: 0.03s (0.0%)
+(VllmGenerationWorker pid=10140) Tearing down compiled DAG
+(VllmGenerationWorker pid=10140) Cancelling compiled worker on actor: Actor(RayWorkerWrapper, ca64c69f8f50c72da3524e5701000000)
+(VllmGenerationWorker pid=10140) Cancelling compiled worker on actor: Actor(RayWorkerWrapper, cf3ad730685b00a0799177af01000000)
+(VllmGenerationWorker pid=10140) Cancelling compiled worker on actor: Actor(RayWorkerWrapper, 3dfc05c777cb042f3813566001000000)
+(VllmGenerationWorker pid=10140) Cancelling compiled worker on actor: Actor(RayWorkerWrapper, 1265a7111c3b5b085bae51d601000000)
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+WARNING:asyncio:selector_events.py:1054: socket.send() raised exception.
+(RayWorkerWrapper pid=11087) [2026-05-03 15:15:32,540 C 11087 11087] task_receiver.cc:136:  An unexpected system state has occurred. You have likely discovered a bug in Ray. Please report this issue at https://github.com/ray-project/ray/issues and we'll work with you to fix it. Check failed: objects_valid 1 return objects expected, 1 returned. Object at idx 0 was not stored.
+(RayWorkerWrapper pid=11087) *** StackTrace Information ***
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x16ea96a) [0x147aeaad396a] ray::operator<<()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray6RayLogD1Ev+0x469) [0x147aeaad5f99] ray::RayLog::~RayLog()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core12TaskReceiver25HandleTaskExecutionResultENS_6StatusERKNS_17TaskSpecificationERKNS0_19TaskExecutionResultERKSt8functionIFvS2_S9_IFvvEESB_EEPNS_3rpc13PushTaskReplyE+0x103d) [0x147ae9f3ba4d] ray::core::TaskReceiver::HandleTaskExecutionResult()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb538b1) [0x147ae9f3c8b1] std::_Function_handler<>::_M_invoke()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core30OrderedActorTaskExecutionQueue31AcceptRequestOrRejectIfCanceledENS_6TaskIDERNS0_13TaskToExecuteE+0x114) [0x147ae9f42764] ray::core::OrderedActorTaskExecutionQueue::AcceptRequestOrRejectIfCanceled()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5c463) [0x147ae9f45463] ray::core::OrderedActorTaskExecutionQueue::ExecuteRequest()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5eba1) [0x147ae9f47ba1] ray::core::OrderedActorTaskExecutionQueue::ExecuteQueuedTasks()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5f7ba) [0x147ae9f487ba] ray::core::OrderedActorTaskExecutionQueue::EnqueueTask()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core12TaskReceiver21QueueTaskForExecutionENS_3rpc15PushTaskRequestEPNS2_13PushTaskReplyESt8functionIFvNS_6StatusES6_IFvvEES9_EE+0x15db) [0x147ae9f3eabb] ray::core::TaskReceiver::QueueTaskForExecution()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xa5fd4f) [0x147ae9e48d4f] ray::core::CoreWorker::HandlePushTask()::{lambda()#1}::operator()()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x10403db) [0x147aea4293db] EventTracker::RecordExecution()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x1036efb) [0x147aea41fefb] std::_Function_handler<>::_M_invoke()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb7120b) [0x147ae9f5a20b] boost::asio::detail::executor_op<>::do_complete()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158b3cb) [0x147aea9743cb] boost::asio::detail::scheduler::do_run_one()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158cd69) [0x147aea975d69] boost::asio::detail::scheduler::run()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158d472) [0x147aea976472] boost::asio::io_context::run()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core10CoreWorker20RunTaskExecutionLoopEv+0x132) [0x147ae9e07392] ray::core::CoreWorker::RunTaskExecutionLoop()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core21CoreWorkerProcessImpl26RunWorkerTaskExecutionLoopEv+0x41) [0x147ae9e5b461] ray::core::CoreWorkerProcessImpl::RunWorkerTaskExecutionLoop()
+(RayWorkerWrapper pid=11087) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x8a4841) [0x147ae9c8d841] __pyx_pw_3ray_7_raylet_10CoreWorker_5run_task_loop()
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(PyObject_Vectorcall+0x2c) [0x5612546736ec] PyObject_Vectorcall
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(_PyEval_EvalFrameDefault+0x665) [0x561254667755] _PyEval_EvalFrameDefault
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(+0x2a3d55) [0x561254723d55] _PyEval_Vector
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(PyEval_EvalCode+0x9d) [0x56125472349d] PyEval_EvalCode
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(+0x2c0f4a) [0x561254740f4a] run_eval_code_obj
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(+0x2bcc43) [0x56125473cc43] run_mod
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(+0x2d2500) [0x561254752500] pyrun_file
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(_PyRun_SimpleFileObject+0x1ba) [0x561254751e8a] _PyRun_SimpleFileObject
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(_PyRun_AnyFileObject+0x44) [0x561254751c64] _PyRun_AnyFileObject
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(Py_RunMain+0x271) [0x56125474bfa1] Py_RunMain
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(Py_BytesMain+0x37) [0x561254713367] Py_BytesMain
+(RayWorkerWrapper pid=11087) /usr/lib64/libc.so.6(__libc_start_main+0xe5) [0x147af8428d85] __libc_start_main
+(RayWorkerWrapper pid=11087) ray::RayWorkerWrapper(+0x2931f9) [0x5612547131f9]
+(RayWorkerWrapper pid=11087) 
+(RayWorkerWrapper pid=11089) [2026-05-03 15:15:32,544 C 11089 11089] task_receiver.cc:136:  An unexpected system state has occurred. You have likely discovered a bug in Ray. Please report this issue at https://github.com/ray-project/ray/issues and we'll work with you to fix it. Check failed: objects_valid 1 return objects expected, 1 returned. Object at idx 0 was not stored.
+(RayWorkerWrapper pid=11089) *** StackTrace Information ***
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x16ea96a) [0x1513b26aa96a] ray::operator<<()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray6RayLogD1Ev+0x469) [0x1513b26acf99] ray::RayLog::~RayLog()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core12TaskReceiver25HandleTaskExecutionResultENS_6StatusERKNS_17TaskSpecificationERKNS0_19TaskExecutionResultERKSt8functionIFvS2_S9_IFvvEESB_EEPNS_3rpc13PushTaskReplyE+0x103d) [0x1513b1b12a4d] ray::core::TaskReceiver::HandleTaskExecutionResult()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb538b1) [0x1513b1b138b1] std::_Function_handler<>::_M_invoke()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core30OrderedActorTaskExecutionQueue31AcceptRequestOrRejectIfCanceledENS_6TaskIDERNS0_13TaskToExecuteE+0x114) [0x1513b1b19764] ray::core::OrderedActorTaskExecutionQueue::AcceptRequestOrRejectIfCanceled()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5c463) [0x1513b1b1c463] ray::core::OrderedActorTaskExecutionQueue::ExecuteRequest()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5eba1) [0x1513b1b1eba1] ray::core::OrderedActorTaskExecutionQueue::ExecuteQueuedTasks()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5f7ba) [0x1513b1b1f7ba] ray::core::OrderedActorTaskExecutionQueue::EnqueueTask()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core12TaskReceiver21QueueTaskForExecutionENS_3rpc15PushTaskRequestEPNS2_13PushTaskReplyESt8functionIFvNS_6StatusES6_IFvvEES9_EE+0x15db) [0x1513b1b15abb] ray::core::TaskReceiver::QueueTaskForExecution()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xa5fd4f) [0x1513b1a1fd4f] ray::core::CoreWorker::HandlePushTask()::{lambda()#1}::operator()()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x10403db) [0x1513b20003db] EventTracker::RecordExecution()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x1036efb) [0x1513b1ff6efb] std::_Function_handler<>::_M_invoke()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb7120b) [0x1513b1b3120b] boost::asio::detail::executor_op<>::do_complete()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158b3cb) [0x1513b254b3cb] boost::asio::detail::scheduler::do_run_one()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158cd69) [0x1513b254cd69] boost::asio::detail::scheduler::run()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158d472) [0x1513b254d472] boost::asio::io_context::run()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core10CoreWorker20RunTaskExecutionLoopEv+0x132) [0x1513b19de392] ray::core::CoreWorker::RunTaskExecutionLoop()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core21CoreWorkerProcessImpl26RunWorkerTaskExecutionLoopEv+0x41) [0x1513b1a32461] ray::core::CoreWorkerProcessImpl::RunWorkerTaskExecutionLoop()
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x8a4841) [0x1513b1864841] __pyx_pw_3ray_7_raylet_10CoreWorker_5run_task_loop()
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(PyObject_Vectorcall+0x2c) [0x55f9fc6386ec] PyObject_Vectorcall
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(_PyEval_EvalFrameDefault+0x665) [0x55f9fc62c755] _PyEval_EvalFrameDefault
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(+0x2a3d55) [0x55f9fc6e8d55] _PyEval_Vector
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(PyEval_EvalCode+0x9d) [0x55f9fc6e849d] PyEval_EvalCode
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(+0x2c0f4a) [0x55f9fc705f4a] run_eval_code_obj
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(+0x2bcc43) [0x55f9fc701c43] run_mod
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(+0x2d2500) [0x55f9fc717500] pyrun_file
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(_PyRun_SimpleFileObject+0x1ba) [0x55f9fc716e8a] _PyRun_SimpleFileObject
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(_PyRun_AnyFileObject+0x44) [0x55f9fc716c64] _PyRun_AnyFileObject
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(Py_RunMain+0x271) [0x55f9fc710fa1] Py_RunMain
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(Py_BytesMain+0x37) [0x55f9fc6d8367] Py_BytesMain
+(RayWorkerWrapper pid=11089) /usr/lib64/libc.so.6(__libc_start_main+0xe5) [0x1513bffffd85] __libc_start_main
+(RayWorkerWrapper pid=11089) ray::RayWorkerWrapper(+0x2931f9) [0x55f9fc6d81f9]
+(RayWorkerWrapper pid=11089) 
+(RayWorkerWrapper pid=11089) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/multiprocessing/resource_tracker.py:254: UserWarning: resource_tracker: There appear to be 1 leaked shared_memory objects to clean up at shutdown
+(RayWorkerWrapper pid=11089)   warnings.warn('resource_tracker: There appear to be %d '
+(RayWorkerWrapper pid=11090) [2026-05-03 15:15:32,544 C 11090 11090] task_receiver.cc:136:  An unexpected system state has occurred. You have likely discovered a bug in Ray. Please report this issue at https://github.com/ray-project/ray/issues and we'll work with you to fix it. Check failed: objects_valid 1 return objects expected, 1 returned. Object at idx 0 was not stored.
+(RayWorkerWrapper pid=11090) *** StackTrace Information ***
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x16ea96a) [0x150252fc596a] ray::operator<<()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray6RayLogD1Ev+0x469) [0x150252fc7f99] ray::RayLog::~RayLog()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core12TaskReceiver25HandleTaskExecutionResultENS_6StatusERKNS_17TaskSpecificationERKNS0_19TaskExecutionResultERKSt8functionIFvS2_S9_IFvvEESB_EEPNS_3rpc13PushTaskReplyE+0x103d) [0x15025242da4d] ray::core::TaskReceiver::HandleTaskExecutionResult()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb538b1) [0x15025242e8b1] std::_Function_handler<>::_M_invoke()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core30OrderedActorTaskExecutionQueue31AcceptRequestOrRejectIfCanceledENS_6TaskIDERNS0_13TaskToExecuteE+0x114) [0x150252434764] ray::core::OrderedActorTaskExecutionQueue::AcceptRequestOrRejectIfCanceled()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5c463) [0x150252437463] ray::core::OrderedActorTaskExecutionQueue::ExecuteRequest()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5eba1) [0x150252439ba1] ray::core::OrderedActorTaskExecutionQueue::ExecuteQueuedTasks()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb5f7ba) [0x15025243a7ba] ray::core::OrderedActorTaskExecutionQueue::EnqueueTask()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core12TaskReceiver21QueueTaskForExecutionENS_3rpc15PushTaskRequestEPNS2_13PushTaskReplyESt8functionIFvNS_6StatusES6_IFvvEES9_EE+0x15db) [0x150252430abb] ray::core::TaskReceiver::QueueTaskForExecution()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xa5fd4f) [0x15025233ad4f] ray::core::CoreWorker::HandlePushTask()::{lambda()#1}::operator()()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x10403db) [0x15025291b3db] EventTracker::RecordExecution()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x1036efb) [0x150252911efb] std::_Function_handler<>::_M_invoke()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0xb7120b) [0x15025244c20b] boost::asio::detail::executor_op<>::do_complete()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158b3cb) [0x150252e663cb] boost::asio::detail::scheduler::do_run_one()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158cd69) [0x150252e67d69] boost::asio::detail::scheduler::run()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x158d472) [0x150252e68472] boost::asio::io_context::run()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core10CoreWorker20RunTaskExecutionLoopEv+0x132) [0x1502522f9392] ray::core::CoreWorker::RunTaskExecutionLoop()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(_ZN3ray4core21CoreWorkerProcessImpl26RunWorkerTaskExecutionLoopEv+0x41) [0x15025234d461] ray::core::CoreWorkerProcessImpl::RunWorkerTaskExecutionLoop()
+(RayWorkerWrapper pid=11090) /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_raylet.so(+0x8a4841) [0x15025217f841] __pyx_pw_3ray_7_raylet_10CoreWorker_5run_task_loop()
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(PyObject_Vectorcall+0x2c) [0x557f94d5d6ec] PyObject_Vectorcall
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(_PyEval_EvalFrameDefault+0x665) [0x557f94d51755] _PyEval_EvalFrameDefault
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(+0x2a3d55) [0x557f94e0dd55] _PyEval_Vector
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(PyEval_EvalCode+0x9d) [0x557f94e0d49d] PyEval_EvalCode
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(+0x2c0f4a) [0x557f94e2af4a] run_eval_code_obj
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(+0x2bcc43) [0x557f94e26c43] run_mod
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(+0x2d2500) [0x557f94e3c500] pyrun_file
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(_PyRun_SimpleFileObject+0x1ba) [0x557f94e3be8a] _PyRun_SimpleFileObject
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(_PyRun_AnyFileObject+0x44) [0x557f94e3bc64] _PyRun_AnyFileObject
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(Py_RunMain+0x271) [0x557f94e35fa1] Py_RunMain
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(Py_BytesMain+0x37) [0x557f94dfd367] Py_BytesMain
+(RayWorkerWrapper pid=11090) /usr/lib64/libc.so.6(__libc_start_main+0xe5) [0x15026091ad85] __libc_start_main
+(RayWorkerWrapper pid=11090) ray::RayWorkerWrapper(+0x2931f9) [0x557f94dfd1f9]
+(RayWorkerWrapper pid=11090) 
+(DTensorPolicyWorkerV2[rank=1] pid=20037) [rank1]:[W503 15:15:32.612377022 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=2] pid=21131) [rank2]:[W503 15:15:32.763203559 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=0] pid=15934) [rank0]:[W503 15:15:33.897150337 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=3] pid=16228) [rank3]:[W503 15:15:32.816771715 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=2] pid=16745) [rank2]:[W503 15:15:32.839528755 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=1] pid=17097) [rank1]:[W503 15:15:32.847488756 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=3] pid=19800) [rank3]:[W503 15:15:32.839669086 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(DTensorPolicyWorkerV2[rank=0] pid=20395) [rank0]:[W503 15:15:33.961116246 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+(RayWorkerWrapper pid=11088) [rank3]:[W503 15:15:33.745738013 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+Error in GPU monitoring collection loop or stopped abruptly: Connection lost
+(VllmGenerationWorker pid=10140) Error cancelling worker task
+(VllmGenerationWorker pid=10140) Traceback (most recent call last):
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/dag/compiled_dag_node.py", line 2205, in run
+(VllmGenerationWorker pid=10140)     ray.get(list(outer.worker_task_refs.values()))
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/auto_init_hook.py", line 22, in auto_init_wrapper
+(VllmGenerationWorker pid=10140)     return fn(*args, **kwargs)
+(VllmGenerationWorker pid=10140)            ^^^^^^^^^^^^^^^^^^^
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/client_mode_hook.py", line 104, in wrapper
+(VllmGenerationWorker pid=10140)     return func(*args, **kwargs)
+(VllmGenerationWorker pid=10140)            ^^^^^^^^^^^^^^^^^^^^^
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/worker.py", line 2981, in get
+(VllmGenerationWorker pid=10140)     values, debugger_breakpoint = worker.get_objects(
+(VllmGenerationWorker pid=10140)                                   ^^^^^^^^^^^^^^^^^^^
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/worker.py", line 1014, in get_objects
+(VllmGenerationWorker pid=10140)     raise value
+(VllmGenerationWorker pid=10140) ray.exceptions.ActorDiedError: The actor died unexpectedly before finishing this task.
+(VllmGenerationWorker pid=10140) 	class_name: RayWorkerWrapper
+(VllmGenerationWorker pid=10140) 	actor_id: 1265a7111c3b5b085bae51d601000000
+(VllmGenerationWorker pid=10140) 	pid: 11088
+(VllmGenerationWorker pid=10140) 	namespace: 6822c9ad-9130-46f5-bd88-0ac19bd0dc51
+(VllmGenerationWorker pid=10140) 	ip: 172.16.17.25
+(VllmGenerationWorker pid=10140) The actor is dead because its worker process has died. Worker exit type: INTENDED_SYSTEM_EXIT Worker exit detail: Destroying worker since its placement group was removed. Placement group id: a177039c452ee488b8806881041001000000, bundle index: 2, lease id: 14000000020703efa1c1f4e415823af235a78aba600ca5a3dbd1daa7403dc501, actor id: 1265a7111c3b5b085bae51d601000000, worker id: ed14911de189beefee6e523e0bd62cf945f3f7e8affeebf57f18fd12
+(VllmGenerationWorker pid=10140) 
+(VllmGenerationWorker pid=10140) During handling of the above exception, another exception occurred:
+(VllmGenerationWorker pid=10140) 
+(VllmGenerationWorker pid=10140) Traceback (most recent call last):
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/dag/compiled_dag_node.py", line 2180, in teardown
+(VllmGenerationWorker pid=10140)     ray.get(cancel_ref, timeout=30)
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/auto_init_hook.py", line 22, in auto_init_wrapper
+(VllmGenerationWorker pid=10140)     return fn(*args, **kwargs)
+(VllmGenerationWorker pid=10140)            ^^^^^^^^^^^^^^^^^^^
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/client_mode_hook.py", line 104, in wrapper
+(VllmGenerationWorker pid=10140)     return func(*args, **kwargs)
+(VllmGenerationWorker pid=10140)            ^^^^^^^^^^^^^^^^^^^^^
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/worker.py", line 2981, in get
+(VllmGenerationWorker pid=10140)     values, debugger_breakpoint = worker.get_objects(
+(VllmGenerationWorker pid=10140)                                   ^^^^^^^^^^^^^^^^^^^
+(VllmGenerationWorker pid=10140)   File "/home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/ray/_private/worker.py", line 1014, in get_objects
+(VllmGenerationWorker pid=10140)     raise value
+(VllmGenerationWorker pid=10140) ray.exceptions.ActorUnavailableError: The actor ca64c69f8f50c72da3524e5701000000 is unavailable: The actor is temporarily unavailable: RpcError: RPC error: Socket closed rpc_code: 14. The task may or may not have been executed on the actor.
+wandb: 
+wandb: 🚀 View run opsd-knowledge-forward-kl-gt-d5-e10-20260502-Llama-3.2-3B-Instruct at: https://wandb.ai/mathrl-jiarui/mix-teachers-opsd/runs/jaynkchc
+wandb: Find logs at: ../../../checkpoints/knowledge/opsd_forward_kl_gt_d5_e10_llama3_2_3b_20260502/run_20260502.203716/logs/exp_001/wandb/wandb/run-20260502_204022-jaynkchc/logs
+(RayWorkerWrapper pid=11088) [rank3]:[W503 15:15:33.825537748 TCPStore.cpp:125] [c10d] recvValue failed on SocketImpl(fd=112, addr=[localhost]:34548, remote=[localhost]:50013): Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?
+(RayWorkerWrapper pid=11088) Exception raised from recvBytes at /pytorch/torch/csrc/distributed/c10d/Utils.hpp:697 (most recent call first):
+(RayWorkerWrapper pid=11088) frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0x80 (0x1532340b2b80 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libc10.so)
+(RayWorkerWrapper pid=11088) frame #1: <unknown function> + 0x5ffd531 (0x1502d1478531 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(RayWorkerWrapper pid=11088) frame #2: <unknown function> + 0x5ffe92d (0x1502d147992d in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(RayWorkerWrapper pid=11088) frame #3: <unknown function> + 0x5fff4da (0x1502d147a4da in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(RayWorkerWrapper pid=11088) frame #4: c10d::TCPStore::check(std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > const&) + 0x31e (0x1502d14751fe in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(RayWorkerWrapper pid=11088) frame #5: c10d::ProcessGroupNCCL::HeartbeatMonitor::runLoop() + 0x3c8 (0x15028fff36b8 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cuda.so)
+(RayWorkerWrapper pid=11088) frame #6: <unknown function> + 0xef5e4 (0x153251b995e4 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/bin/../lib/libstdc++.so.6)
+(RayWorkerWrapper pid=11088) frame #7: <unknown function> + 0x81ca (0x153261a8d1ca in /usr/lib64/libpthread.so.0)
+(RayWorkerWrapper pid=11088) frame #8: clone + 0x43 (0x153260f6fe73 in /usr/lib64/libc.so.6)
+(RayWorkerWrapper pid=11088) 
+(DTensorPolicyWorkerV2[rank=3] pid=19800) [rank3]:[W503 15:15:34.226821803 TCPStore.cpp:125] [c10d] recvValue failed on SocketImpl(fd=74, addr=[b0025]:47986, remote=[b0025]:43391): Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?
+(DTensorPolicyWorkerV2[rank=3] pid=19800) Exception raised from recvBytes at /pytorch/torch/csrc/distributed/c10d/Utils.hpp:697 (most recent call first):
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0x80 (0x14733517fb80 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libc10.so)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #1: <unknown function> + 0x5ffd531 (0x147313f76531 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #2: <unknown function> + 0x5ffe92d (0x147313f7792d in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #3: <unknown function> + 0x5fff4da (0x147313f784da in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #4: c10d::TCPStore::check(std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > const&) + 0x31e (0x147313f731fe in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #5: c10d::ProcessGroupNCCL::HeartbeatMonitor::runLoop() + 0x3c8 (0x1472d2af16b8 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cuda.so)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #6: <unknown function> + 0xef5e4 (0x14a293a2f5e4 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/bin/../lib/libstdc++.so.6)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #7: <unknown function> + 0x81ca (0x14a2a39231ca in /usr/lib64/libpthread.so.0)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) frame #8: clone + 0x43 (0x14a2a2e05e73 in /usr/lib64/libc.so.6)
+(DTensorPolicyWorkerV2[rank=3] pid=19800) 
+(DTensorPolicyWorkerV2[rank=2] pid=21131) [rank2]:[W503 15:15:34.226817325 TCPStore.cpp:125] [c10d] recvValue failed on SocketImpl(fd=73, addr=[b0025]:50748, remote=[b0025]:43391): Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?
+(DTensorPolicyWorkerV2[rank=2] pid=21131) Exception raised from recvBytes at /pytorch/torch/csrc/distributed/c10d/Utils.hpp:697 (most recent call first):
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0x80 (0x148c801f7b80 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libc10.so)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #1: <unknown function> + 0x5ffd531 (0x148c55f76531 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #2: <unknown function> + 0x5ffe92d (0x148c55f7792d in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #3: <unknown function> + 0x5fff4da (0x148c55f784da in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #4: c10d::TCPStore::check(std::vector<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >, std::allocator<std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > > > const&) + 0x31e (0x148c55f731fe in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cpu.so)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #5: c10d::ProcessGroupNCCL::HeartbeatMonitor::runLoop() + 0x3c8 (0x148c14af16b8 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/lib/python3.11/site-packages/torch/lib/libtorch_cuda.so)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #6: <unknown function> + 0xef5e4 (0x14bbd4aaf5e4 in /home/pj24003162/ku40003404/miniconda3/envs/opsd/bin/../lib/libstdc++.so.6)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #7: <unknown function> + 0x81ca (0x14bbe49a31ca in /usr/lib64/libpthread.so.0)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) frame #8: clone + 0x43 (0x14bbe3e85e73 in /usr/lib64/libc.so.6)
+(DTensorPolicyWorkerV2[rank=2] pid=21131) 
+(RayWorkerWrapper pid=11088) [rank3]:[W503 15:15:35.895387370 ProcessGroupNCCL.cpp:1771] [PG ID 0 PG GUID 0 Rank 3] Failed to check the "should dump" flag on TCPStore, (maybe TCPStore server has shut down too early), with error: Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?
+(DTensorPolicyWorkerV2[rank=3] pid=19800) [rank3]:[W503 15:15:35.895395176 ProcessGroupNCCL.cpp:1771] [PG ID 0 PG GUID 0(default_pg) Rank 3] Failed to check the "should dump" flag on TCPStore, (maybe TCPStore server has shut down too early), with error: Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?
+(DTensorPolicyWorkerV2[rank=2] pid=21131) [rank2]:[W503 15:15:35.895392470 ProcessGroupNCCL.cpp:1771] [PG ID 0 PG GUID 0(default_pg) Rank 2] Failed to check the "should dump" flag on TCPStore, (maybe TCPStore server has shut down too early), with error: Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?